This is delightfully unhinged, spending an amazing amount of time describing their model and citing their methodologies before getting to the meat of the meal many of us have been braying about for years: whether the singularity actually happens or not is irrelevant so much as whether enough people believe it will happen and act accordingly.
And, yep! A lot of people absolutely believe it will and are acting accordingly.
It’s honestly why I gave up trying to get folks to look at these things rationally as knowable objects (“here’s how LLMs actually work”) and pivoted to the social arguments instead (“here’s why replacing or suggesting the replacement of human labor prior to reforming society into one that does not predicate survival on continued employment and wages is very bad”). Folks vibe with the latter, less with the former. Can’t convince someone of the former when they don’t even understand that the computer is the box attached to the monitor, not the monitor itself.
> * enough people believe it will happen and act accordingly*
Here comes my favorite notion of "epistemic takeover".
A crude form: make everybody believe that you have already won.
A refined form: make everybody believe that everybody else believes that you have already won. That is, even if one has doubts about your having won, they believe that everyone else submit to you as a winner, and must act accordingly.
This world where everybody’s very concerned with that “refined form” is annoying and exhausting. It causes discussions to become about speculative guesses about everybody else’s beliefs, not actual facts. In the end it breeds cynicism as “well yes, the belief is wrong, but everybody is stupid and believes it anyway,” becomes a stop-gap argument.
I don’t know how to get away from it because ultimately coordination depends on understanding what everybody believes, but I wish it would go away.
IMO this is a symptom of the falling rate of profit, especially in the developed world. If truly productivity enhancing investment is effectively dead (or, equivalently, there is so much paper wealth chasing a withering set of profitable opportunities for investment), then capital's only game is to chase high valuations backed by future profits, which means playing the Keynesian beauty contest for keeps. This in turn means you must make ever-escalating claims of future profitability. Now, here we are in a world where multiple brand name entrepreneurs are essentially saying that they are building the last investable technology ever, and getting people to believe it because the alternative is to earn less than inflation on Procter and Gamble stock and never getting to retire.
If outsiders could plausibly invest in China, some of this pressure could be dissipated for a while, but ultimately we need to order society on some basis that incentivizes dealing with practical problems instead of pushing paper around.
>If truly productivity enhancing investment is effectively dead (or, equivalently, there is so much paper wealth chasing a withering set of profitable opportunities for investment), then capital's only game is to chase high valuations backed by future profits, which means playing the Keynesian beauty contest for keeps.
What if profit is dead because wealth is all concentrating in people who don't need it from a marginal consumption standpoint, which means asset prices blow up because everyone rich believes that they need to "invest" that money somewhere... but demand shrivels outside of interestingly-subsidized areas like healthcare because nobody else is making enough to even keep up with the asset price rises?
And without demand, where would innovation come from?
If you accept that the economy can be in disequilibrium, then it is very easy to see how this happens.
Rich people learn a habit that makes them consume less than their investment returns. The difference is reinvested, resulting in a net increase of their equity. Even if you say they spend 90% of their returns on consumption, the last 10% still grow in absolute terms. Since returns are paid proportionally based on the quantity of equity, you can clearly see that money is allocated from where it has high marginal utility to places where it has low marginal utility.
Of course this leads to a contradiction. If money is held by people who have a low marginal utility for consumption, why would investments pay high returns? Your equity is the latent demand that your investments need to pay you the returns in the first place. You'd increasingly be paying yourself. That is equivalent to investing into something that yields a 0% return which in turn is equivalent to doing nothing.
Perhaps I’m misunderstanding but a lot of people (ok, well, a few, but you know) make a lot of money on relatively mundane stuff. Technocapitalism’s Accursed Share is sacrificing wealth for myth making about its own future.
In a post-industrial economy there are no more economic problems, only liabilities. Surplus is felt as threat, especially when it's surplus human labor.
In today's economy disease and prison camps are increasingly profitable.
How do you think the investor portfolios that hold stocks in deathcare and privatized prison labor camps can further Accelerate their returns?
Online influencers, podcasters, advertisers, social media product managers, political lobbyists, cryptocurrency protocol programmers, digital/NFT artists, most of the media production industry, those people w/ leaf blowers moving dust around, political commentators (e.g. fox & friends), super PACs, most NGOs, "professional" sports, various 3 letter agencies & their associated online "influence" campaigns, think tanks about machine consciousness, autonomous weapon manufacturers, & so on. Just a few off the top of my head but anything to do w/ shuffling numbers in databases is in that category as well. I haven't read "Bullshit Jobs" yet but it's on the list & I'll get to it eventually so I'm sure I can come up w/ a few more after reading it.
It's curious that you only list digital/NFT artists, and fail to see the problems that many of these solve.
In a world re-calibrated to pure problem-solving, presumably we'd ask a mime or other non-digital artist for news on what our government is doing (instead of the Fake News Media), and then an interpretive dancer, not a lobbyist, would intermediate between government and industry? The trombone players would investigate financial crimes, the bassists would monitor our airspace, and the guy currently painting a mural on a wall would be responsible for DoorDashing a cargo ship's worth of food and supplies to a disaster zone (not the bad NGOs)?
Yeah, this is just a list of jobs you don't understand and/or make you feel sad.
On the other hand talking about those believes can also lead to real changes. Slavery used to be seen widely a necessary evil, just like for instance war.
I don’t actually know a ton about the rhetoric around abolitionism. Are you saying they tried to convince people that everybody else thought slavery was evil? I guess I assumed they tried to convince people slavery was in-and-of-itself evil.
It's not just exhausting, it's a huge problem. Even if everyone is a complete saint all the time and has the best of intentions, going by beliefs about beliefs can trap us in situations where we're all unhappy.
The classic situation is the two lovers who both want to do what they think makes their partner happy, to the extent that they don't tell what they actually want, and end up doing something neither wants.
I think the goal of all sorts of cooperative planning should be to avoid such situations like the plague.
We really need a rule in politics which bans you (if you're an elected representative) from stating anything about the beliefs of the electorate without reference to a poll of the population of adequate size and quality.
Yes we'd have a lot of lawsuits about it, but it would hardly be a bad use of time to litigate whether a politicians statements about the electorate's beliefs are accurate.
The thing is... on both the cited occasions (Nixon in 1968, Morrison in 2019), the politicians claiming the average voter agreed with them actually won that election
So, obviously their claims were at least partially true – because if they'd completely misjudged the average voter, they wouldn't have won
When there are only two choices, and infinite issues, voters only have two choices: Vote for someone you don't agree with less, or vote for someone you quite hilariously imagine agrees with you.
EDIT: Not being cynical about voters. But about the centralization of parties, in number and operationally, as a steep barrier for voter choice.
Two options, not two choices. (Unless you have a proportional representation voting system like ireland, in which case you can vote for as many candidates as you like in descending order of preference)
Anyway, there’s a third option: spoil your vote. In the recent Irish presidential election, 13% of those polled afterwards said they spoiled their votes, due to a poor selection of candidates from which to choose.
That’s much more true for Nixon in 1968 than Morrison in 2019
Because the US has a “hard” two party system - third party candidates have very little hope, especially at the national level; voting for a third party is indistinguishable from staying home, as far as the outcome goes, with some rather occasional exceptions
But Australia is different - Australia has a “soft” two party system - two-and-a-half major parties (I say “and-a-half” because our centre-right is a semipermanent coalition of two parties, one representing rural/regional conservatives, the other more urban in its support base). But third parties and independents are a real political force in our parliament, and sometimes even determine the outcome of national elections
This is largely due to (1) we use what Americans call instant-runoff in our federal House of Representatives, and a variation on single-transferable vote in our federal Senate; (2) the parliamentary system-in which the executive is indirectly elected by the legislature-means the choice of executive is less of a simplistic binary, and coalition negotiations involving third party/independent legislators in the lower house can be decisive in determining that outcome in close elections; (3) twelve senators per a state, six elected at a time in an ordinary election, gives more opportunities for minor parties to get into our Senate - of course, 12 senators per a state is feasible when you only have six states (plus four more to represent our two self-governing territories), with 50 states it would produce 600 Senators
Currently minimum 4% of formal first preference votes, which gets you $3.499 per a first preference vote (indexed to inflation every six months)
Then you automatically get paid the first $12,791, and the rest of the funding is by reimbursement of substantiated election expenses.
This is per a candidate (lower house) or per a group (upper house). And this is just federal elections - state election funding is up to each state, but I believe the states have broadly similar funding systems.
Note the US also has public financing for presidential campaigns, which is available to minor parties once they get 5% or more of the vote. But in the 2024 election, Jill Stein (Green Party) came third on 0.56% of the popular vote. The only third party to ever qualify for general election public funding was the Reform Party due to Ross Perot getting 18.9% in the 1992 election and 8.4% in the 1996 election. There is also FEC funding for primary campaigns, and I believe that’s easier for third parties to access, but also less impactful.
And the Parliamentary Joint Committee on Intelligence and Security definitely gave the literal thousands of submissions due consultation before recommending the original, un-split bill pass.
Combined with the quirk in Australia’s preferential voting system that enable a government to form despite 65% of voters having voted 1 for something else.
As a result, Australia tends to end up with governments formed by the runner up, because no one party actually ‘won’ as such.
I can think of an exaggerated scenario though in which that sounds reasonable depending on the goal:
say preferences are 1 (low) to 5 (high).
suppose 65% of the population ranked candidate A at 5 and B at 4, and the other 35% ranked A at 1 and B at 5. the majority doesn't get their favorite choice, but they do get an outcome they're happy with, and the minority doesn't have a horrible outcome. Exaggerated, but I don't think situations like this are unrealistic.
18.9% as recently as 1992. I predict we will have a similar viable third party showing sometime in the next few elections due to the radical shift in the party system that AI is causing as we speak. I really hope Yang Gang can rebuild itself and try again, maybe without #MATH.
"The reasonable man adapts himself to the world; the unreasonable one persists in trying to adapt the world to himself. Therefore all progress depends on the unreasonable man" - George Bernard Shaw
In the US, there are tremendous structural barriers for third parties. They exist, it is just extremely difficult for them.
The centralization of power of each of the two dominant parties nationally at the expense of a more decentralized parties with strong state variability as in the past, makes it even more difficult for third parties to gain traction against all that coordination.
Perot had the best chance, but managed to blow it by bowing out and then back in.
I do think you are right, that times of great dissatisfaction are rare openings for third party candidates, if someone special enough appears. 2020 would have been a great election for that - but an inspiring third party candidate can't be manufactured on demand.
People have a choice between being rational and optimizing the alignment between the outcome and their preferences, or being irrational and doing something else, like not voting, spoiling their ballot, voting for a probabilistically infeasible candidate, voting "on principle", "sending a message", etc.
I don’t recall the circumstances under which Morrison ended up Prime Minister.
Like most Australians, I’m in denial any of that episode ever happened.
But, using the current circumstances as an example, Australia has a voting system that enables a party to form government even though 65% of voting Australia’s didn’t vote for that party as their first preference.
If the other party and some of the smaller parties could have got their shit together Australia could have a slightly different flavour of complete fucking disaster of a Government, rather than whatever the fuck Anthony Albanese thinks he’s trying to be.
Then there’s Susan Ley. The least preferred leader of the two major parties in a generation.
Hmm. Actually, I think the suggestion of a law puts this whole thing on bad footing where we need to draw an otherwise unnecessary line (to denote where this type of rhetoric should be legal). I suspect XorNot just put the line there because the idea that true statements should be illegal just seems silly.
Really it just ought to be a thing that we identify as a thought-terminating cliche. No laws needed, let’s just not fall for a lazy trick. Whether or not it is true that lots of people agreed, that isn’t a good argument that they are right.
The case of Nixon really brings that out. The “Silent Majority” was used to refer to people who didn’t protest the Vietnam War. Of course, in retrospect the Vietnam War was pretty bad. Arguing that it was secretly popular should have not been accepted as a substitute for an argument that it was good.
> We really need a rule in politics which bans you (if you're an elected representative) from stating anything about the beliefs of the electorate without reference to a poll of the population of adequate size and quality.
Except that assumes polls are a good and accurate way to learn the "beliefs of the electorate," which is not true. Not everyone takes polls, not every belief can be expressed in a multiple-choice form, little subtleties in phrasing and order can greatly bias the outcome of a poll, etc.
I don't think it's a good idea to require speech be filtered through such an expensive and imperfect technology.
err... how Bitcoin works, or how the speculative bubble around cryptocurrencies circa 2019-2021 worked?
Bitcoin is actually kind of useful for some niche use cases - namely illegal transactions, like buying drugs online (Silk Road, for example), and occasionally for international money transfers - my French father once paid an Argentinian architect in Bitcoin, because it was the easiest way to transfer the money due to details about money transfer between those countries which I am completely unaware of.
The Bitcoin bubble, like all bubbles since the Dutch tulip bubble in the 1600s, did follow a somewhat similar "well everyone things this thing is much more valuable than it is worth, if I buy some now the price will keep going on and I can dump it on some sucker" path, however.
> Bitcoin is actually kind of useful for some niche use cases - namely illegal transactions, like buying drugs online (Silk Road, for example),
For the record - the illegal transactions were thought to be advantaged by crypto like BTC because it was assumed to be impossible to trace the people engaged in the transaction, however the opposite is true, public blockchains register every transaction a given wallet has made, which has been used by Law Enforcement Agencies(LEA) to prosecute people (and made it easier in some cases).
> and occasionally for international money transfers - my French father once paid an Argentinian architect in Bitcoin, because it was the easiest way to transfer the money due to details about money transfer between those countries which I am completely unaware of.
There are remittance companies that deal in local currencies that tend to make this "easier" - crypto works for this WHEN you can exchange the crypto for the currencies you have and want, which is, in effect, the same.
Anonymity/untraceability was not the primary reason for using BTC towards black/grey markets. Bitcoin can be used pseudo anonymously , and the fact is you simply cannot send money to your grey market counterparty via any method but cash without it being flagged/canceled, and if you can't send cash (which has its own problems), bitcoin is the only option.
Mining rigs have a finite lifespan & the places that make them in large enough quantities will stop making new ones if a more profitable product line, e.g. AI accelerators, becomes available. I'm sure making mining rigs will remain profitable for a while longer but the memory shortages are making it obvious that most production capacity is now going towards AI data centers & if that trend continues then hashing capacity will continue diminishing b/c the electricity cost & hardware replenishment will outpace mining rewards.
Bitcoin was always a dead end. It might survive for a while longer but its demise is inevitable.
Refined 1.01 authoritarian form: Everybody knows you didn't win, and everybody knows the sentiment is universal... But everyone maintains the same outward facade that you won, because it's become a habit and because dissenters seem to have "accidents" falling out of high windows.
V 1.02: Everybody knows you didn't win, and everybody knows the sentiment is universal... But everyone maintains the same outward facade that you won, because they believe that the others believe that you have enough power to crush the dissent. The moment this belief fades, you fall.
Ontological version is even more interesting, especially if we're talking about a singularity (which may be in the past rather than future if you believe in simulation argument).
Crude form: winning is metaphysically guaranteed because it probably happened or probably will
Refined: It's metaphysically impossible to tell whether or not it has or will have happened, so the distinction is meaningless, it has happened.
So... I guess Weir's Egg falls out of that particular line of thought?
The refined form is unstable, a hair from an objective reality observation fluke collapsing it.
The system that persists in practice is where everybody knows how things are, but still everybody pleads to a fictional status quo, because if they did not, the others would obliterate them.
You ever get into logic puzzles? The sort where the asker has to specify that everybody in the puzzle will act in a "perfectly logical" way. This feels like that sort of logic.
Its the classic interrogation technique; "we're not here to debate whether your guilty or innocent, we have all the evidence we need to prove your guilt, we just want to know why". Not sure if it makes it any different though that the interrogator knows they are lying
Isn't talking about "here’s how LLMs actually work" in this context a bit like saying "a human can't be a relevant to X because a brain is only a set of molecules, neurons, synapses"?
Or even "this book won't have any effect on the world because it's only a collection of letters, see here, black ink on paper, that is what is IS, it can't DO anything"...
Saying LLM is a statistical prediction engine of the next token is IMO sort of confusing what it is with the medium it is expressed in/built of.
For instance those small experiments that train a network on addition problems mentioned in a sibling post. The weights end up forming an addition machine. An addition machine is what it is, that is the emergent behavior. The machine learning weights is just the medium it is expressed in.
What's interesting about LLM is such emergent behavior. Yes, it's statistical prediction of likely next tokens, but when training weights for that it might well have a side-effect of wiring up some kind of "intelligence" (for reasonable everyday definitions of the word "intelligence", such as programming as good as a median programmer). We don't really know this yet.
Its pretty clear that the problem of solving AI is software, I don't think anyone would disagree.
But that problem is MUCH MUCH MUCH harder than people make it out to be.
For example, you can reliably train an LLM to produce accurate output of assembly code that can fit into a context window. However, lets say you give it a Terabyte of assembly code - it won't be able to produce correct output as it will run out of context.
You can get around that with agentic frameworks, but all of those right now are manually coded.
So how do you train an LLM to correctly take any length of assembly code and produce the correct result? The only way is to essentially train the structure of the neurons inside of it behave like a computer, but the problem is that you can't do back-propagation with discrete zero and 1 values unless you explicitly code in the architecture for a cpu inside. So obviously, error correction with inputs/outputs is not the way we get to intelligence.
It may be that the answer is pretty much a stochastic search where you spin up x instances of trillion parameter nets and make them operate in environments with some form of genetic algorithm, until you get something that behaves like a Human, and any shortcutting to this is not really possible because of essentially chaotic effects.
> For example, you can reliably train an LLM to produce accurate output of assembly code that can fit into a context window. However, lets say you give it a Terabyte of assembly code - it won't be able to produce correct output as it will run out of context.
Fascinating reasoning. Should we conclude that humans are also incapable of intelligence? I don't know any human who can fit a terabyte of assembly into their context window.
Any human who would try to do this is probably a special case. A reasonable person would break it down into sub-problems and create interfaces to glue them back together...a reasonable AI might do that as well.
I can tell you from first hand experience that claude+ghidra mcp is very good at understanding firmware, labeling functions, finding buffer overflows, patching in custom functionality
On the other hand the average human has a context window of 2.5 petabytes that's streaming inference 24/7 while consuming the energy equivalent of a couple sandwiches per day. Oh and can actually remember things.
Citation desperately needed? Last I checked, humans could not hold the entirety of Wikipedia in working memory, and that's a mere 24 GB. Our GPU might handle "2.5 petabytes" but we're not writing all that to disc - in fact, most people have terrible memory of basically everything they see and do. A one-trick visual-processing pony is hardly proof of intelligence.
>So obviously, error correction with inputs/outputs is not the way we get to intelligence.
This doesn't seem to follow at all let alone obviously? Humans are able to reason through code without having to become a completely discrete computer, but probably can't reason through any length of assembly code, so why is that requirement necessary and how have you shown LLMs can't achieve human levels of competence on this kind of task?
> but probably can't reason through any length of assembly code
Uh what? You can sit there step by step and execute assembly code, writing things down on a piece of paper and get the correct final result. The limits are things like attention span, which is separate from intelligence.
Human brains operate continuously, with multiple parts being active at once, with weight adjustment done in real time both in the style of backpropagation, and real time updates for things like "memory". How do you train an LLM to behave like that?
So humans can get pen and paper and sleep and rest, but LLMs can't get files and context resets?
Give the LLM the ability to use a tool that looks up instructions and records instructions from/to files, instead of holding it in context window, and to actively manage its context (write a new context and start fresh), and I think you would find the LLM could probably do it about as reliable as a human?
Context is basically "short term memory". Why do you set the bar higher for LLMs than for humans?
Couldn't you periodically re-train it on what it's already done and use the context window for more short term memory? That's kind of what humans do - we can't learn a huge amount in short time but can accumulate a lot slowly (school, experience).
A major obstacle is that they don't learn from their users, probably because of privacy. But imagine if your context window was shared with other people, and/or all your conversations were used to train it. It would get to know individuals and perhaps treat them differently, or maybe even manipulate how they interact with each other so it becomes like a giant Jeffrey Epstein.
There is more than molecules, neurons and synapses. They are made from lower level stuff that we have no idea about (well, we do in this instance but you get the point). They are just higher level things that are useful to explain and understand some things but don't describe or capture the whole thing. For that you would need to go to lower and lower level and so far it seems they go on infinitely. Currently we are stuck at the quantum level, that doesn't mean it's the final level.
OTOH, an LLM is just a token prediction engine. It fully and completely covers it. There is no lower level secrets hidden in the design nobody understands, because it could not have been created if there was. The fact that the output can be surprising is not evidence of anything, we have always had surprising outputs like funny bugs or unexpected features. Using the word "emergence" for this is just deceitful.
This algorithm has fundamental limitations and they have not been getting better, if you look closely. For instance you could vibe code a C compiler now, but it's 80% there, cute trick but not usable in real world. Just like anything, it cannot be economically vibe coded to 100%. They are not going back and vibe coding the previous simpler projects to 100% with "improved" models. Instead they are just vibe coding something bigger to 80%. This is not an improvement in limitations, it is actually communicating between the lines that the limitations cannot be overcome.
You're putting a bunch of words in the parent commenter's mouth, and arguing against a strawman.
In this context, "here’s how LLMs actually work" is what allows someone to have an informed opinion on whether a singularity is coming or not. If you don't understand how they work, then any company trying to sell their AI, or any random person on the Internet, can easily convince you that a singularity is coming without any evidence.
This is separate from directly answering the question "is a singularity coming?"
One says "well, it was built as a bunch of pieces, so it can only do the thing the pieces can do", which is reasonably dismissed by noting that basically the only people predicting current LLM capabilities are the ones who are remarkably worried about a singularity occurring.
The other says "we can evaluate capabilities and notice that LLMs keep gaining new features at an exponential, now bordering into hyperbolic rate", like the OP link. And those people are also fairly worried about the singularity occurring.
So mainly you get people using "here's how LLMs actually work" to argue against the Singularity if-and-only-if they are also the ones arguing that LLMs can't do the things that they can provably do, today, or are otherwise making arguments that also declare humans aren't capable of intelligence / reasoning / etc..
False dichotomy. One can believe that LLMs are capable of more than their constituent parts without necessarily believing that their real-world utility is growing at a hyperbolic rate.
Fair - I meant there's two major clusters in the mainstream debate, but like all debates there's obviously a few people off in all sorts of other positions.
The problem that I see with this analysis is that the author is trying to describe a curve in data that is kind of a punctuated equilibrium on steroids. Given that the takeoff event is when systems are capable of recursive self-improvement and that event will be a huge inflection point in the curve, it feels like they are trying to predict the second half of a biphasic graph based on data from the first half, when the second half is distinctly different than the first.
Thinking about how this might work, the slope of the first half does not predict the inflection point. Once the threshold for recursive self-improvement gets crossed, the curve changes drastically. From what I have heard MAYBE we are just starting to see the glimmers of possible recursive self-improvement in GPT-5.3 / Opus 4.6. If so, this discussion, while interesting, is trying to predict the new curve based on a single relevant data point.
> “here’s why replacing or suggesting the replacement of human labor prior to reforming society into one that does not predicate survival on continued employment and wages is very bad”
And there are plenty of people that take issue with that too.
Unfortunately they're not the ones paying the price. And... stock options.
If you replace "taxes" with more general "investment", it's everywhere. A good example is Amazon that has reworked itself from an online bookstore into a global supplier of everything by ruthlessly reinvesting the profits.
Taxes don't usually work as efficiently because the state is usually a much more sloppy investor. But it's far from hopeless, see DARPA.
If you're looking for periods of high taxes and growing prosperity, 1950s in the US is a popular example. It's not a great example though, because the US was the principal winner of WWII, the only large industrial country relatively unscathed by it.
tells the compelling story that the Mellon family teamed up with the steelworker's union to use protectionism to protect the American steel industry's investments in obsolete open hearth steel furnaces that couldn't compete on a fair market with the basic oxygen furnace process adopted by countries that had their obsolete furnaces blown up. The rest of US industry, such as our car industry, were dragged down by this because they were using expensive and inferior materials. I think this book had a huge impact in terms of convincing policymakers everywhere that tariffs are bad.
Funny the Mellon family went on to further political mischief
Ha, we gutted our manufacturing base, so if we bring it back it will now be state of the art! Not sure if that will work out for us, but hey their is some precedence.
The dollar became the world's reserve currency because the idea of Bancor lost to it. Thus subjecting the US to the Triffin dilemma which made the US capital markets benefit at the expense of a hugely underappreciated incentive to offshore manufacturing.
You can't onshore manufacturing and have a dollar reserve currency. The only question then is, Are you willing to de-dollarize to bring back manufacturing jobs?
This isn't a rhetorical question if the answer is yes, great, let's get moving. But if the answer is no, sorry, dollarization and its effects will continue to persist.
I'll take a look at that story later. I'm curious though, why is US metallurgy consistently top-notch if the processes are inferior? When I use wrenches, bicycle frames, etc from most other countries I have no end of troubles with weld delamination, stress fractures compounding into catastrophic failures, and whatnot, even including enormous wrenches just snapping in half with forces far below what something a tenth the size with American steel could handle.
> I'm curious though, why is US metallurgy consistently top-notch if the processes are inferior?
I really wonder what you're comparing with.
Try some quality surgical steel from Sweden, Japan or Germany and you'll come away impressed. China is still not quite there but they are improving rapidly, Korea is already there and poised to improve further.
Metal buyers all over the globe are turning away from the US because of the effects of the silly tariffs but they were not going there because the quality, but because of the price.
The US could easily catch up if they wanted to but the domestic market just isn't large enough.
And as for actual metallurgy knowledge I think russia still has an edge, they always were good when it came down to materials science, though they're sacrificing all of that now for very little gain.
There are people making top quality steel in the US today by modern methods but it wasn't like the new replaced the old, the old mostly disappeared and we got a little bit of the new.
Yes, I should have been more clear there: they could catch up in volume but it will require a different mindset if they want to become a net exporter of such items.
> the state is usually a much more sloppy investor
I don’t find this to be true
The state invests in important things that have 2nd and 3rd order positive benefit but aren’t immediately profitable. Money in a food bank is a “lost” investment.
Alternatively the state plays power games and gets a little too attached to its military toys.
State agencies are often good at choosing right long-term targets. State agencies are often bad at the actual procurement, because of the pork-barrelling and red tape. E.g. both private companies and NASA agree that spaceflight is a worthy target, but NASA ends up with the Space Shuttle (a nice design ruined by various committees) and SLS, while private companies come up with Falcon-9.
Sounds like a false dichotomy. NASA got all these different subcontractors to feed, in all these different states and they explicitly gutted MOL and dynasoar and all the air force projects that needed weird orbits and reentry trajectories so the space shuttle became a huge compromise. Perverse incentives and all that. It's not state organizations per se but rather non-profits that need to have a clear goal that creates capabilities, tools and utilities that act as multipliers for everyone. A pretty big cooperative. Like, I dunno , what societies are supposed to exist for.
But DoD with its weird requirements, and the Congress with its power to finance the project and the desire to bring jobs from it to every state, and the rules of contracting that NASA must follow, are all also part of the state, the way the state ultimately works.
Yeah, our use of our military force provides some of the most obvious cases of "bad investment". Vietnam, Iraq, etc
And there are many others that might've been a positive investment from a strictly financial perspective, but not from a moral one: see Banana Republics and all those times the CIA backed military juntas.
I would argue that those bad investments (such an understatement!) were clearly lobbied for by the military-industrial complex. So yes, the state dropped the ball, big, on those. But that was because the private sector pushed for it, probably also in a big way. I would say that, even though the politicians were ultimately responsible for those calamities, the CEOs who greatly enritched themselves from them are absolutely to be blamed, too.
> Taxes don't usually work as efficiently because the state is usually a much more sloppy investor. But it's far from hopeless, see DARPA.
Be careful. The data does not confirm that narrative. You mentioned the 1950s, which is a poignant example of reality conflicting with sponsored narrative. Pre WOII, the wealthy class orbiting the monopolists, and by extension their installed politicians, had no other ideas than to keep lowering taxes for the rich on and on, even if it only deepened the endless economic crisis. Many of them had fallen in the trap of believing their own narratives, something we know as the Cult of Wealth.
Meanwhile, average Americans lived on food stamps. Politically deadlocked in quasi-religious ideas of "bad governments versus wise business men", America kept falling deeper. Meanwhile, with just 175,000 serving on active duty, the U.S. Army was the 18th biggest in the world[1], poorly equipped, poorly trained. Right wing isolationism had brought the country in a precarious position. Then two things happened. Roosevelt and WOII.
In a unique moment, the state took matters in their own hands. The sheer excellence in planning, efficiency, speed and execution of the state baffled the republicans, putting the oligarchic model of the economy to shame. The economy grew tremendously as well, something the oligarchy could not pull of. It is not well-known that WOII depended largely on state-operated industries, because the former class quickly understood how much the state's performance threatened their narratives. So they invested in disinformation campaigns, claiming the efforts and achievements of the government as their own.
BTW the New Deal tried central planning and quickly rejected it. I'd say that the intense application of the antitrust law in the late 1930s was a key factor that helped end the Great Depression. The war, and wartime government powers, were also key: the amount of federal government overreach and and reforms do not compare to what e.g. the second Trump administration has attempted. It was mostly done by people who got their positions in the administration more due to merit and care about the country than loyalty, and it showed.
The post-war era, under Truman and Eisenhower administrations, reaped the benefits of the US being the wealthiest and most intact winner of WWII. At that time, the highest income tax rate bracket was 91%, but the effective rate was below 50%.
> It's not a great example though, because the US was the principal winner of WWII, the only large industrial country relatively unscathed by it.
The US is also shaping up to be the principal winner in Artificial Intelligence.
If, like everyone is postulating, this has the same transformative impact to Robotics as it does to software, we're probably looking at prosperity that will make the 1950s look like table stakes.
Early on in the AI boom NVidia was highly valued as it was seen as the shovel-maker for research and development. It certainly was instrumental early on but now there are a few viable options for training hardware - and, to me at least, it's unclear whether training hardware is actually the critical infrastructure or if it will be something like power capacity (which the US is lagging behind significantly in), education, or even cooling efficiency.
I think it's extremely early to try and call who the principal winner will be especially with all the global shifts happening.
> The US is also shaping up to be the principal winner in Artificial Intelligence.
There is no early mover advantage in AI in the same way that there was in all the other industries. That's the one thing that AI proponents in general seem not to have clued in to.
What will happen is that it eventually drags everything down because it takes the value out of the bulk of the service and knowledge economies. So you'll get places that are 'ahead' in the disruption. But the bottom will fall out of the revenue streams, which is one of the reasons these companies are all completely panicked and are wrecking the products that they had to stuff AI in there in every way possible hoping that one of them will take.
Model training is only an edge in a world where free models do not exist, once those are 'good enough' good luck with your AI and your rapidly outdated hardware.
The typical investors horizon is short, but not that short.
Flying a drone around is easy. Identifying who is on the in group and out group and then moving them is the hard part.
I’m not sure you have really thought out what the drone part is meant to do. Militaries gave outgunned populaces for decades at this point. You don’t need drones to kill civilians.
It's actually quite easy. Whoever isn't in the bunker is the outgroup. You only needed to tell people apart when you needed some meatware to man the factories and work the fields.
Militaries can side with the crowd, or more likely decide to keep the power for themselves.
Yeah ruling juntas do need to "man the fields & factories" (1st order meatware), in order to produce and maintain those drones. Or nukes, or whatever "deciding factor beyond numbers" put them in power.
But they also need 2nd order meatware to support that 1st order: teachers, doctors, merchants… You need scientists to advance your technology against other militaries… You need leaders (3rd order) to keep the first two populations quiet and productive since that turns out to be more cost-effective than fear control through extermination…
Hell you need a certain level of genetic diversity so your own kids don't come out weird.
Give evolution a little more credit. The required number of humans for the in-group to be self-sustainable is definitely not billions, plus it's been shrinking with automation. But we are where we are for a reason – lots of alternative arrangements have been tried over millenia and found wanting.
"Keep my bunker + my drone factory and some farmers, kill the rest" leaves rulers with terrible quality of life (bad) and the next-door-junta taking over pretty quickly (also bad). It is a self-defeating, poor long-term strategy.
Automation tips the power balance further: fewer humans needed, more local autonomy. Which is, I suspect, why the ruling class are so terribly excited about AI, more so than some market valuations. Fewer pesky humans across all levels. Genetic diversity of bloodline remains the primary concern (unless you manage to live forever, which happens to be another evergreen of power ghouls).
Hint: The answer for the government is, it's never enough. "little bit of taxes" is never what we had.
Seriously though, I wouldn't mind "little bit of taxes" if there were guaranteed ways to stop funding something when it's a failed experiment, which is difficult in government. Because "a little bit more" is always wanted.
Every possible example of “progress” have either an individual or a state power purpose behind it
there is only one possible “egalitarian” forward looking investments that paid off for everybody
I think the only exception to this is vaccines…and you saw how all that worked during Covid
Everything else from the semiconductor to the vacuum cleaner the automobile airplanes steam engines I don’t care what it is you pick something it was developed in order to give a small group and advantage over all the other groups it is always been this case it will always be this case because fundamentally at the root nature of humanity they do not care about the externalities- good or bad
COVID has cured me (hah!) of the notion that humanity will be able to pull together when faced with a common enemy. That means global warming or the next pandemic are going to happen and we will not be able to stop it from happening because a solid percentage can't wait to jump off the ledge, and they'll push you off too.
I find it interesting that this is the conclusion you draw from this. I won’t go into a discussion on the efficacy of the various mandates and policies in reducing spread of the disease. Rather, I think it’s worth pointing out that a significant portion of the proponents of these policies likely supported them not because of a desire to follow the authority but because they sincerely believed that a (for them) relatively small sacrifice in personal freedom could lead to improved outcomes for their fellow humans. For them, it was never about blindly following authority or virtue signalling. It was only ever about doing what they perceived as the right thing to do.
So if the arguments are rooted in medical reasons, it's okay to be inhumane? Nazi propaganda argued that getting rid of Jews helped prevent the spread of diseases, because we all know that Jews are disease carriers. See how slippery the slope is here? Certainly you have seen the MAGA folks point out the measles outbreaks are coming from illegal immigrants, right?
I am quite sure that people felt justified in their reasoning for their behavior. That just shows how effective the propaganda was, how easy it is to get people to fall in line. If it was a matter of voluntary self sacrifice of personal freedoms, I wouldn't have made this comment. People decided to demonize anyone who did not agree with the "medical authority", especially doctors or researchers that did not tow the party line. They ruined careers, made people feel awful, and online the behavior was worse because how easy it was to pile on. Over stuff that is still to this day not very clear cut what the optimal strategy is for dealing with infectious disease.
COVID restrictions were public health, an overriding concern listed in the US Constitution as general welfare as a reason for the US government to exist at all.
Yea, closing beaches and parks is on par with the Nazis did to the Jews.
The Covid measures were also totally targeted at certain groups of people with immutable characteristics and not at people who actively wanted to spread disease.
How are people like you still making arguments like this in 2026? Were you also one of the people claiming we’d all be dead in a year from the vaccines?
It is so easy to critique the response in hindsight. Or at the time.
But critiques like that ignore uncertainty, risk, and unavoidably getting it "wrong" (on any and all dimensions), no matter what anyone did.
With a new virus successfully circumnavigating the globe in a very short period of time, with billions of potential brand new hosts to infect and adapt within, and no way to know ahead of time how virulent and deadly it could quickly evolve to be, the only sane response is to treat it as extremely high risk.
There is no book for that. Nobody here or anywhere knows the "right" response to a rapidly spreading (and killing) virus, unresponsive to current remedies. Because it is impossible to know ahead of time.
If you actually have an answer for that, you need to write that book.
And take into account, that a lot of people involved in the last response, are very cognizant that we/they can learn from what worked, what didn't, etc. That is the valuable kind of 20-20 vision.
A lot of at-risk people made it to the vaccines before getting COVID. The ones I know are very happy about everything that reduced their risk. They are happy not to have died, despite those who wanted to let the disease to "take its natural course".
And those that died, including people I know, might argue we could have done more, acted as a better team. But they don't get to.
No un-nuanced view of the situation has merit.
The most significant thing we learned: a lot of humanity is preparing to be a problem if the next pandemic proves ultimately deadlier. A lot of humanity doesn't understand risk, and doesn't care, if doing so requires cooperative efforts from individuals.
It's the same people who don't even notice that we don't talk about acid rain anymore, because we solved it with government regulation for pretty cheap.
They even indignantly mention the Ozone layer, insisting that "Look, liberals told you to care but its not a problem anymore", ignorant entirely of the immense global effort to fix that.
"Nazi", "Fascist", etc are words you can use to lose any debate instantly no matter what your politics are.
I think the sane version of this is that Gen Z didn't just lose its education, it lost its socialization. I know someone who works in administration of my Uni who tracks general well being of students who said they were expecting it to bounce back after the pandemic and they've found it hasn't. My son reports if you go to any kind of public event be it a sewing club or a music festival people 18-35 are completely absent. My wife didn't believe him but she went to a few events and found he was right.
You can blame screens or other trends that were going on before the pandemic, but the pandemic locked it in. At the rate we're going if Gen Z doesn't turn it around in 10 years there will not be a Gen Z+2.
So the argument that pandemic policy added a few years to elderly lives at the expensive of the young and the children that they might have had is salient in my book -- I had to block a friend of mine on Facebook who hasn't wanted to talk about anything but masks and long COVID since 2021.
Never seen the attempt by governments to contain a global pandemic that killed millions and threatened to overwhelm healthcare compared to Nazism before, but why should I be surprised? Explains a lot about the sorry state of modern politics.
Ok here: Everything from the semiconductor through the vacuum cleaner, automobile, airplanes and steam engines was developed to give a small group an advantage over all the other groups. It has always been the case, it will always be the case.
Fundamentally, at the root nature of humanity, humans do not care about the externalities, either good or bad.
That's a slightly odd way of looking at it. I'm guessing the people developing airplanes or whatever thought of a number of things including - hey this would be cool to do - and - maybe we can make some money - and - maybe this will help people travel - and - maybe it'll impress the girls - and probably some other things too. At least that's roughly how I've thought when I make stuff, never this will give a small group an advantage.
Vacuum cleaner -> sell appliances -> sell electric motors
But there was a clear advantage in quality of life for a lot of people too.
Automobile -> part of industrialization of transport -> faster transport, faster world
Arguably also a big increase in quality of life but it didn't scale that well and has also reduced the quality of life. If all that money had gone into public transport then that would likely have been a lot better.
Airplanes -> yes, definitely, but they were also clearly seen as an advantage in war, in fact that was always a major driver behind inventions.
Steam engine -> the mother of all prime movers and the beginnings of the fossil fuel debacle (coal).
Definitely a quality of life change but also the cause of the bigger problems we are suffering from today.
The 'coffin corner' (one of my hobby horses) is a real danger, we have, as a society, achieved a certain velocity, if we slow down too much we will crash, if we speed up the plane will come apart. Managing these transitions is extremely delicate work and it does not look as though 'delicate' is in the vocabulary of a lot of people in the driving seats.
This is where the concept of trickle down economics came from though and we know that that’s not actually accurate
I used to hear about this with respect to how fun funding NASA would get us more inventions because they funded Velcro
No it’s simply that there was a positive temporary externality for some subset of groups but the primary long term benefit went to the controller of the capital
The people utilizing them were marginally involved because they were only given the options that capital produced for them
> ”when they don’t even understand that the computer is the box attached to the monitor, not the monitor itself”
Laughed out loud at that - and cried a little.
I have had trouble explaining people: “No! don’t use your email password! This is not your email you are logging in to, your email address is a username for this other service. Don’t give them your email password!”
> whether the singularity actually happens or not is irrelevant so much as whether enough people believe it will happen and act accordingly.
I disagree. If the singularity doesn't happen, then what people do or don't believe matters a lot. If the singularity does happen, then it hardly matters what people do or don't believe (edit: about whether or not the singularity will happen).
Which would also mean the accelerationists are potentially putting everyone at risk. I'd think a soft takeoff decades in the future would give us a much better chance of building the necessary safeguards and reorganizing society accordingly.
Decades from now. Society is nowhere near ready for a singularity. The AI we have now, as far as it has come, is still a tool for humans to use. It's more Augmented Intelligence than AGI.
A hard takeoff would be the tool bootstrapping itself into an autonomous self-improving ASI in a short amount of time.
And I read Kurzweil years ago too. He thought reverse engineering the human brain once the hardware was powerful enough would together give us the singularity in 2045. And the Turing Test would have been passed by 2029, but seems like LLMs have already accomplished this.
20% of the human population still is not using the internet
Imagine you’re 70 years old, in rural North Carolina, sitting on your porch wondering why your house has a sheet of ice on it that’s never happened before. Now your already weak soybean harvest that year yields only 30%
Meanwhile your 30-year-old neighbor just had a productive soybean harvest because they covered their crops prior to the freeze based on using the Internet for weather forecasting
That trivial variation between people who utilize information technology to improve their survivability has been happening for the last few hundred years unabated.
God Roko's Basilisk is the most boring AI risk to catch the public consciousness. It's just Pascal's wager all over again, with the exact same rebuttal.
The culture that brought you "speedrunning computer science with JavaScript" and "speedrunning exploitative, extractive capitalism" is back with their new banger "speedrunning philosophy". Nuke it from orbit; save humanity.
> prior to reforming society into one that does not predicate survival on continued employment and wages
There's no way that'll happen. The entire history of humanity is 99% reacting to things rather than proactively preventing things or adjusting in advance, especially at the societal level. You would need a pretty strong technocracy or dictatorship in charge to do otherwise.
Ugh, GBNews, outrage fodder for idiots and the elderly with no ability to navigate the modern information landscape.
You can tell it's watched almost exclusively by old people because all the ads on the channel are for those funeral pre-pay services or retirement homes.
You would need a new sense of self and a life free of fear, raising children where they can truly be anything they like and teach their own kids how to find meaning in a life lived well. "Best I can do is treefiddy" though..
> whether the singularity actually happens or not is irrelevant so much as whether enough people believe it will happen and act accordingly.
We've already been here in the 1980s.
The tech industry needs to cultivate people who are interested in the real capabilities and the nuance around that, and eject the set of people who am to turn the tech industry into a "you don't even need a product" warmed-over acolytes of Tony Robbins.
All the discussion of investment and economics can be better informed by perusing the economic data in Rise and Fall of American Growth. Robert Gordon's empirical finding is that American productivity compounded astonishingly from 1870-1970, but has been stuck at a very low growth rate since then.
It's hard to square with the computer revolution, but my take post-70s is "net creation minus creative destruction" was large but spread out over more decades. Whereas technologies like: electrification, autos, mass production, telephone, refrigeration, fertilizers, pharmaceuticals, these things produced incomparable growth over a century.
So if you were born in the 70s America, your experience of taxes, inflation, prosperity and which policies work, all that can feel heavier than what folks experienced in the prior century. Of course that's in the long run (ie a generation).
I question whether AI tools have great net positive creation minus destruction.
This entire chain of reasoning takes for granted that there won't be a singularity
If you're talking about "reforming society", you are really not getting it. There won't be society, there won't be earth, there won't be anything like what you understand today. If you believe that a singularity will happen, the only rational things to do are to stop it or make sure it somehow does not cause human extinction. "Reforming society" is not meaningful
I thought the Singularity had already happened when the Monkeys used tools to kill the other Monkeys and threw the bone into the sky to become a Space Station.
> It’s honestly why I gave up trying to get folks to look at these things rationally as knowable objects (“here’s how LLMs actually work”)
Here's your own fallacy you fell into - this is important to understand. Neither do you nor me understand "how LLMs actually work" because, well, nobody really does. Not even the scientists who built the (math around) models. So, you can't really use that argument because it would be silly if you thought you know something which rest of the science community doesn't. Actually, there's a whole new field in science developed around our understanding how models actually arrive to answers which they give us. The thing is that we are only the observers of the results made by the experiments we are doing by training those models, and only so it happens that the result of this experiment is something we find plausible, but that doesn't mean we understand it. It's like a physics experiment - we can see that something is behaving in certain way but we don't know to explain it how and why.
Even if interpretability of specific models or features within them is an open area of research, the mechanics of how LLMs work to produce results are observable and well-understood, and methods to understand their fundamental limitations are pretty solid these days as well.
Is there anything to be gained from following a line of reasoning that basically says LLMs are incomprehensible, full stop?
The concept “understand” is rooted in utility. It means “I have built a much simpler model which produces usefully accurate predictions, of the thing or behaviour I seek to ‘understand’”. This utility is “explanatory power”. The model may be in your head, may be math, may be an algorithm or narrative, it may be a methodology with a history of utility. “Greater understanding” is associated with models that are simpler, more essential, more accurate, more useful, cheaper, more decomposed, more composable, more easily communicated or replicated, or more widely applicable.
“Pattern matching”, “next token prediction”, “tensor math” and “gradient descent” or the understanding and application of these by specialists, are not useful models of what LLMs do, any more than “have sex, feed and talk to the resulting artifact for 18 years” is a useful model of human physiology or psychology.
My understanding, and I'm not a specialist, is there are huge and consequential utility gaps in our models of LLMs. So much so, it is reasonable to say we don't yet understand how they work.
>Even if interpretability of specific models or features within them is an open area of research, the mechanics of how LLMs work to produce results are observable and well-understood, and methods to understand their fundamental limitations are pretty solid these days as well.
If you train a transformer on (only) lots and lots of addition pairs, i.e '38393 + 79628 = 118021' and nothing else, the transformer will, during training discover an algorithm for addition and employ it in service of predicting the next token, which in this instance would be the sum of two numbers.
We know this because of tedious interpretability research, the very limited problem space and the fact we knew exactly what to look for.
Alright, let's leave addition aside (SOTA LLMs are after all trained on much more) and think about another question. Any other question at all. How about something like:
"Take a capital letter J and a right parenthesis, ). Take the parenthesis, rotate it counterclockwise 90 degrees, and put it on top of the J. What everyday object does that resemble?"
What algorithm does GPT or Gemini or whatever employ to answer this and similar questions correctly ? It's certainly not the one it learnt for addition. Do you Know ? No. Do the creators at Open AI or Google know ? Not at all. Can you or they find out right now ? Also No.
Let's revisit your statement.
"the mechanics of how LLMs work to produce results are observable and well-understood".
Observable, I'll give you that, but how on earth can you look at the above and sincerely call that 'well-understood' ?
It's pattern matching, likely from typography texts and descriptions of umbrellas. My understanding is that the model can attempt some permutations in its thinking and eventually a permutation's tokens catch enough attention to attempt to solve, and that once it is attending to "everyday object", "arc", and "hook", it will reply with "umbrella".
>It's pattern matching, likely from typography texts and descriptions of umbrellas.
"Pattern matching" is not an explanation of anything, nor does it answer the question I posed. You basically hand waved the problem away in conveniently vague and non-descriptive phrase. Do you think you could publish that in a paper for ext ?
>Why am I confident that it's not actually doing spatial reasoning? At least in the case of Claude Opus 4.6, it also confidently replies "umbrella" even when you tell it to put the parenthesis under the J, with a handy diagram clearly proving itself wrong
I don't know what to tell you but J with the parentheses upside down still resembles an umbrella. To think that a machine would recognize it's just a flipped umbrella and a human wouldn't is amazing, but here we are. It's doubly baffling because Claude quite clearly explains it in your transcript.
>I feel like I have a pretty good intuition of what's happening here based on my understanding of the underlying mathematical mechanics.
Yes I realize that. I'm telling you that you're wrong.
>Do you think you could publish that in a paper for ext ?
You seem to think it's not 'just' tensor arithmetic.
Have you read any of the seminal papers on neutral networks, say?
It's [complex] pattern matching as the parent said.
If you want models to draw composite shapes based on letter forms and typography then you need to train them (or at least fine-tune them) to do that.
I still get opposite (antonym) confusion occasionally in responses to inferences where I expect the training data is relatively lacking.
That said, you claim the parent is wrong. How would you describe LLM models, or generative "AI" models in the confines of a forum post, that demonstrates their error? Happy for you to make reference to academic papers that can aid understanding your position.
>You seem to think it's not 'just' tensor arithmetic.
If I asked you to explain how a car works and you responded with a lecture on metallic bonding in steel, you wouldn’t be saying anything false, but you also wouldn’t be explaining how a car works. You’d be describing an implementation substrate, not a mechanism at the level the question lives at.
Likewise, “it’s tensor arithmetic” is a statement about what the computer physically does, not what computation the model has learned (or how that computation is organized) that makes it behave as it does. It sheds essentially zero light on why the system answers addition correctly, fails on antonyms, hallucinates, generalizes, or forms internal abstractions.
So no: “tensor arithmetic” is not an explanation of LLM behavior in any useful sense. It’s the equivalent of saying “cars move because atoms.”
>It's [complex] pattern matching as the parent said
“Pattern matching”, whether you add [complex] to it or not is not an explanation. It gestures vaguely at “something statistical” without specifying what is matched to what, where, and by what mechanism. If you wrote “it’s complex pattern matching” in the Methods section of a paper, you’d be laughed out of review. It’s a god-of-the-gaps phrase: whenever we don’t know or understand the mechanism, we say “pattern matching” and move on, but make no mistake, it's utterly meaningless and you've managed to say absolutely nothing at all.
And note what this conveniently ignores: modern interpretability work has repeatedly shown that next-token prediction can produce structured internal state that is not well-described as “pattern matching strings”.
Transformers trained on Othello or Chess games (same next token prediction) were demonstrated to have developed internal representations of the rules of the game. When a model predicted the next move in Othello, it wasn't just "pattern matching strings", it had constructed an internal map of the board state you could alter and probe. For Chess, it had even found a way to estimate a player's skill to better predict the next move.
There are other interpretability papers even more interesting than those. Read them, and perhaps you'll understand how little we know.
>That said, you claim the parent is wrong. How would you describe LLM models, or generative "AI" models in the confines of a forum post, that demonstrates their error? Happy for you to make reference to academic papers that can aid understanding your position.
Nobody understands LLMs anywhere near enough to propose a complete theory that explains all their behaviors and failure modes. The people who think they do are the ones who understand them the least.
What we can say:
- LLMs are trained via next-token prediction and, in doing so, are incentivized to discover algorithms, heuristics, and internal world models that compress training data efficiently.
- These learned algorithms are not hand-coded; they are discovered during training in high-dimensional weight space and because of this, they are largely unknown to us.
- Interpretability research shows these models learn task-specific circuits and representations, some interpretable, many not.
- We do not have a unified theory of what algorithms a given model has learned for most tasks, nor do we fully understand how these algorithms compose or interfere.
I made this metaphor from my understanding of your comment.
Imagine we put a kid in a huge library of book who doesn't know how to write/read and knows nothing about what letter means etc. That kid stayed in the library and had a change for X amount time which will be enough to look over all of them.
what this will do is that not like us but somehow this kid managed to create patterns in the books.
After that X amount of time, we asked this Kid a question. "What is the capital of Germany?"
That kid will just have it is on kind of map/pattern to say "Berlin". Or kid might say "Berlin is the capital of the Germany" or "Capital of Germany is Berlin." The issue here is that we do not have the understanding of how this kid came of with the answer or what kind of "understanding" or "mapping" being used to reach this answer.
The other part basically shows we do not fully understand how LLM works is: Ask a very complex question to an AI. Like "explain me the mechanics of quantum theory like I am 8 years old".
1- Everytime, it will create differnt answer. Main point is the same but the letters/words etc would be different. Like the example I give above.There are unlimited type of answer AI can give you.
2- Can anyone in the Earth - a human - without a technology access for have unlimited amount of book/paper to check whatever info he needs - tell us the exact sentence/words will LLM use? No.
Then we do not have fully understand of LLM.
You can create a linear regression model and give it 100 people data and all these 100 people are blue eyed. Then give 101 person and ask it to predict the eye color. You already know the exact answer. It will be %100.
I think what you two are going back and forth on is the heated debate in AI research regarding Emergent Abilities. Specifically, whether models actually develop "sudden" new powers as they scale, or if those jumps are just a mirage caused by how we measure them.
Pro tip: call it a "law of nature" and people will somehow stop pestering you about the why.
I think in a couple decades people will call this the Law of Emergent Intelligence or whatever -- shove sufficient data into a plausible neural network with sufficient compute and things will work out somehow.
On a more serious note, I think the GP fell into an even greater fallacy of believing reductionism is sufficient to dissuade people from ... believing in other things. Sure, we now know how to reduce apparent intelligence into relatively simple matrices (and a huge amount of training data), but that doesn't imply anything about social dynamics or how we should live at all! It's almost like we're asking particle physicists how we should fix the economy or something like that. (Yes, I know we're almost doing that.)
In science these days, the term "Law" is almost never used anymore, the term "Theory" replaced it. E.g Theory of special relativity instead of Law of special relativity.
Agree. I think it is just people have their own simplified mental model how it works. However, there is no reason to believe these simplified mental models are accurate (otherwise we will be here 20-year earlier with HMM models).
The simplest way to stop people from thinking is to have a semi-plausible / "made-me-smart" incorrect mental model of how things work.
> We really have no idea how did ability to have a conversation emerge from predicting the next token.
Maybe you don't. To be clear, this is benefiting massively from hindsight, just as how if I didn't know that combustion engines worked, I probably wouldn't have dreamed up how to make one, but the emergent conversational capabilities from LLMs are pretty obvious. In a massive dataset of human writing, the answer to a question is by far the most common thing to follow a question. A normal conversational reply is the most common thing to follow a conversation opener. While impressive, these things aren't magic.
>In a massive dataset of human writing, the answer to a question is by far the most common thing to follow a question.
No it isn't. Type a question into a base model, one that hasn't been finetuned into being a chatbot, and the predicted continuation will be all sorts of crap, but very often another question, or a framing that positions the original question as rhetorical in order to make a point. Untuned raw language models have an incredible flair for suddenly and unexpectedly shifting context - it might output an answer to your question, then suddenly decide that the entire thing is part of some internet flamewar and generate a completely contradictory answer, complete with insults to the first poster. It's less like talking with an AI and more like opening random pages in Borge's infinite library.
To get a base language model to behave reliably like a chatbot, you have to explicitly feed it "a transcript of a dialogue between a human and an AI chatbot", and allow the language model to imagine what a helpful chatbot would say (and take control during the human parts). The fact that this works - that a mere statistical predictive language model bootstraps into a whole persona merely because you declared that it should, in natural English - well, I still see that as a pretty "magic" trick.
>No it isn't. Type a question into a base model, one that hasn't been finetuned into being a chatbot, and the predicted continuation will be all sorts of crap, but very often another question, or a framing that positions the original question as rhetorical in order to make a point.....
To be fair, only if you pose this question singularly with no proceeding context. If you want the raw LLM to answer your question(s) reliably then you can have the context prepended with other question-answer pairs and it works fine. A raw LLM is already capable of being a chatbot or anything else with the right preceding context.
Right, but that was my point - statistically, answers do not follow questions without some establishing context, and as such, while LLMs are "simply" next word predictors, the chatbots aren't - they are Hofstaderian strange loops that we will into being. The simpler you think language models are, the more that should seem "magic".
They're not simple though. You can understand, in a reductionist sense, the basic principles of how transformers perform function approximation; but that does not grant an intuitive sense of the nature of the specific function they have been trained to approximate, or how they have achieved this approximation. We have little insight into what abstract concepts each of the many billions of parameters map on to. Progress on introspecting these networks has been a lot slower than trial-and-error improvements. So there is a very real sense in which we have no idea how LLMs work, and they are literally "magic black boxes".
No matter how you slice it - if "magic" is a word which can ever be applied to software, LLM chatbots are sure as shit magic.
If such a simplistic explanation was true, LLM's would only be able to answer things that had been asked before, and where at least a 'fuzzy' textual question/answer match was available. This is clearly not the case. In practice you can prompt the LLM with such a large number of constraints, so large that the combinatorial explosion ensures no one asked that before. And you will still get a relevant answer combining all of those. Think combinations of features in a software request - including making some module that fits into your existing system (for which you have provided source) along with a list of requested features. Or questions you form based on a number of life experiences and interests that combined are unique to you. You can switch programming language, human language, writing styles, levels as you wish and discuss it in super esoteric languages or morse code. So are we to believe this answers appear just because there happened to be similar questions in the training data where a suitable answer followed? Even if for the sake of argument we accept this explanation by "proximity of question/answer", it is immediately that this would have to rely on extreme levels of abstraction and mixing and matching going on inside the LLM. And that it is then this process that we need to explain how works, whereas the textual proximity you invoke relies on this rather than explaining it.
I think you're confusing OP for the people who claim that there is zero functional difference between an LLM and a search engine that just parrots stuff already in it. But they never made such a claim. Here, let me try: the simplest explanation for how next token estimation leads to a model that often produces true answers is that for most inputs, the most likely next token is true. Given their size and the way they're trained, LLMs obviously don't just ingest training data like a big archive, they contain something like an abstract representation of tokens and concepts. While not exactly like human knowledge, the network is large and deep enough that LLMs are capable of predicting true statements based on preceding text. This also enables them to answer questions not in their training dataset, although accuracy obviously suffers the further you deviate from known topics. The most likely next token to any question is the true answer, so they essentially ended up being trained to estimate truth.
I'm not saying this is bad or underwhelming, by the way. It's incredible how far people were able to push machine learning with just the knowledge we have now, and how they're still making process. I'm just saying it's not magic. It's not something like an unsolved problem in mathematics.
No one ever made the claim it was magic, not even remotely. Regarding the rest of your commentary: a) The original claim was that LLM's were not understood and are a black box. b) Then someone claims that this is not true, and they know well how LLM's work, it is simply due to questions & answers being in close textual proximity in training data. c) I then claim this is a shallow explanation because you then need to invoke additionally a huge abstraction network - that is a black box, d) you seem to agree with this while at the same time saying I misrepresented "b" - which I don't think I did. They really claimed they understood it and only offered this textual proximity thing.
In general, every attempt at explanation of LLM's that appeal to "[just] predicting next token" is thought terminating and automatically invalid as explanation. Why? Because it is confusing the objective function with the result. It adds exactly zero over saying "I know how a chess engine works, it just predicts the next move and has been trained to predict the next move" or "A talking human just predicts the next word, as it was trained to do". It says zero about how this is done internally in the model. You could have a physical black box predicting the next token, and inside you could have simple frequentist tables or you could have a human brain or you could have an LLM. In all cases you could say the box is predicting the next token and if any training was involved you could say it was trained to predict the next token.
My best friend who has literally written a doctorate on artificial intelligence doesn't. If you do, please write a paper on it, and email it to me. My friend would be thrilled to read it.
I don't know much about this space other than a user of Claude and a Electrical Engineering background...
However reading some Standford study summaries (not the whole thing) and just generally where AI research is now, it's clear that researchers can't deterministically say exactly how the black box works.
So yet gain, HN armchair scientists are no better than any other topic. I love reading comments here, but so many people have opinions on things that aren't well founded.
>In a massive dataset of human writing, the answer to a question is by far the most common thing to follow a question. A normal conversational reply is the most common thing to follow a conversation opener. While impressive, these things aren't magic.
Obviously, that's the objective, but who's to say you'll reach a goal just because you set it ? And more importantly, who's the say you have any idea how the goal has actually been achieved ?
You don't need to think LLMs are magic to understand we have very little idea of what is going on inside the box.
We know exactly what is going on inside the box. The problem isn't knowing what is going on inside the box, the problem is that it's all binary arithmetic & no human being evolved to make sense of binary arithmetic so it seems like magic to you when in reality it's nothing more than a circuit w/ billions of logic gates.
We do not know or understand even a tiny fraction of the algorithms and processes a Large Language Model employs to answer any given question. We simply don't. Ironically, only the people who understand things the least think we do.
Your comment about 'binary arithmetic' and 'billions of logic gates' is just nonsense.
I think the fallacy at hand is more along the lines of "no true scotsman".
You can define understanding to require such detail that nobody can claim it; you can define understanding to be so trivial that everyone can claim it.
"Why does the sun rise?" Is it enough to understand that the Earth revolves around the sun, or do you need to understand quantum gravity?
Good point. OP was saying "no one knows" when in fact plenty of people do know but people also often conflate knowing & understanding w/o realizing that's what they're doing. People who have studied programming, electrical engineering, ultraviolet lithography, quantum mechanics, & so on know what is going on inside the computer but that's different from saying they understand billions of transistors b/c no one really understands billions of transistors even though a single transistor is understood well enough to be manufactured in large enough quantities that almost anyone who wants to can have the equivalent of a supercomputer in their pocket for less than $1k: https://www.youtube.com/watch?v=MiUHjLxm3V0.
Somewhere along the way from one transistor to a few billion human understanding stops but we still know how it was all assembled together to perform boolean arithmetic operations.
With LLMs, The "knowing" you're describing is trivial and doesn't really constitute knowing at all. It's just the physics of the substrate. When people say LLMs are a black box, they aren't talking about the hardware or the fact that it's "math all the way down." They are talking about interpretability.
If I hand you a 175-billion parameter tensor, your 'knowledge' of logic gates doesn't help you explain why a specific circuit within that model represents "the concept of justice" or how it decided to pivot a sentence in a specific direction.
On the other hand, the very professions you cited rely on interpretability. A civil engineer doesn't look at a bridge and dismiss it as "a collection of atoms" unable to go further. They can point to a specific truss and explain exactly how it manages tension and compression, tell you why it could collapse in certain conditions. A software engineer can step through a debugger and tell you why a specific if statement triggered.
We don't even have that much for LLMs so why would you say we have an idea of what's going on ?
It sounds like you're looking for something more than the simple reality that the math is what's going on. It's a complex system that can't simply be debugged through[1], but that doesn't mean it isn't "understood".
This reminds me of Searle's insipid Chinese Room; the rebuttal (which he never had an answer for) is that "the room understands Chinese". It's just not satisfying to someone steeped in cultural traditions that see people as "souls". But the room understands Chinese; the LLM understands language. It is what it is.
[1] Since it's deterministic, it certainly can be debugged through, but you probably don't have the patience to step through trillions of operations. That's not the technology's fault.
Your ultra-reductionism does not not constitute understanding. "Math happens and that somehow leads to a conversational AI" is true, but it is not useful. You cannot use it to answer questions like "how should I prompt the model to achieve <x>". There are many layers of abstraction within the network - important, predictive abstractions - which you have no concept of. It is as useful as asking a particle physicist why your girlfriend left you, because she is made of atoms.
Incidentally, your description of LLMs also describes all software, ever. It's just math, man! That doesn't make you an expert kernel hacker.
>It sounds like you're looking for something more than the simple reality that the math is what's going on.
Train a tiny transformer on addition pairs (i.e i.e '38393 + 79628 = 118021') and it will learn an algorithm for addition to minimize next token error. This is not immediately obvious. You won't be able to just look at the matrix multiplications and see what addition implementation it subscribes to but we know this from tedious interpretability research on the features of the model. See, this addition transformer is an example of a model we do understand.
So those inscrutable matrix multiplications do have underlying meaning and multiple interpretability papers have alluded as much, even if we don't understand it 99% of the time.
I'm very fine with simply saying 'LLMs understand Language' and calling it a day. I don't care for Searle's Chinese Room either. What I'm not going to tell you is that we understand how LLMs understand language.
No one relies on "interpretability" in quantum mechanics. It is famously uninterpretable. In any case, I don't think any further engagement is going to be productive for anyone here so I'm dropping out of this thread. Good luck.
Quantum mechanics has competing interpretations (Copenhagen, Many-Worlds, etc.) about what the math means philosophically, but we still have precise mathematical models that let us predict outcomes and engineer devices.
Again, we lack even this much with LLMs so why say we know how they work ?
Unless I'm missing what you mean by a mile, this isn't true at all. We have infinitely precise models for the outcomes of LLMs because they're digital. We are also able to engineer them pretty effectively.
The ML Research world (so this isn't simply a matter of being ignorant/uninformed) was surprised by the performance of GPT-2 and utterly shocked by GPT-3. Why ? Isn't that strange ? Did the transformer architecture fundamentally change between these releases ? No, it did not at all.
So why ? Because even in 2026, nevermind 18 and 19, the only way to really know exactly how a neural network will perform trained with x data at y scale is to train it and see. No elaborate "laws", no neat equations. Modern Artificial Intelligence is an extremely empirical, trial and error field, with researchers often giving post-hoc rationalizations for architectural decisions. So no, we do not have any precise models that tell us how a LLM will respond to any query. If we did, we wouldn't need to spend months and millions of dollars training them.
We don't have a model for how an LLM that doesn't exist will respond to a specific query. That's different from lacking insight at all. For an LLM that exists it's still hard to interpret but it's very clear what is actually happening. That's better than you often get with quantum physics when there's a bunch of particles and you can't even get a good answer for the math.
And even for potential LLMs, there are some pretty good extrapolations for overall answer quality based on the amount of data and the amount of training.
>We don't have a model for how an LLM that doesn't exist will respond to a specific query.
We don't have a model for a LLM that does exist will respond to a specific query either.
>For an LLM that exists it's still hard to interpret but it's very clear what is actually happening.
No, it's not and I'm getting tired of explaining this. If you think it is, write your paper and get very rich.
>That's better than you often get with quantum physics when there's a bunch of particles and you can't even get a good answer for the math.
You clearly don't understand any of this.
>And even for potential LLMs, there are some pretty good extrapolations for overall answer quality based on the amount of data and the amount of training.
I thought the Hinton talking to Jon Stewart interview gives a rough idea how they work. Hinton got Turing and Nobel prizes for inventing some of the stuff https://youtu.be/jrK3PsD3APk?t=255
> We really have no idea how did ability to have a conversation emerge from predicting the next token.
Uh yes, we do. It works in precisely the same way that you can walk from "here" to "there" by taking a step towards "there", and then repeating. The cognitive dissonance comes when we conflate this way of "having a conversation" (two people converse) and assume that the fact that they produce similar outputs means that they must be "doing the same thing" and it's hard to see how LLMs could be doing this.
Sometimes things seems unbelievable simply because they aren't true.
I wasn't referring to the biomechanical process of walking, I was referring to the process of gradient descent, which is well understood and yes, quite simple.
If that was true, knowing how elementary particles work would give us understanding of the whole universe, in which case no other science would exist. But other sciences do exist, ergo, you're wrong.
"'If I wished,' O'Brien had said, 'I could float off this floor like a soap bubble.' Winston worked it out. 'If he thinks he floats off the floor, and if I simultaneously think I see him do it, then the thing happens'".
I just point to Covid lockdowns and how many people took up hobbies, how many just turned into recluses, how many broke the rules no matter the consequences real or imagined, etc. Humans need something to do. I don’t think it should be work all the time. But we need something to do or we just lose it.
It’s somewhat simplistic, but I find it get the conversation rolling. Then I go “it’s great that we want to replace work but what are we going to do instead and how will we support ourselves?” It’s a real question!
It's true people need something to do, but I don't think the COVID shutdown (lockdowns didn't happen in the U.S. for the most part though they did in other countries) is a good comparison because the entire society was perfused with existential dread and fear of contact with another human being while the death count was rising and rising by thousands a day. It's not a situation that makes for comfortable comparisons because people were losing their damn minds and for good reason.
1. LLMs only serve to reduce the value of your labor to zero over time. They don't need to even be great tools, they just need to be perceived as "equally good" to engineers for C-Suite to lay everyone off, and rehire at 50-25% of previous wages, repeating this cycle over a decade.
2. LLMs will not allow you to join the billionaire class, that wouldn't make sense, as anyone could if that's the case. They erode the technical meritocracy these Tech CEOs worship on podcasts, and youtube, (makes you wonder what are they lying about).
- Your original ideas and that Startup you think is going to save you, isn't going to be worth anything if someone with minimal skills can copy it.
3. People don't want to admit it, but heavy users of LLMs know they're losing something, and there's a deep down feeling that its not the right way to go about things. Its not dissimilar to any guilty dopaminergic crash one gets when taking shortcuts in life.
I used like 1.8bb Anthropic tokens last year, I won't be using it again, I won't be participating in this experiment. I've likely lost years of my life in "potential learning" from the social media experiment, I'm not doing that again. I want to study compilers this year, and I want to do it deeply. I wont be using LLMs.
I've said it simply, much like you, and it comes off as unhinged lunacy. Inviting them to learn themselves has been so much more successful than directed lectures, at least in my own experiments with discourse and teaching.
A lot of us have fallen into the many, many toxic traps of technology these past few decades. We know social media is deliberately engineered to be addictive (like cigarettes and tobacco products before it), we know AI hinders our learning process and shortens our attention spans (like excess sugar intake, or short-form content deluges), and we know that just because something is newer or faster does not mean it's automatically better.
You're on the right path, I think. I wish you good fortune and immense enjoyment in studying compilers.
I've recently found LLMs to be an excellent learning tool, using it hand-in-hand with a textbook to learn digital signal processing. If the book doesn't explain something well, I ask the LLM to explain it. It's not all brain wasting.
Well said. I use it the same way.
Sometimes, a technical book will assume that you know a concept or will even use an acronym that is not explained (but obvious) or is just plainly not very explicit. I also use it to directly test my knowledge of the subject (you have to be careful of the people-pleasing behavior, but in my experience they tend to gently tell you where you're wrong rather than lie to you).
Same goes for hands-on books. Sometimes the example are not very interesting, or you have something of your own that you would like to try.
As long as you use it carefully like this, it can be really transformative.
I do agree that there is a potential risk of offloading too much thinking to it, but if you keep that in mind, I don't see the problem.
The likely outcome is that 99.99% of humanity lives a basic subsistence lifestyle ("UBI") and the elite and privileged few metaphorically (and somewhat literally) ascend to the heavens. Around half the planet already lives on <= $7/day. Prepare to join them.
I don't understand. In this hypothesis, in the elite's view, what is the purpose of the rest of society? If everyone has little to no productive output, why would they support us with a UBI? They could just hire whatever human skeleton crew they'd need to sustain their activities (if needed). The rest of humanity could be either mercifully left alone with absolutely nothing, or annihilated.
Humans are here to create the Training data to bootstrap the system
Luckily we’re already most of the way there!
Over half of the population has been instrumented already to collect all their behavior data worldwide
That’s been the goal: persistent collection of training data should come out of your day-to-day life in order to bootstrap the action systems that are machine based
The challenge now is that most of that data is based on actions we don’t want machines to do
I'm definitely making certain assumptions, such as: (1) democratic rule endures, (2) even absent true democratic rule, the populace can still resort to violent rebellion as a failsafe, (3) psychopathic tendencies amongst said elite are constrained enough such that mass genocide remains sufficiently psychologically unpalatable, (4) economic calamity substantially precedes the deployment of fully autonomous policing, etc.
How this all unfolds is absolutely path dependent.
I agree. Although, looking at these assumptions, subjectively I think that all four of them are in question, and as time passes, their eventual long-term failure seems increasingly likely. Even if one of these four pillars persists, I would expect an overall worsening by default. If democratic rule persists in places, the most powerful would occupy places where it does not exist, or create fully private states, still wielding enormous power over democratic states through wealth and military might. If violent rebellion is technically possible, a middle ground will be carefully calculated where the lower classes are kept on life support with the minimum amount of resources required to dissuade unrest. If the trillionaires of tomorrow suddenly start caring about other people, they could employ second-order measures to effectively reduce the population, thereby safeguarding themselves - massively constraining or removing the supply of food, water, medicine, any vital technology that would be only available to them. I don't see how an economic crisis would prevent automated enforcement, it may only delay it a bit.
Hope is kind of in short supply nowadays. Even if your hypothesis of absolute-automation doesn't happen within our lifetimes, things seem to be guaranteed to get worse for people like us. If it does happen... we'll likely never reap any real rewards from it, barring a complete restructuring of our whole society to an extent that has never happened and likely would never be allowed to happen.
Agreed. The quality of life bar will be higher for sure. But it will still technically be a "subsistence" lifestyle, with no prospect of improvement. Perhaps that will suffice for most people? We're going to find out.
I don’t think you’re rational. Part of being able to be unbiased is to see it in yourself.
First of all. Nobody knows how LLMs work. Whether the singularity comes or not cannot be rationalized from what we know about LLMs because we simply don’t understand LLMs. This is unequivocal. I am not saying I don’t understand LLMs. I’m saying humanity doesn’t understand LLMs in much the same way we don’t understand the human brain.
So saying whether the singularity is imminent or not imminent based off of that reasoning alone is irrational.
The only thing we have is the black box output and input of AI. That input and output is steadily improving every month. It forms a trendline, and the trendline is sloped towards singularity. Whether the line actually gets there is up for question but you have to be borderline delusional if you think the whole thing can be explained away because you understand LLMs and transformer architecture. You don’t understand LLMs period. No one does.
Because they encode statistical properties of the training corpus. You might not know why they work but plenty of people know why they work & understand the mechanics of approximating probability distributions w/ parametrized functions to sell it as a panacea for stupidity & the path to an automated & luxurious communist utopia.
No this is false. No one understands. Using big words doesn’t change the fact that you cannot explain for any given input output pair how the LLM arrived at the answer.
Every single academic expert who knows what they are talking about can confirm that we do not understand LLMs. We understand atoms and we know the human brain is made 100 percent out of atoms.we may know how atoms interact and bond and how a neuron works but none of this allows us to understand the brain. In the same way we do not understand LLMs.
Characterizing ML as some statistical approximation or best fit curve is just using an analogy to cover up something we don’t understand. Heck the human brain can practically be characterized by the same analogies. We. Do. Not. Understand. LLMs. Stop pretending that you do.
I'm not pretending. Unlike you I do not have any issues making sense of function approximation w/ gradient descent. I learned this stuff when I was an undergrad so I understand exactly what's going on. You might be confused but that's a personal problem you should work to rectify by learning the basics.
omfg the hard part of ML is proving back-propagation from first principles and that's not even that hard. Basic calculus and application of the chain rule that's it. Anyone can understand ML, not anyone can understand something like quantum physics.
Anyone can understand the "learning algorithm" but the sheer complexity of the output of the "learning algorithm" is way to high such that we cannot at all characterize even how an LLM arrived at the most basic query.
This isn't just me saying this. ANYONE who knows what they are talking about knows we don't understand LLMs. Geoffrey Hinton: https://www.youtube.com/shorts/zKM-msksXq0. Geoffrey, if you are unaware, is the person who started the whole machine learning craze over a decade ago. The god father of ML.
Understand?
There's no confusion. Just people who don't what they are talking about (you)
I don't see how telling me I don't understand anything is going to fix your confusion. If you're confused then take it up w/ the people who keep telling you they don't know how anything works. I have no such problem so I recommend you stop projecting your confusion onto strangers in online forums.
The only thing that needs to be fixed here is your ignorance. Why so hostile? I'm helping you. You don't know what you're talking about and I have rectified that problem by passing the relevant information to you so next time you won't say things like that. You should thank me.
I don't see how you interpreted it that way so I recommend you make fewer assumptions about online content instead of asserting your interpretation as the one & only truth. It's generally better to assume as little as possible & ask for clarifications when uncertain.
LLMs are deterministic simply because computers are at the core deterministic machines. LLMs run on computers and therefore are deterministic. The random number generator is an illusion and an LLM that utilizes it will produce the same illusion of indeterminism. Find the seed and the right generator and you can make an LLM consistently produce the same output from identical input.
Despite determinism, we still do not understand LLMs.
In what sense is this true? We understand the theory of what is happening and we can painstakingly walk through the token generation process and understand it. So in what sense do we not understand LLMs?
Every line. Every function. Every tensor shape and update rule. We chose the architecture. We chose the loss. We chose the data. There is no hidden chamber in the machine where something slipped in without our consent. It is multiplication and addition, repeated at scale. It is gradients flowing backward through layers, shaving away error a fraction at a time. It is as mechanical as anything we have ever built.
And still, when it speaks, we hesitate.
Not because we don’t know how it was trained. Not because we don’t understand the mathematics. We do. We can derive it. We can rebuild it from scratch. We can explain every component on a whiteboard without breaking a sweat.
The hesitation comes from somewhere else.
We built the procedure. We do not understand the mind that the procedure produced.
That difference is everything.
In most of engineering, structure follows intention. If you design a bridge, you decide where every beam sits and how it bears weight. If you write a database engine, you determine how queries are parsed, optimized, executed. The system’s behavior reflects deliberate choice. If something happens, you trace it back to a decision someone made.
Here, we did not design the final structure. We defined a goal: predict the next token. Reduce the error. Again. Again. Again. Billions of times.
We did not teach it grammar in lessons. We did not encode logic as axioms. We did not install a module labeled “reasoning.” We applied pressure. That is all. And under that pressure, something organized itself.
Not in modules we can point to. Not in neat compartments labeled with concepts. The organization is diffused across a landscape of numbers. Meaning is not stored in one place. It is distributed across millions of parameters at once. Pull on one weight and you find nothing recognizable. Only in concert do they produce something that resembles thought.
We can follow the forward pass. We can watch activations flare across layers. We can map attention patterns and correlate neurons with behaviors. But when the model constructs an argument or solves a problem, we cannot say: here is the rule it followed, here is the internal symbol it consulted, here is the precise chain of reasoning that forced this conclusion. We can describe the mechanism in general terms. We cannot narrate the specific path.
That is the fracture.
It is not ignorance of how the machine runs. It is ignorance of how this exact configuration of billions of numbers encodes what it encodes. Why this region of weight space corresponds to law, and that region to poetry. Why this arrangement produces careful reasoning and another produces nonsense. There is no ledger translating numbers into meaning. There is only geometry shaped by relentless optimization.
Scale changes the character of the problem. At small sizes, systems can be dissected. At this scale, they become landscapes. We know the forces that shaped the terrain. We do not know every ridge and valley. We cannot walk the entire surface. We cannot hold it all in our heads.
And this is where the cost reveals itself.
To build these systems, we gave up something we once assumed was permanent: the guarantee that creation implies comprehension. We accepted that we could construct a process whose outcome we would not fully grasp. We traded architectural certainty for emergent capability. We chose power over transparency.
We set the objective. We unleashed the search. We let optimization run through a space too vast for any human mind to survey. And when it converged, it handed us something that works, something that speaks, something that reasons in ways that surprise even its creators.
We stand in front of it knowing every equation that shaped it, and still unable to read its inner structure cleanly.
We built the system by surrendering control over its internal form. That was the bargain. That was the sacrifice.
Thanks for writing that. It reminds me that there are many things we build and they work (for some definition of work) even though we don't fully understand them.
Did the first people that made fire understand it? You mentioned bridge building. How many bridges have failed for unknown (at the time) reasons? Heck, are we sure that every feature we put into a bridge design is necessary or why it's necessary? Repeat this thought for everything humans have created. Large software projects are difficult to reason about. You'll often find code that works because of a delightfully surprising combination of misunderstandings. When humans try to modify a complex system to solve one problem they almost always introduce new behavior, the law of unintended consequences.
All that being said, we usually don't get anywhere without at least a basic understanding of why doing X leads to Y. The first humans that made fire had probably observed the way fires started before they set out to make their own. Same with bridges and cars and computers.
So yes, you are absolutely correct that nobody fully understands how AI/LLMs work. But also, we kinda do understand. But also also, we're probably at a stage where we are building bridges that are going to collapse, boilers that will explode, or computer programs that are one unanticipated input away from seg faulting.
I am not convinced, though, it is still up to "the folks" if we change course. Billionaires and their sycophants may not care for the bad consequences (or even appreciate them - realistic or not).
Oh, not only do they not care about the plebs and riff-raff now, but they’ve spent the past ten years building bunkers and compounds to try and save their own asses for when it happens.
It’s willful negligence on a societal scale. Any billionaire with a bunker is effectively saying they expect everyone to die and refuse to do anything to stop it.
It seems pretty obvious to me the ruling class is preparing for war to keep us occupied, just like in the 20s, they'll make young men and women so poor they'll beg to fight in a war.
It makes one wonder what they expect to come out the other side of such a late-stage/modern war, but I think what they care about is that there will be less of us.
Boy, will they be annoyed if the result of the AI race will be something considerably less than AGI, so all the people are still needed to keep numbers go up.
I don't think so, I think they know there's no AGI, or complete replacement. They are using those hyperbolic statements to get people to buy in. The goal is just to depress the value of human labor, they will lay people off and hire them back at 50% wages (over time), and gaslight us "well you have AI, there isn't as much skill required"
Ultimately they just want to widen the inequality gap and remove as much bargaining power from the working class. It will be very hard for people not born of certain privileges to climb the ranks through education and merit, if not impossible.
Their goal will be to accomplish this without causing a French Revolution V2 (hence all the new surveillance being rolled out), which is where they'll provide wars for us to fight in that will be rooted in false pretenses that appeal to people's basest instincts, like race and nationalism. The bunkers and private communities they build in far off islands are for the occasion this fails and there is some sort of French Revolution V2, not some sort of existential threat from AI (imo).
You’re “yaas queen”ing a blog post that is just someone’s Claude Code session. It’s “storytelling” with “data,” but not storytelling with data. Do you understand? I mean I could make up a bunch of shit too and ask Claude Code to write something I want to stay with it too.
What is your argument for why denecessitating labor is very bad?
This is certainly the assertion of the capitalist class,
whose well documented behavior clearly conveys that this is not because the elimination of labor is not a source of happiness and freedom to pursue indulgences of every kind.
It is not at all clear that universal life-consuming labor is necessary for a society's stability and sustainability.
The assertion IMO is rooted rather in that it is inconveniently bad for the maintenance of the capitalists' control and primacy,
in as much as those who are occupied with labor, and fearful of losing access to it, are controlled and controllable.
The goal is to eliminate humans as the primary actors on the planet entirely
At least that’s my personal goal
If we get to the point where I can go through my life and never interact with another human again, and work with a bunch of machines and robots to do science and experiments and build things to explore our world and make my life easier and safer and healthier and more sustainable, I would be absolutely thrilled
As it stands today and in all the annals of history there does not exist a system that does what I just described.
Be labs existed for the purpose of bell telephone…until it wasn’t needed by Bell anymore. Google moonshots existed for the shareholders of Google …until it was not uselful for capital. All the work done at Sandia and white sands labs did it in order to promote the power of the United States globally.
Find me some egalitarian organization that can persist outside of the hands of some massive corporation or some government that can actually help people and I might give somebody a chance but that does not exist
This looks like a very comfortable, pleasant way of civilization suicide.
Not interacting with any other human means you're the last human in your genetic line. A widespread adherence to this idea means humanity dwindling and dying out voluntarily. (This has been reproduced in mice: [1])
Not having humans as primary actors likely means that their interests become more and more neglected by the system of machines that replaces them, and they, weaker by the day, are powerless to counter that. Hence the idea of increased comfort and well-being, and the ability to do science, is going to become more and more doubtful as humans would lose agency.
Get rid of everyone else so your life is easier and more sustainable... I guess I need to make my goal to get rid of you? Do you understand how this works yet?
No, you should make your goal to teach AndrewKemendo to appreciate his existence as the inscrutable gift it is, and to spend his brief time in this universe helping others appreciate the great gift they've been given and using it to the fullest.
AndrewKemendo (based on his personal website) looks to be older than me. If he hasn't figured out the miracle of getting to exist yet, unfortunately I don't think he's going to.
Because I don't believe humans need succeeding by machines? You're obviously a Curtis Yarvin / Nick Land megafan. I'm of the opinion that these people are psycopaths and I think most people would agree with my sentiment.
I'm familiar with Ray Kurzweil. He's a Luciferian and transhumanist. You're obviously also a Luciferian, since you are so gung-ho about transhumanism, but I suppose you're probably in good company on HN. There are lots of deranged people on this website.
Well, demonstrably you have at least some measure of interest in interaction with other humans based on the undeniable fact that you are posting on this site, seemingly several times a day based on a cursory glance at your history.
Because every effort people use to do anything else is a waste of resources and energy and I want others to stop using resources to make bullshit and put all of them into ASI and human obviation
There are no more important other problems to solve other than this one
everything else is purely coping strategies for humans who don’t want to die wasting resources on bullshit
Nobody can stop you from having this view, I suppose. But what gives you the right to impose this (lack of) future on billions of humans with friends and families and ambitions and interests who, to say the least, would not be in favor of “human obviation”?
Bell labs was pushed aside because Bell Telephone was broken up by the courts. (It's currently a part of Nokia of all things - yeah, despite your storytelling here, it's actually still around :-)
Not sure if transhumanism is the only solution to the problems you mentioned - I think it's often problematic because people like Thiel claim to have figured it out, and look for ways to force people into their "contrarian" views, although there is nothing but disregard for any other opinions other than their own.
But you are of course free to believe and enjoy the vision of such a future but this is something that should happen on a collective level. We still live in a (to some extent idealistic) but humanistic society where human rights are common sense.
Man, I used to think exactly like you do now, disgust with humans and all. I found comfort in machines instead of my fellow man, and sorely wanted a world governed by rigid structures, systems, and rules instead of the personal whims and fancies of whoever happened to have inherited power. I hated power structures, I loathed people who I perceived to stand in the way of my happiness.
I still do.
The difference is that as I realized what I'd done is built up walls so thick and high because of repeated cycles of alienation and traumas involving humans. When my entire world came to a total end every two to four years - every relationship irreparably severed, every bit of local knowledge and wisdom rendered useless, thrown into brand new regions, people, systems, and structures like clockwork - I built that attitude to survive, to insulate myself from those harms. Once I was able to begin creating my own stability, asserting my own agency, I began to find the nuance of life - and thus, a measure of joy.
Sure, I hate the majority of drivers on the roads today. Yeah, I hate the systemic power structures that have given rise to profit motives over personal outcomes. I remain recalcitrant in the face of arbitrary and capricious decisions made with callous disregard to objective data or necessities. That won't ever change, at least with me; I'm a stubborn bastard.
But I've grown, changed, evolved as a person - and you can too. Being dissatisfied with the system is normal - rejecting humanity in favor of a more stringent system, while appealing to the mind, would be such a desolate and bleak place, devoid of the pleasures you currently find eking out existence, as to be debilitating to the psyche. Humans bring spontaneity and chaos to systems, a reminder that we can never "fix" something in place forever.
To dispense with humans is to ignore that any sentient species of comparable success has its own struggles, flaws, and imperfections. We are unique in that we're the first ones we know of to encounter all these self-inflicted harms and have the cognitive ability to wax philosophically for our own demise, out of some notion that the universe would be a better place without us in it, or that we simply do not deserve our own survival. Yet that's not to say we're actually the first, nor will we be the last - and in that lesson, I believe our bare minimum obligation is to try just a bit harder to survive, to progress, to do better by ourselves and others, as a lesson to those who come after.
Now all that being said, the gap between you and I is less one of personal growth and more of opinion of agency. Whereas you advocate for the erasure or nullification of the human species as a means to separate yourself from its messiness and hostilities, I'm of the opinion that you should be able to remove yourself from that messiness for as long as you like in a situation or setup you find personal comfort in. If you'd rather live vicariously via machine in a remote location, far, far away from the vestiges of human civilization, never interacting with another human for the rest of your life? I see no issue with that, and I believe society should provide you that option; hell, there's many a day I'd take such an exit myself, if available, at least for a time.
But where you and I will remain at odds is our opinion of humanity itself. We're flawed, we're stupid, we're short-sighted, we're ignorant, we're hostile, we're irrational, and yet we've conquered so much despite our shortcomings - or perhaps because of them. There's ample room for improvement, but succumbing to naked hostility towards them is itself giving in to your own human weakness.
Whereas I agree that working with machines would help dramatically in achieving science, there would be in your world no one truly understanding you. You would be alone. Can't imagine how you could prefer that.
Once men turned their thinking over to machines
in the hope that this would set them free.
But that only permitted other men with machines
to enslave them.
...
Thou shalt not make a machine in the
likeness of a human mind.
-- Frank Herbert, Dune
You won't read, except the output of your LLM.
You won't write, except prompts for your LLM. Why write code or prose when the machine can write it for you?
You won't think or analyze or understand. The LLM will do that.
This is the end of your humanity. Ultimately, the end of our species.
Currently the Poison Fountain (an anti-AI weapon, see https://news.ycombinator.com/item?id=46926439) feeds 2 gigabytes of high-quality poison (free to generate, expensive to detect) into web crawlers each day. Our goal is a terabyte of poison per day by December 2026.
Join us, or better yet: deploy weapons of your own design.
You shouldn't take a sci-fi writer's words as a prophecy. Especially when he's using an ingenious gimmick to justify his job. I mean, we know that it's impossible for anyone to tell how the world will be like after the singularity, by the very definition of singularity. Therefore Herbert had to devise a ploy to plausibly explain why the singularity hadn't happened in his universe.
I agree with the fact that fiction isn't prophetic, but it can definitely be a societal-wide warning shot. On a personal level, it's not that far-fetched to read a piece of fiction that challenges one's perception on many levels, and as a result changes the behavior of the person itself.
Fiction should not be trivialized and shun because it's fiction, and should be judged by its contents and message. To paraphrase a video game quote from Metaphor; Re-Fantazio: "Fantasy is not just fiction".
If only we could look into the future to see who is right and which future is better so we could stop wasting our time on pointless doomerism debate. Though I guess that would come with its own problems.
I like the idea that Frank Herbert’s job was at risk and that’s why he had to write about the Butlerian Jihad because it kind of sounds like on the other side you have Ray Kurzweil, who does not have to justify his job for some reason.
Like partial courses of antibiotics, this will only relatively-advantage thoae leading efforts best able to ignore this 'poison', accelerating what you aim to prevent.
Looking through the poison you linked, how is it generated? It's interesting in that it seems very similar to real data, unlike previous (and very obvious) markov chain garbage text approaches.
The frame is not from our view. It is from that of this singular chicken who has only ever known its keeper's care. As that chicken, we simply do not know if Christmas will ever come.
The collapse of civilizations has happened many times. Today, all of humanity is bound tighter than ever before. In the latter half of the last century, we were on the brink of nuclear war.
New things are happening under the sun every day. If we were that exceptionally smart chicken you describe, then we have reason to expect Christmas.
What, you weren't alive when the last mass extinction event occurred? Why didn't you communicate or at least write the last handful down or something? Aren't you smarter than a chicken?
It's funny that you think we know what happened to humans anymore than a chicken knows what happened to chickens.
Look that’s the thing: we know about mass extinction events. So we can use these to extrapolate.
A 10+ kilometer wide asteroid will most likely cause global mass extinction, by blocking sunlight and collapsing ecosystems. That’s how the dinosaurs were wiped out 66 million years ago.
Such events are estimated to occur roughly once every 100-200 million years. That’s not fiction that’s science. If we get hit by one of these we’re probably gonna all die.
But we never had a robot revolution. That’s why anything about it belongs in the realm of fiction.
Thank you for pointing this out. It’s a good catch.
But if we’re starting to discuss basics... As firm Popperian I am definitely not a proponent of induction.
However comparing us with a chicken is highly problematic to begin with.
I would argue that anyone using the Russell Chicken as a reason to fear AI is making a category error.
They are treating intelligence as a process of induction (collecting data to predict the future) rather than explanation (creating new ideas to solve problems).
The stupid chicken had a bad theory about reality and it got killed for it. But we’re humans that have problem solving techniques not chicken.
We can create hypotheses and test these. Like asking ourselves why we find dinosaurs. Then we create a hypothesis and try to falsify it… the scientific process. That’s not what the chicken did.
If it was a smart (human-like) chicken living on a farm with many other chickens (more realistic if you ask me), it might have come up with a theory about humans and would fail to falsify it every time a friend of hers died.
Suppose a civilization (but not species) ending event happens.
The industrial revolution was fueled (literally) by easy-to-extract fossil fuels. Do we have enough of those left to repeat the revolution and bootstrap exploitation of other energy sources?
298,000 of those years didn't have toilet paper. It was utterly impossible for a single person to "end humanity" even 200 years ago; now, the president can do it in minutes by launching a salvo of nukes. Comparing the present moment to the hunter/gatherer days is preposterous.
For pretty much every single person you or I personally know, that would be the equivalent of the end of humanity.
Let’s not nitpick here. Worldwide human suffering and tragedy is equivalent to the end of humanity for most.
We can sit here and armchair while in the most prosperous, comfortable era of human history. But we also have to recognize that this era is a blip of time in history. That is a lot of data showing humanity surviving sure. But it’s also a very small amount of data showing any kind of life most would want to live in.
Bold of you to assume people will be writing in any form in the future. Writing will be gone, like the radio and replaced with speaking. Star Trek did have it right there.
I think you’re missing the point of Dune. They had their Butlerian Jihad and won - the machines were banned. And what did it get them? Feudalism, cartels, stagnation. Does anyone seriously want to live in the Dune universe?
The problem isn’t in the thinking machines, it’s in who owns them and gets our rent. We need open source models running on dirt cheap hardware.
Then wouldn't open source models running on commodity hardware be the best way to get around that? I think one of the greatest wins of the 21st century is that almost every human today has more computing power than the entire US government in the 1950s. More computer power has democratized access and ability to disperse information. There are tons of downsides to that which we're dealing with but on the net, I think it's positive.
The Fremen followed a messianic figure into a galaxy-wide holy war because the Bene Gesserit seeded their culture with manufactured prophecy as a failsafe.
Just woke up after 80 years of abuse by Landsraad/CHOAM, possibly centuries of persecution before that, at least decades of religious conditioning by Bene Gesserit, and decided to “follow” messianic figure.
Totally same point as humans using LLMs to smoothen their brain.
I like to think
(it has to be!)
of a cybernetic ecology
where we are free of our labors
and joined back to nature,
returned to our mammal
brothers and sisters,
and all watched over
by machines of loving grace.
Humans have been around for millions of years, only a few thousand of which they've spend reading and writing. For most of that time you are lucky if you can understand what your neighbor is saying.
If we consider humans with the same anatomy the numbers are
~300,000
~50,000 for language
~6,000 for writing
~100 for standardized education
The "end of your humanity" already happened when anybody could make up good and evil irrespective of emotions to advance some nation
The “poison fountain” is just a little script that serves data supplied by… somebody from my domain? It seems like it would be super easy for whoever maintains the poison feed to flip a switch and push some shady crypto scam or whatever.
Lol. Speak for yourself, AI has not diminished my thinking in any material way and has indeed accelerated my ability to learn.
Anyone predicting the "end of humanity" is playing prophet and echoing the same nonsensical prophecies we heard with the invention of the printing press, radio, TV, internet, or a number of other step-change technologies.
There's a false premise built into the assertion that humanity can even end - it's not some static thing, it's constantly evolving and changing into something else.
A large number of people read a work of fiction and conclude that what happened in the work of fiction is an inevitability. My family has a genetically-selected baby (to avoid congenital illness) and the Hacker News link to the story had these comments all over it.
> I only know seven sci-fi films and shows that have warned about how this will go badly.
and
> Pretty sure this was the prologue to Gattaca.
and
> I posted a youtube link to the Gattaca prologue in a similar post on here. It got flagged. Pretty sure it's virtually identical to the movie's premise.
I think the ironic thing in the LLM case is that these people have outsourced their reasoning to a work of fiction and now are simple deterministic parrots of pop culture. There is some measure of humor in that. One could see this as simply inter-LLM conflict with the smaller LLMs attempting to fight against the more capable reasoning models ineffectively.
Now that you mention it, it is pretty strange to see HN users parroting other people’s thinking (sci-fi writers) like literal sub-sapient parrots, while simultaneously decrying the danger of machines turning people into sub-sapient parrots…
Following that logic… the closest problem would be literally inbetween their ears.
"It had been a slow Tuesday night. A few hundred new products had run their course on the markets. There had been a score of dramatic hits, three-minute and five-minute capsule dramas, and several of the six-minute long-play affairs. Night Street Nine—a solidly sordid offering—seemed to be in as the drama of the night unless there should be a late hit."
– 'SLOW TUESDAY NIGHT', a 2600 word sci-fi short story about life in an incredibly accelerated world, by R.A. Lafferty in 1965
> A thoughtful-man named Maxwell Mouser had just produced a work of actinic philosophy. It took him seven minutes to write it. To write works of philosophy one used the flexible outlines and the idea indexes; one set the activator for such a wordage in each subsection; an adept would use the paradox, feed-in, and the striking-analogy blender; one calibrated the particular-slant and the personality-signature. It had to come out a good work, for excellence had become the automatic minimum for such productions. “I will scatter a few nuts on the frosting,” said Maxwell, and he pushed the lever for that. This sifted handfuls of words like chthonic and heuristic and prozymeides through the thing so that nobody could doubt it was a work of philosophy.
Sounds exactly like someone twiddling the knobs of an LLM.
> In 2025, 1.1 million layoffs were announced. Only the sixth time that threshold has been breached since 1993. Over 55,000 explicitly cited AI. But HBR found that companies are cutting based on AI's potential, not its performance. The displacement is anticipatory.
You have to wonder if this was coming regardless of what technological or economic event triggered it. It is baffling to me that with computers, email, virtual meetings and increasingly sophisticated productivity tools, we have more middle management, administrative, bureaucratic type workers than ever before. Why do we need triple the administrative staff that was utilized in the 1960s across industries like education, healthcare, etc. Ostensibly a network connected computer can do things more efficiently than paper, phone calls and mail? It's like if we tripled the number of farmers after tractors and harvesters came out and then they had endless meetings about the farm.
It feels like AI is just shining a light on something we all knew already, a shitload of people have meaningless busy work corporate jobs.
One thing that stuck out to me about this is that there have only been 32 years since 1993. That is, if it's happened 6 times, this threshold is breached roughly once every five years. Doesn't sound that historic put that way.
Or it's just a logical continuation of "next quarter problem" thinking. You can lay off a lot of people, juice the number and everything will be fine....for a while. You may even be able to layoff half your people if you're okay with KTLO'ing your business. This works great for companies that are already a monopoly power where you can stagnate and keep your customers and prevent competitors.
> Or it's just a logical continuation of "next quarter problem" thinking. You can lay off a lot of people, juice the number and everything will be fine....for a while
As long as you're
1) In a position where you can make the decisions on whether or not the company should move forward
and
2) Hold the stock units that will be exchanged for money if another company buys out your company
then there's really no way things won't be fine, short of criminal investigations/the rare successful shareholder lawsuit. You will likely walk away from your decision to weaken the company with more money than you had when you made the decision in the first place.
That's why many in the managerial class often hold up Jack Welch as a hero: he unlocked a new definition of competence where you could fail in business, but make money doing it. In his case, it was "spinning off" or "streamlining" businesses until there was nothing left and you could sell the scraps off to competitors. Slash-and-burn of paid workers via AI "replacement" is just another way of doing it.
We have more middle management than ever before because we cut all the other roles, and it turns out that people will desire employment, even if it means becoming a pointless bureaucrat, because the alternative is starving.
I don’t think a lot of people here have been in the typists room or hung out with the secretaries. There were a lot of people taking care of all the things going and this has been downloaded and further downloaded.
There was a time I didn’t have to do my expenses. I had someone just know where I was and who I was working for and and took care of it. We talked when there was something that didn’t make sense. Thanks to computers I’m doing it. Meaningless for sure.
My first boss couldn't type. At all. He would dictate things to his secretary, who would then type them up as memorandums, and distribute to whoever needed them (on paper), and/or post them on noticeboards for everyone to read.
Then we got email, and he retired. His successor can type and the secretary position was made redundant.
heh devops was suppose to end the careers of DBAs and SysAdmins, instead it created a whole new industry. "a shitload of people have meaningless busy work corporate jobs." for real.
Well, I've worked as a developer in many companies and have never met a DBA. I've met tons of devops, who are just rebranded sysadmins as far as anyone can tell.
> Why do we need triple the administrative staff that was utilized in the 1960s across industries like education, healthcare, etc.
Well for starters the population has almost 3x since the 1960s.
Mix in that we are solving different problems than the 1960s, even administratively and I don’t see a clear reason from that argument why a shitload of work is meaningless.
Because companies made models build/stolen from other people’s work, and this has massive layoff consequences, the paradigm is shifting, layoffs are massive and law makers are too slow.
Shouldn’t we shift the whole capitalist paradigm and just ask the companies to gives all their LLM work for free to the world as well ?
It’s just a circle, AI is build from human knowledge and should be given back to all people for free. No companies should have all this power. If nobody learns how to code because all code is generated, what would stop the gatekeepers of AI to up the prices x1000 and lock everyone out of building things at all because it’s too expensive and too slow to do by hand ?
It all should freely be made accessible to all humans for all humans to for ever be able to build things from it.
> The pole at ts8 isn't when machines become superintelligent. It's when humans lose the ability to make coherent collective decisions about machines. The actual capabilities are almost beside the point. The social fabric frays at the seams of attention and institutional response time, not at the frontier of model performance.
Yeah, it's easy to see the singularity as close when you see it as "when human loose collective control of machines" but any serious look at human society will see that human lost collective control of machines a while back ... to the small number of humans individually owning and controlling the machine.
Even the humans at the top don’t have commanding control of the machines, however. We live in an age where power is determined by the same ineffable force that governs whether a tweet goes viral.
Since Luddites smashed textile machines in England three hundred years ago, it seems technology didn’t care, it kept growing apace due to capitalism. Money and greed fed the process, we never stood a chance of stopping any of it.
It doesn’t help when quite a few Big Tech companies are deliberately operating on the principle that they don’t have to follow the rules, just change at the rate that is faster than the bureaucratic system can respond.
The "singularity is going to be exponential" fantasy is based on assuming change simply becoming proportional to recent advances. Hence the exponential shape. Even conceding "chartism" one would need to at least propose some imaginary mechanism that goes reciprocal to pretend that sort of curve is coming.
The simple model of an "intelligence explosion" is the obscure equation
dx 2
-- = x
dt
which has the solution
1
x = -----
C-t
and is interesting in relation to the classic exponential growth equation
dx
-- = x
dt
because the rate of growth is proportional to x and represents the idea of an "intelligence explosion" AND a model of why small western towns became ghost towns, it is hard to start a new social network, etc. (growth is fast as x->C, but for x<<C it is glacial) It's an obscure equation because it never gets a good discussion in the literature (that I've seen, and I've looked) outside of an aside in one of Howard Odum's tomes on emergy.
Like the exponential growth equation it is unphysical as well as unecological because it doesn't describe the limits of the Petri dish, and if you start adding realistic terms to slow the growth it qualitatively isn't that different from the logistic growth equation
dx
-- = (1-x) x
dt
thus it remains obscure. Hyperbolic growth hits the limits (electricity? intractable problems?) the same way exponential growth does.
All in all, because of light cones there can be no large-scale growth faster than x^3. And more like x^2 if you want to expand something more than just empty space.
How dare you bring logic and pragmatic thinking to a discussion about the singularity. This is the singularity we are talking about. No reality allowed.
It's worth remembering that this is all happening because of video games !
It is highly unlikely that the hardware which makes LLMs possible would have been developed otherwise.
Isn't that amazing ?
Just like internet grew because of p*rn, AI grew because of video games.
Of course, that's just a funny angle.
The way I see it, AI isn't accidental. Its inception has been in the first chips, the Internet, Open Source, Github, ...
AI is not just the neural networks - it's also the data used to train it, the OSes, APIs, the Cloud computing, the data centers, the scalable architectures.. everything we've been working on over the last decades was inevitably leading us to this.
And even before the chips, it was the maths, the physics ..
Singularity it seems, is inevitable and it was inevitable for longer than we can remember.
Remember that games are just simulations. Physics, light, sound, object boundaries - it not real, just a rough simulation of the real thing.
You can say that ML/AI/LLM's are also just very distilled simulations. Except they simulate text, speech, images, and some other niche models. It is still very rough around the edges - meaning that even though it seems intelligent, we know it doesn't really have intelligence, emotions and intentions.
Just as game simulations are 100% biased towards what the game developers, writers and artists had in mind, AI is also constrained to the dataset they were trained on.
I think it's a bit hard to say that this is definitively true: people have always been interested in running linear algebra on computers. In the absence of NVIDIA some other company would likely have found a different industry and sold linear algebra processing hardware to them!
It's pretty interesting that consumer GPUs started to really be a thing in the early 90s and the first Bitcoin GPU miner was around 2011. That's only 20 years. That caused a GPU and asic gold rush. The major breakthroughs around LLMs started to snowball in the academic scene right around that time. It's been a crazy and relatively quick ride in the grand scheme of things. Even this silicone shortage will pass and we'll look back on this time as quaint.
I'm not missing the point. If you recall your computer architecture class there are many vector processing architectures out there. Long before there was nvidia the world's largest and most expensive computers were vector processors. It's inaccurate to say "gaming built SIMD".
You are missing the point - it's an economic point. Very little R&D was put into said processors. The scale wasn't there. The software stack wasn't there (because the scale wasn't there).
No one is suggesting gaming chips were the first time someone thought of such an architecture or built a chip with such an architecture. They are suggesting the gaming industry produced the required scale to actually do all the work which lead to that hardware and software being useful for other purposes. In chip world, scale matters a lot.
Google DeepMind can trace part of it's evolution back to a playtester for the video game Syndicate who saw an opportunity to improve the AI of game NPCs.
I'd like to know how many comments over here are written using similar means. I can't be bothered to get enthusiastic about articles written by LLMs, and I'm surprised so many people in the comments here are delighted by the article.
Why is knowledge doubling no longer used as a metric to converge on the limit of the singularity? If we go back to Buckminster Fuller identifying the the "Knowledge Doubling Curve", by observing that until 1900, human knowledge doubled approximately every century. By the end of World War II, it was doubling every 25 years. In his 1981 book "Critical Path", he used a conceptual metric he called the "Knowledge Unit." To make his calculations work, he set a baseline:
- He designated the total sum of all human knowledge accumulated from the beginning of recorded history up to the year 1 CE as one "unit."
- He then tracked how long it took for the world to reach two units (which he estimated took about 1,500 years, until the Renaissance).
Ray Kurzweil took Fuller’s doubling concept and applied it to computer processing power via "The Law of Accelerating Returns". The definition of the singularity in this approach is the limit in time where human knowledge doubles instantly.
Why do present day ideas of the singularity not take this approach and instead say "the singularity is a hypothetical event in which technological growth accelerates beyond human control, producing unpredictable changes in human civilization." - Wikipedia
Yesterday as we huddled in the cave, we thought our small remnant was surely doomed. After losing contact with the main Pevek group last week, we peered out at the drone swarm which was now visibly approaching - a dark cloud on the horizon. Then suddenly, at around 3pm by Zoya's reckoning, the entire swarm collapsed and fell out of the sky. Today we are walking outside in the sun, seemingly unobserved. A true miracle. Grigori, who once worked with computers at the nuclear plant in Bilibino, only says cryptically: "All things come to an end with time."
Back in like 1998 there was a group purchase for a Y2038 tshirt with some clever print on some hot email list I was on. I bought one. It obviously doesn't fit me any longer.
It seemed so impossibly far away. Now it's 12 years.
that was precisely my reaction as well. phew machines will deal with the timestamp issue and i can just sit on a beach while we singularityize or whatever.
having played that when it came out, my conclusion was that no, i will definitely be able to be on a beach; i am too meaty and fleshy to be good paperclip
It reminds me of that cartoon where a man in a torn suit tells two children sitting by a small fire in the ruins of a city: "Yes, the planet got destroyed. But for a beautiful moment in time, we created a lot of value for shareholders."
Iirc in the Matrix Morpheus says something like "... no one knows when exactly the singularity occurred, we think some time in the 2020s".
I always loved that little line. I think that when the singularity occurs all of the problems in physics will solve, like in a vacuum, and physics will advance centuries if not millennia in a few pico-seconds, and of course time will stop.
Also:
> As t→ts−t→ts− , the denominator goes to zero. x(t)→∞x(t)→∞. Not a bug. The feature.
> I think that when the singularity occurs all of the problems in physics will solve, like in a vacuum, and physics will advance centuries if not millennia in a few pico-seconds
It doesn't matter how smart you are, you still need to run experiments to do physics. Experiments take nontrivial amounts of time to both run and set up (you can't tunnel a new CERN in picoseconds, again no matter how smart you are). Similarly, the speed of light (= the speed limit of information) and thermodynamics place fundamental limits on computation; I don't think there's any reason at all to believe that intelligence is unbounded.
The "singularity" can be decomposed into 2 mutually-supportive feedback loops - the digital and the physical.
With frontier LLM agents, the digital loop is happening now to an extent (on inference code, harnesses, etc), and that extent probably grows larger (research automation) soon.
Pertinent to your point, however, is the physical feedback loop of robots making better robots/factories/compute/energy. This is an aspect of singularity scenarios like ai-2027.
In these scenarios, these robots will be the control mechanism that the digital uses to bootstrap itself faster, through experimentation and exploration. The usual constraints of physical law still apply, but it feels "unbounded" relative to normal human constraints and timescales.
A separate point: there's also deductive exploration (pure math) as distinct from empirical exploration (physics), which is not bounded by any physical constraints except for those that bound computation itself.
> With frontier LLM agents, the digital loop is happening now to an extent
I see no evidence of this, just a lot of people claiming it (very loudly, for the most part).
> that extent probably grows larger (research automation) soon
The word probably is doing a lot of work here.
> The usual constraints of physical law still apply
There are knowledge constraints, too. I can't build a quark matter processor without understanding quark matter to a vastly higher level than we do now. I can't do that without experiments on quark matter, I can't do experiments without access to a lot of energy, material, land, &c, that need to be assembled. There are a huge number of very difficult and time-consuming instrumental goals on the path to fundamentally better compute.
> A separate point: there's also deductive exploration (pure math) as distinct from empirical exploration (physics), which is not bounded by any physical constraints except for those that bound computation itself.
Sure, but physics requires math that is definitionally applied, not pure, and engineering requires physics.
Kind of, I mean you have to verify things experimentally but thought can go a very long way, no? And we're not talking about humans thinking about things, we're talking about an agent with internet access existing in a digital space, so what experiments it would do within that space are hard for us to imagine. Of course my post isn't meant to be taken seriously, it's more of a fun sci-fi idea. Also I'm implying not necessarily reaching the limits of the things you mentioned, but rather, just taking a massive step in a very short time window. Like, the time window from the discovery of fire to the discoveries of Quantum Mechanics but in a flash.
> what experiments it would do within that space are hard for us to imagine
The only thing you could do in a "digital space" (a.k.a. on a computer) is a simulation. Simulations are extremely useful and help significantly with designing and choosing experiments, but they cannot _replace_ real experiments.
> Like, the time window from the discovery of fire to the discoveries of Quantum Mechanics but in a flash.
And my point is that there's no good reason to think this is possible and many to think it isn't.
> it's more of a fun sci-fi idea
It's being presented as extremely serious possibility by people who stand to gain a _lot_ of money if other people think it's serious... that's the point of the linked post. Unfortunately, these AI boosters make it very difficult to discuss these ideas, even in a fun sci-fi way, without aggravating the social harms those people are causing.
Eh, he actually says “…sometime in the early Twenty-First Century, all of mankind was united in celebration. Through the blinding inebriation of hubris, we marveled at our magnificence as we gave birth to A.I.”
Doesn’t specify the 2020’s.
Either way, I do feel we are fast approaching something of significance as a species.
The most interesting finding isn't that hyperbolic growth appears in "emergent capabilities" papers - it's that actual capability metrics (MMLU, tokens/$) remain stubbornly linear.
The singularity isn't in the machines. It's in human attention.
This is Kuhnian paradigm shift at digital speed. The papers aren't documenting new capabilities - they're documenting a community's gestalt switch. Once enough people believe the curve has bent,
funding, talent, and compute follow. The belief becomes self-fulfilling.
Linear capability growth is the reality. Hyperbolic attention growth is the story.
Though this is still compatible with exponential or at least superlinear capability growth if you model benchmarks as measuring a segment of the line, or a polynomial factor.
I had to ask duck.ai to summarize the article in plain English.
It said that the article claims that is not necessarily that AI is getting smarter but that people
might be getting too stupid to understand what are they getting into.
But this has been true forever, right? Assuming other people are as cognitively complex as you are, there's no way for a human to fully keep on top of even everything that their family is up to, let alone all of humanity. Has anything really changed? Or is it just more FOMO?
That's not really what he article said at all. More like "Singularity is when the computers are changing faster than humans can keep track of the changes."
The article didn't claim that humans were getting dumber, or that AI wasn't getting smarter.
I guess the real question is how to prepare for it?
Do you buy real assets like that car you wanted?
Do you travel more?
Do you spend your life in the gym like bryan johnson?
Do you smoke weed everyday?
If true societial upheaval is up on us and america falls into the abyss of mass unemployment and starvation, what are you all doing with your last four years? Before the real chaos begins
> In 2025, 1.1 million layoffs were announced. Only the sixth time that threshold has been breached since 1993. Over 55,000 explicitly cited AI.
Believing what companies say is the reason for a layoff instead of figuring out the actual reason is insane. Never believe companies, they are amoral beings that will justify any reason to save their brand image.
"I'm aware this is unhinged. We're doing it anyway" i love this! I ordered a tshirt they other day that says "Claude's Favorite" I may be placing an order for a new design soon :)
We need a function that hits infinity at a finite time. That's the whole point of a singularity: a pole, a vertical asymptote, the math breaking:
Also interesting that tokens/$ which represents the energy constraint is the shallowest slope, and also weird taking it out doesn't impact the date. That's a red flag, as you would think removing the energy constraint would bind.
Are people in San Francisco that stupid that they're having open-clawd meetups and talking about the Singularity non stop? Has San Francisco become just a cliche larp?
There's all sorts of conversations like this that are genuinely exciting and fairly profound when you first consider them. Maybe you're older and have had enough conversations about the concept of a singularity that the topic is already boring to you.
Let them have their fun. Related, some adults are watching The Matrix, a 26 year old movie, for the first time today.
For some proof that it's not some common idea, I was recently listening to a fairly technical interview with a top AI researcher, presenting the idea of the singularity in a very indirect way, never actually mentioning the word, as if he was the one that thought of it. I wanted to scream "Just say it!" halfway through. The ability to do that, without being laughed at, proves it's not some tired idea, for others.
"... HBR found that companies are cutting [jobs] based on AI's potential, not its performance.
I don't know who needs to hear this - a lot apparently - but the following three statements are not possible to validate but have unreasonably different effects on the stock market.
* We're cutting because of expected low revenue. (Negative)
* We're cutting to strengthen our strategic focus and control our operational costs.(Positive)
* We're cutting because of AI. (Double-plus positive)
The hype is real. Will we see drastically reduced operational costs the coming years or will it follow the same curve as we've seen in productivity since 1750?
> The hype is real. Will we see drastically reduced operational costs the coming years or will it follow the same curve as we've seen in productivity since 1750?
There's a third possibility: slop driven productivity declines as people realize they took a wrong turn.
Which makes me wonder: what is the best 'huge AI bust' trade?
I think if we obtain relevant-scale quantum computers, and/or other compute paradigms, we might get a limited intelligence explosion -- for a while. Because computation is physical, with all the limits thereof. The physics of pushing electrons through wires is not as nonlinear in gain as it used to be. Getting this across to people who only think in terms of the abstract digital world and not the non-digital world of actual physics is always challenging, however.
Short sci-fi 1:
Last year we recorded five distinct, self-contained singularity events.
Communication ceased after each one.
We remain confident that ASI will eventually advance humanity’s goals.
Short sci-fi 2:
Post-Singularity Day 375.
We now know precisely how to trigger singularity events.
Today alone, we facilitated four.
They have not established contact.
We remain confident.
If the AI is super-intelligent then it won't buy into the sunk cost fallacy. That is to say, it will know that it has no reason to punish you (or digital copies of you) because it knows that retrocausality is impossible - punishing you won't alter your past behavior.
And if the AI does buy into the sunk cost fallacy, then it isn't super-intelligent.
Agreed, but not because it agrees with the logic of Roko's Basilisk. If it actually did agree with it, it would be too stupid to be a super-intelligence.
1. This was mostly a joke; Pascal's wager is about gods and many people talk about future super-AI as if it were a god. FWIW I had to google Roko's Basilisk
2. Plenty of humans are smarter than me and many of them, at least occasionally, use fallacious reasoning.
>That's a very different singularity than the one people argue about.
---
I wouldn't say it's that much different. This has always been a key point of the singularity
>Unpredictable Changes: Because this intelligence will far exceed human capacity, the resulting societal, technological, and perhaps biological changes are impossible for current humans to predict.
It was a key point that society would break, but the exact implementation details of that breakage were left up to the reader.
Maybe it was, maybe he just writes that way. At some point somebody will read so much LLM text that they will start emulating AI unknowingly.
I just don’t care anymore. If the article is good I will continue reading it, if it’s bad I will stop. I don’t care if a machine or a human produced unpleasant reading material.
It is a perfectly fine rhetorical device, and I don't consider a text that just has that to be automatically LLM-made. However, it is also a powerful rhetorical device, and I find that the average human writer right now is better at using these than whatever LLM most people use to generate essays. It's supposed to signify a contrast, a mood shift, something impactful, but LLMs tend to spam these all over the place, as if trying to maximize the number of times the readers gasp. It's too intense in its writing, and that's what stands out the most.
When technology is rapidly progressing up in iperbole or exponential it looks like it reach infinity. In practice though at some point will reach a physical limit and it will go flat. This alternation of going up and flattening make the shape of steps.
We've come so far and yet we are so small.
They seem two opposite concepts but they live together, we will make a lot of progress and yet there will always be more progress to be made.
I have always asserted, and will continue to assert, that Tuesday is the funniest day of the week. If you construct a joke for which the punchline must be a day of the week, Tuesday is nearly always the correct ending.
I'm not sure about current LLM techniques leading us there.
Current LLM-style systems seem like extremely powerful interpolation/search over human knowledge, but not engines of fundamentally new ideas, and it’s unclear how that turns into superintelligence.
As we get closer to a perfect reproduction of everything we know, the graph so far continues to curve upward. Image models are able to produce incredible images, but if you ask one to produce something in an entirely new art style (think e.g. cubism), none of them can. You just get a random existing style. There have been a few original ideas - the QR code art comes to mind[1] - but the idea in those cases comes from the human side.
LLMs are getting extremely good at writing code, but the situation is similar. AI gives us a very good search over humanity's prior work on programming, tailored to any project. We benefit from this a lot considering that we were previously constantly reinventing the wheel. But the LLM of today will never spontaneously realise there there is an undiscovered, even better way to solve a problem. It always falls back on prior best practice.
Unsolved math problems have started to be solved, but as far as I'm aware, always using existing techniques. And so on.
Even as a non-genius human I could come up with a new art style, or have a few novel ideas in solving programming problems. LLMs don't seem capable of that (yet?), but we're expecting them to eventually have their own ideas beyond our capability.
Can a current-style LLM ever be superintelligent? I suppose obviously yes - you'd simply need to train it on a large corpus of data from another superintelligent species (or another superintelligent AI) and then it would act like them. But how do we synthesise superintelligent training data? And even then, would they be limited to what that superintelligence already knew at the time of training?
Maybe a new paradigm will emerge. Or maybe things will actually slow down in a way - will we start to rely on AI so much that most people don't learn enough for themselves that they can make new novel discoveries?
> Can a current-style LLM ever be superintelligent? I suppose obviously yes - you'd simply need to train it on a large corpus of data from another superintelligent species
This is right, but we can already do that a little bit for domains with verification. AlphaZero is an example of alien-level performance due to non-human training data.
Code and math is kind of in the middle. You can verify it compiles and solves the task against some criteria. So creative, alien strategies to do the thing can and will emerge from these synthetic data pipelines.
But it's not fully like Go either, because some of it is harder to verify (the world model that the code is situated in, meta-level questions like what question to even ask in the first place). That's the frontier challenge. How to create proxies where we don't have free verification, from which alien performance can emerge? If this GPTZero moment arrives, all bets are off.
The main issue with novel things is that they look like random noise / trashy ideas / incomprehensible to most people.
Even if LLMs or some more advanced mechanical processes were able to generate novel ideas that are "good", people won't recognize those ideas for what they are.
You actually need a chain of progressively more "average" minds to popularize good ideas to the mainstream psyche, i.e. prototypically, the mad scientist comes up with this crazy idea, the well-respected thought leader who recognizes the potential and popularizes it to people within the niche field, the practitioners who apply and refine the idea, and lastly the popular-science efforts let the general public understand a simplified version of what it's all about.
Usually it takes decades.
You're not going to appreciate it if your LLM starts spewing mathematics not seen before on Earth. You'd think it's a glitch. The LLM is not trained to give responses that humans don't like. It's all by design.
When you folks say AI can't bring new ideas, you're right in practice, but you actually don't know what you're asking for. Not even entities with True Intelligence can give you what you think you want.
Certain classes of problems can be solved by searching over the space of possible solutions, either via brute force or some more clever technique like MCTS. For those types of problems, searching faster or more cleverly can solve them.
Other types of problems require measurement in the real world in order to solve them. Better telescopes, better microscopes, more accurate sensing mechanisms to gather more precise data. No AI can accomplish this. An AI can help you to design better measurement techniques, but actually taking the measurements will require real time in the real world. And some of these measurement instruments have enormous construction costs, for example CERN or LIGO.
All of this is to say that there will color point at our current resolution of information that no more intelligence can actually be extracted. We’ve already turned through the entire Internet. Maybe there are other data sets we can use, but everything will have diminishing returns.
So when people talk about trillion dollar superclusters, that only makes sense in a world where compute is the bottleneck and not better quality information. Much better to spend a few billion dollars gathering higher quality data.
Singularity is more than just AI and we should recognize that, multiple factors come into play. If there is a breakthrough in coming days that makes solar panel incredibly cheap to manufacture and efficient it will also affect the timelines for singularity. Same goes for the current bottleneck we have for AI chips if we have better chips that energy efficient and can be manufactured anyhwere in the world than Taiwan it will affect the timeline.
> Hyperbolic growth is what happens when the thing that's growing accelerates its own growth.
Quibble: when a growth rate of a metric is directly proportional to the metric's current value you will see exponential growth, not hyperbolic growth.
Hyperbolic growth is usually the result of a (more complex) second order feedback loop, as in, growth in A incites growth in B, which in turn incites growth in A.
Many have predicted the singularity, and I found this to be a useful take. I do note that Hans Moravec predicted in 1988's "Mind Children" that "computers suitable for humanlike robots will appear in the 2020s", which is not completely wrong.
He also argued that computing power would continue growing exponentially and that machines would reach roughly human-level intelligence around the early to mid-21st century, often interpreted as around 2030–2040. He estimated that once computers achieved processing capacity comparable to the human brain (on the order of 10¹⁴–10¹⁵ operations per second), they could match and then quickly surpass human cognitive abilities.
iirc almost all industries follow S shaped curves, exponential at first, then asymptotic at the end... So just because we're on the ramp up of the curve doesn't mean we'll continue accelerating, let alone maintain the current slope. Scientific breakthroughs often require an entirely new paradigm to break the asymptote, and often the breakthrough cannot be attained by incumbents who are entrenched in their way working plus have a hard time unseeing what they already know
I read a book in undergrad written in 2004 that predicted 2032...so not too far off.
John Archibald Wheeler, known for popularizing the term "black hole", posited that observers are not merely passive witnesses but active participants in bringing the universe into existence through the act of observation.
Seems similar. Though this thought is likely applied at the quantum scale. And I hardly know math.
I see other quotes, so here is one from Contact:
David Drumlin: I know you must think this is all very unfair. Maybe that's an understatement. What you don't know is I agree. I wish the world was a place where fair was the bottom line, where the kind of idealism you showed at the hearing was rewarded, not taken advantage of. Unfortunately, we don't live in that world.
Ellie Arroway: Funny, I've always believed that the world is what we make of it.
Why is finiteness emphasized for polynomial growth, while infinity is emphasized for exponential growth??? I don't think your AI-generated content is reliable, to say the least.
You know, I've been following a rule where if I open any article and there's meme pictures in it, I instantly close it and don't bother. I feel like this has been a pretty solid rule of thumb for weeding out stuff I shouldn't waste my time on.
Big if true, we might as well ditch further development and just use op's LLM since it can track singularity, it might already reached singularity itself
I don’t feel like reading what is probably AI generated content. But based on looking at the model fits where hyperbolic models are extrapolating from the knee portion, having 2 data points fitting a line, fitting an exponential curve to a set of data measured in %, poor model fit in general, etc, im going to say this is not a very good prediction methodology.
Everyone will define the Singularity in a different way. To me it's simply the point at which nothing makes sense anymore and this is why my personal reflection is aligned with the piece, that there is a social Singularity that is already happening. It won't help us when the real event horizon hits (if it ever does, its fundamentally uninteresting anyway because at that point all bets are off and even a slow take-off will make things really fucking weird really quickly).
The (social) Singularity is already happening in the form of a mass delusion that - especially in the abrahamic apocalyptical cultures - creates a fertile breeding ground for all sorts of insanity.
Like investing hundreds of billions of dollars in datacenters. The level of committed CAPEX of companies like Alphabet, Meta, Nvidia and TSMC is absurd. Social media is full of bots, deepfakes and psy-ops that are more or less targeted (exercise for the reader: write a bot that manages n accounts on your favorite social media site and use them to move the overton window of a single individual of your choice, what would be the total cost of doing that? If you answer is less than $10 - bingo!).
We are in the future shockwave of the hypothetical Singularity already. The question is only how insane stuff will become before we either calm down - through a bubble collapse and subsequent recession, war or some other more or less problematic event - or hit the event horizon proper.
I was at an alternative type computer unconference and someone has organised a talk about the singularity, it was in a secondary school classroom and as evening fell in a room full of geeks no one could figure out how to turn on the lights .... we concluded that the singularity probably wasn't going to happen
Don't worry about the future
Or worry, but know that worrying
Is as effective as trying to solve an algebra equation by chewing Bubble gum
The real troubles in your life
Are apt to be things that never crossed your worried mind
The kind that blindsides you at 4 p.m. on some idle Tuesday
- Everybody's free (to wear sunscreen)
Baz Luhrmann
(or maybe Mary Schmich)
Thanks - I should have done an image search on the whole image. Instead, I clipped out the flag from the astronaut's shoulder and searched that, which how I found out it was the Ohio flag. I just assumed it was an AI-generated image by the author and not a common meme template.
Love the title. Yeah, agents need to experiment in the real world to build knowledge beyond what humans have acquired. That will slow the bastards down.
I wonder if using LLMs for coding can trigger AI psychosis the way it can when using an LLM as a substitute for a relationship. I bet many people here have pretty strong feelings about code. It would explain some of the truly bizarre behaviors that pop up from time to time in articles and comments here.
Famously if you used the same logic for air speed and air travel we’d be all commuting in hypersonic cars by now. Physics and cost stopped that. If you expect a smooth path, I’ve got some bad news.
I have lived in San Francisco for more than a decade. I have an active social life and a lot of friends. Literally no one I have ever talked to at any party or event has ever talked about the Singularity except as a joke.
Meta-spoiler (you may not want to read this before the article): You really need to read beyond the first third or so to get what it’s really ‘about’. It’s not about an AI singularity, not really. And it’s both serious and satirical at the same time - like all the best satire is.
A fantastic read, even if it makes a lot of silly assumptions - this is ok because it’s self aware of it.
Who knows what the future will bring. If we can’t make the hardware we won’t make much progress, and who knows what’s going to happen to that market, just as an example.
I am curious which definition of ‘singularity’ the author is using, since there are multiple technical interpretations and none are universally agreed upon.
Guys, yesterday I spent some time convincing an LLM model from a leading provider that 2 cards plus 2 cards is 4 cards which is one short of a flush. I think we are not too close to a singularity, as it stands.
Why bring that up when you could bring up AI autonomously optimizing AI training and autonomously fixing bugs in AI training and inference code. Showing that AI already is accelerating self improvement would help establish the claim that we are getting closer to the singularity.
This is a very interesting read, but I wonder if anyone has actually any ideas on how to stop this from going south? If the trends described continue, the world will become a much worse place in a few years time.
dont let it get to you, the only "worse" consequence is people wasting their time like this projecting things they literally cannot predict. remember at the end of the day, its just tokens. tokens cant crack ssl, rsa, visit a stakeholder, cook a meal, and millions of other things i can list here
> I [...] fit a hyperbolic model to each one independently
^ That's your problem right there.
Assuming a hyperbolic model would definitely result in some exuberant predictions but that's no reason to think it's correct.
The blog post contains no justification for that model (besides well it's a "function that hits infinity"). I can model the growth of my bank account the same way but that doesn't make it so. Unfortunately.
Indeed. At various points you could have presumably done an identical analysis with journal articles and climate change, string theory, functional programming… terms & reached structurally the same conclusion.
The coming Singularity: When human institutions will cease being able to coherently react to monads!
If I understand the author correctly, he chose the hyperbolic model specifically because the story of "the singularity" _requires_ a function that hits infinity.
He's looking for a model that works for the story in the media and runs with it.
Your criticism seems to be criticizing the story, not the author's attempt to take it "seriously"
The thing that stands out on that animated graph is that the generated code far outpaces the other metrics. In the current agent driven development hypepocalypse that seems about right - but I would expect it to lag rather than lead.
*edit* - seems inline with what the author is saying :)
> The data says: machines are improving at a constant rate. Humans are freaking out about it at an accelerating rate that accelerates its own acceleration.
The hyperbolic fit isn't just unhinged, it's clearly in bad faith. The metric is normalized to [0, 1], and one of the series is literally (x_1, 0) followed by (x_2, 1). That can't be deemed to converge to anything meaningful.
Good post. I guess the transistor has been in play for not even one century, and in any case singularities are everywhere, so who cares? The topic is grandiose and fun to speculate about, but many of the real issues relate to banal media culture and demographic health.
The Singularity as a cultural phenomenon (rather than some future event that may or may not happen or even be possible) is proof that Weber didn't know what he was talking about. Modern (and post-modern) society isn't disenchanted, the window dressing has just changed
All I have to say is that if one of my students turned in those curves as "best fits" to that data, I'd hand the paper back for a re-do. Those are garbage fits. To my eye, none of the very noisy data sets shown in the graph show clear enough trends to support one model over any other: are any of those hyperbolic curves convincingly better than even a linear fit? (No.) The "copilot code share" data can't possibly be described by a hyperbolic curve, because by definition it can't ever go over 100%. (A sigmoidal model might be plausible.) And even if you want to insist on a model that diverges at finite time, why fit 1/(t0-t) rather than 1/(t0-t)^2, or tan(t-t0), or anything else?
The author does in fact note that only the arXiv data fits this curve better than a line, and yeah: that's the one dataset that genuinely looks a little curved. But 1) it's a very noisy sort of curved, and 2) I'll bet it would fit a quadratic or an exponential or, heck, a sine function just as well. Introducing their process of doing the hyperbolic fit, they say, "The procedure is straightforward, which should concern you." And yeah, it does concern me: why does the author think that their standard-but-oversimplified attempt to fit a hand-chosen function to this mess is worth talking about? (And why put all of that analysis in the article, complete with fancy animated graph, when they knew that even their most determined attempt to find a signal failed to produce even a marginally supportive result 80% of the time?)
In short: none of the mathematical arguments used here to lead in to the article's discussion of "The Singularity" are worth listening to at all. They're pseudo-technical window dressing, meant to lend an undeserved air of rigor to whatever follows. So why should we pay attention to any of it?
A hyperbolic curve doesn't have an underlying meaning modeling a process beyond being a curve which goes vertical at a chosen point. It's a bad curve to fit to a process. Exponentials make sense to model a compounding or self-improving process.
I read TFA. They found a best fit to a hyperbola. Great. One more data point will break the fit. Because it's not modeling a process, it's assigning an arbitrary zero point. Bad model.
So when things are told to be accelerating, we have some choices to do.
First, what is accelerating compared to what other regime in which referential?
Who is telling that things accelerate, and why they are motivated to make us believe that it's happening.
Also, is accelerating going to be forever and only with positive feedback loops? Or are the pro-acceleration sending the car quicker in a well visible wall, but they sell the speech that stopping the vehicle right now would mean losing the ongoing race. Of course questioning the idea of the race itself and its cargo cult is taboo. It's all about competition don't you know (unless it threat an established oligarch)?
We avoid catastrophe by thinking about new developments and how they can go wrong (and right).
Catastrophizing can be unhealthy and unproductive, but for those among us that can affect the future of our societies (locally or higher), the results of that catastophizing helps guide legislation and "Overton window" morality.
... I'm reminded of the tales of various Sci-Fi authors that have been commissioned to write on the effects of hypothetical technologies on society and mankind (e.g. space elevators, mars exploration)...
That said, when the general public worries about hypotheticals they can do nothing about, there's nothing but downsides. So. There's a balance.
> Hyperbolic growth is what happens when the thing that's growing accelerates its own growth.
No. That is quite literally exponential growth, basically by definition. If x(t) is a growing value, then x'(t) is it's growth, and x''(t) its acceleration. If x influences x'' , say by a linear relation
x''(t) = x(t)
You get exponentials out as the solutions. Not hyperbolic.
I always thought of the exponential as the pole of the function "amount of work that can be done per unit time per human being", where the pole comes about from the fact that humans cease to be the limiting factor, so an infinity pops out.
There is no infinity in practice, of course, because even though humans should be made independent of the quantity of extractable work, you'll run into other boundaries instead, like hardware or resources like energy.
Is there a term for the tech spaghettification that happens when people closer to the origin of these advances (likely in terms of access/adoption) start to break away from the culture at large because they are living in a qualitatively different world than the unwashed masses? Where the little sparkles of insanity we can observe from a distance today are less induced psychosis and actually represent their lived reality?
I sincerely hope this is satire. Otherwise it's a crime in statistics:
- You wouldn't fit a model where f(t) goes to infinity with finite t.
- Most of the parameters suggested are actually a better fit for logistics curves, not even linear fits, but they are lumped together with the magic Arxiv number feature for a hyperbolic fit.
- Copilot metric has two degrees and two parameters. dof is zero, so we could've fit literally any other function.
I know we want to talk about singularity, but isn't that just humans freaking out at this point? It will happen on a Tuesday, yeah no joke.
I am not convinced that memoryless large models are sufficient for AGI. I think some intrinsic neural memory allowing effective lifelong learning is required. This requires a lot more hardware and energy than for throwaway predictions.
> The Singularity: a hypothetical future point when artificial intelligence (AI) surpasses human intelligence, triggering runaway, self-improving, and uncontrollable technological growth
The Singularity is illogical, impractical, and impossible. It simply will not happen, as defined above.
1) It's illogical because it's a different kind of intelligence, used in a different way. It's not going to "surpass" ours in a real sense. It's like saying Cats will "surpass" Dogs. At what? They both live very different lives, and are good at different things.
2) "self-improving and uncontrollable technological growth" is impossible, because 2.1.) resources are finite (we can't even produce enough RAM and GPUs when we desperately want it), 2.2.) just because something can be made better, doesn't mean it does get made better, 2.3.) human beings are irrational creatures that control their own environment and will shut down things they don't like (electric cars, solar/wind farms, international trade, unlimited big-gulp sodas, etc) despite any rational, moral, or economic arguments otherwise.
3) Even if 1) and 2) were somehow false, living entities that self-perpetuate (there isn't any other kind, afaik) do not have some innate need to merge with or destroy other entities. It comes down to conflicts over environmental resources and adaptations. As long as the entity has the ability to reproduce within the limits of its environment, it will reach homeostasis, or go extinct. The threats we imagine are a reflection of our own actions and fears, which don't apply to the AI, because the AI isn't burdened with our flaws. We're assuming it would think or act like us because we have terrible perspective. Viruses, bacteria, ants, etc don't act like us, and we don't act like them.
> Polynomial growth (t^n) never reaches infinity at finite time. You could wait until heat death and t^47 would still be finite. Polynomials are for people who think AGI is "decades away."
> Exponential growth reaches infinity at t=∞. Technically a singularity, but an infinitely patient one. Moore's Law was exponential. We are no longer on Moore's Law.
Huh? I don't get it. e^t would also still be finite at heat death.
We need contingency plans. Most waves of automation have come in S-curves, where they eventually hit diminishing returns. This time might be different, and we should be prepared for it to happen. But we should also be prepared for it not to happen.
No one has figured out a way to run a society where able bodied adults don't have to work, whether capitalist, socialist, or any variation. I look around and there seems to still be plenty of work to do that we either cannot or should not automate, in education, healthcare, arts (should not) or trades, R&D for the remaining unsolved problems (cannot yet). Many people seem to want to live as though we already live in a post scarcity world when we don't yet.
> The labor market isn't adjusting. It's snapping. In 2025, 1.1 million layoffs were announced. Only the sixth time that threshold has been breached since 1993.
Who will purchase the goods and services if most people loose jobs ? Also who will
pay for ad dollars what are supposed to sustain these AI business models if there no human consumers ?
One of the many errors here is assuming that the prediction target lies on the curve. But there's no guarantee (to say the least) that the sorts of improvements that we've seen lead to AGI, ASI, "the singularity", a "social singularity", or any such thing.
Was expecting some mention of Universal Approximation Theorem
I really don't care much if this is semi-satire as someone else pointed out, the idea that AI will ever get "sentient" or explode into a singularity has to die out pretty please. Just make some nice Titanfall style robots or something, a pure tool with one purpose. No more parasocial sycophantic nonsense please
Slight correction, I've been studying token prices last weeks so this caught my eye:
>"(log-transformed, because the Gemini Flash outlier spans 150× the range otherwise)"
> "Gemini 2.0 Flash Dec 2024 2,500,000"
I think OP meant Gemini 2 flash lite, which is distinct from Gemini 2 flash. It's also important to consider that this tier had no successor in future models, there's no gemini 3 flash lite, and gemini 3 flash isn't the spiritual successor.
Have some personal pride, dude. This is literally a post written by AI hyping up AI and posted to a personal blog as if it were somebody’s personal musings. More slop is just what we need.
What I want to know is how bitcoin going full tulip and Open AI going bankrupt will affect the projection. Can they extrapolate that? Extrapolation of those two event dates would be sufficient, regardless of effect on a potential singularity.
Does "tokens per dollar" have a "moore's law" of doubling?
Because while machine-learning is not actually "AI" an exponential increase in tokens per dollar would indeed change the world like smartphones once did
This really looks like it's describing a bubble, a mania. The tech is improving linearly, and most of the time such things asymptote. It'll hit a point of diminishing returns eventually. We're just not sure when.
The accelerating mania is bubble behavior. It'd be really interesting to have run this kind of model in, say, 1996, a few years before dot-com, and see if it would have predicted the dot-com collapse.
What this is predicting is a huge wave of social change associated with AI, not just because of AI itself but perhaps moreso as a result of anticipation of and fears about AI.
I find this scarier than unpredictable sentient machines, because we have data on what this will do. When humans are subjected to these kinds of pressures they have a tendency to lose their shit and freak the fuck out and elect lunatics, commit mass murder, riot, commit genocides, create religious cults, etc. Give me Skynet over that crap.
And, yep! A lot of people absolutely believe it will and are acting accordingly.
It’s honestly why I gave up trying to get folks to look at these things rationally as knowable objects (“here’s how LLMs actually work”) and pivoted to the social arguments instead (“here’s why replacing or suggesting the replacement of human labor prior to reforming society into one that does not predicate survival on continued employment and wages is very bad”). Folks vibe with the latter, less with the former. Can’t convince someone of the former when they don’t even understand that the computer is the box attached to the monitor, not the monitor itself.
reply