I did a show HN & it didn't get much traffic, but I've been working on https://moral.games/. A kind of debate PvP game, where you try to convince an AI judge of a certain moral position given an ethical conundrum.
Was trying to combine AI with generative storytelling with a card game. It was a fun experiment. To play, you'll have to get a friend to queue up at the same time.
I think OpenAI and Anthropic will go bust, or at least be scrapped for parts in the next 5 years or so. It's clear that the extreme cost used up for training is impossible to recoup, as inference is already being subsidized.
It's also clear that, as Tan indicates, open-weight models will be (and basically already are) just as good as frontier models. It's all about the harness, baby. We will have two main forks in the road, and two new industries created:
- AI hardware (NVidia/Cerebras/etc.), the equivalent of Intel/AMD
- AI software (harnesses, assistants, etc.) the equivalent of Microsoft/Apple
We already saw a glimmer of this with popularity of OpenClaw—the problem is that it's janky, hard to set up, inconsistent, and very hacker-esque. Imo "AI labs" will be a dying breed because there's no real money in the actual models if they get commoditized, which they already kind of are.
Inference is not being subsidized and in fact has pretty high margins.
Similar-sized open weight models on openrouter are 15x cheaper per token than the big labs. This should reflect the isolated cost of inference, since 3rd party hosts have no reason to subsidize and no training costs to amortize.
Only datacenter buildout costs are being subsidized.
The majority of revenue comes from API usage. The majority of usage comes from subscriptions. For any of the numbers to make any sense, subscriptions must be subsidized ergo the majority of usage is subsidized. A single $200 subscription can incur upwards of $10,000 in API equivalent usage (and even more when there are frequent resets).
If it were true that Anthropic and OpenAI were profitable on all inference they wouldn’t need to constantly raise so much money. Anthropic regularly announce huge investments in infrastructure but it is all smoke and mirrors, data center build out costs aren’t being paid by OpenAI and Anthropic, they’re financed externally. Google, for example, are backstopping tens of billions of datacenter build outs that are being financed based on commitments but not investment from Anthropic.
You are underestimating the insanity of subscription subsidization. Being profitable on API inference is meaningless when it is such a small proportion of usage and is only going to fall off a cliff as cheap open weight models become more capable.
The absolute majority of tokens are being subsidized and as soon as the subsidies end usage will fall off a cliff, rendering all the data center buildout a terrible waste of money.
The numbers in the article are forecasts but let’s take them as real. That’s $10bn of revenue, the majority from enterprise customers, let’s say 75% from enterprise API usage: $7.5 billion. If the margin on inference is 80% that means of the $7.5bn in enterprise revenue they’re spending $1.5bn on compute. Yet we know that they actually spend over $5bn per month on compute, which includes the $1.25bn per month to SpaceX.
If $7.5bn is their enterprise revenue and it costs just $1.5bn to generate, that leaves $3.5bn in compute costs to account for. Dario previously said that training costs less than inference so training can’t explain it.
If subscriptions aren’t the majority of usage and aren’t subsidized, where is the money going? Anthropic don’t spend money on data centre build out so that can’t be it either.
> If the margin on inference is 80% that means of the $7.5bn in enterprise revenue they’re spending $1.5bn on compute.
I don't think you can reverse this out like this because the 80% rate is before payments to "distribution partners, including Amazon". I think that payment includes the hosting cost for that those model but it's unclear.
> Dario previously said that training costs less than inference
Do you have a source for that?
Are you sure you aren't conflating the statements Dario has made that training costs less than they make on inference (over the life cycle of a model)?
> I don't think you can reverse this out like this because the 80% rate is before payments to "distribution partners, including Amazon". I think that payment includes the hosting cost for that those model but it's unclear.
The "hosting cost" is paid for by Anthropic and is the largest cost. The money Anthropic pay to Amazon for delivering Anthropic models via Bedrock is separate, independent of compute costs, best thought of as commission.
The forecasted / guessed / estimated 80% number is based what customers pay per token minus the projected compute costs, i.e: the people who believe that Anthropic has 80% margins on tokens believe that Anthropic spend $0.20 on inference compute for every $1 of per-token billed-via-the-api revenue.
We know that there are hundreds of thousands of fixed-price subscriptions being used to their absolute maximum, with many people bragging about how many subscriptions they run in parallel. These tokens are not included in the 80% margins, they are acknowledged to be "subsidized". People like @theo on Twitter post almost daily about how much they're milking Anthropic and OpenAI with leaderboards.
Both Anthropic and OpenAI (more so OpenAI) do "resets" where they increase the limits available to people on their fixed price plans. We know that there are people paying $1,000 per month for multiple subscriptions to generate tokens that would cost $50,000 via the API. Even if Anthropic's margins are 80% on compute for per-token billing, that's still $10,000 of cost to Anthropic generating just $1,000 in revenue. Multiply that by tens of thousands or maybe even hundreds of thousands of subscriptions.
Anthropic and OpenAI have raised over $100 billion each and continue to raise. If they're making 80% or even 50% margins on $10 billion in revenue per month they would not need to raise, they would be shouting for the roof tops about how profitable they are, they wouldn't be delaying their IPOs, yet they're only profitable by non-GAAP metrics like WeWork's classic "Community-adjusted EBITDA" or in this case "per-token-adjusted EBITDA" or whatever they will call it in their IPOs.
Yes, they're selling tokens via the API for more than they cost, they are profitable on per-token billed inference, it has positive margins, but those profits are obliterated when you account for all the inference they're paying for out of pocket on fixed price subscriptions, upon which they keep increasing limits because they desperately need to show growth further harming their profitability (consuming all of the money they make from their API).
If Anthropic and OpenAI needed to be profitable tomorrow, they could be, they could kill off all their fixed price subscription plans and charge only for usage via the API, they'd print money, but they'd lose mindshare because nobody except for enterprises can afford to pay the true cost, all the regular people would switch to cost effective good-enough models, and then within months, the enterprises would start to switch too because no longer would their employees be claude-pilled.
Anthropic and OpenAI cannot turn off subsidization, thus, their margins on per-token API billing are not important in any discussion about their long term financial wellbeing. Just look at the large scale customers like Harvey (~15 trillion tokens per month, ~$50m+ in spend) who are, sensibly, investing in building their own specialized models that are cheap to run so they can cut their spend by 90%. That's profitable revenue for Anthropic / OpenAI today, but completely gone soon.
"This week, Noah Smith and Erik Torenberg are joined by Dario Amodei, CEO and Co-founder of Anthropic. Dario talks about the economics of AI development, the comparative advantage of AI companies like Anthropic, AI safety, and his stance on California's SB 1047 bill. They also discuss the impacts of AI on global power dynamics, competition between the US and China, and inequality in an AI-powered world."
At around 12 minutes in:
"I think actually even if such a model is released one thing you know that's a this analogy to to open- Source software is that these big models they're actually very expensive to run on inference the majority of the cost is is inference not necessarily the training of the model so if you have only you know I don't know 10 20% 30% better way to do inference that can kind of negate the effect so the economics are kind of strange yes there's this giant fixed cost that you have to amortise but then there's also the per unit cost of inference and small differences in that can actually again assuming the thing is deployed widely enough make a very big difference so I don't know quite how that's going to play out"
The scales have changed since then with inference costs falling and more being spent on training but the fundamentals are the same. Inference is expensive, in part, because peak usage dictates capacity whereas capacity can dictate training. Anthropic must pay billions of dollars per month to be able to handle peak inference, hence their efforts to try and shape usage by offering discounts / flexible limits at different times of the day. They can train when capacity permits.
Do we know that? As I understand it, enterprise customers pay more. Do we know the usage breakdown between monthly subscribers vs enterprise accounts? I agree that it's inevitable that subsidized subscriptions are unlikely to last forever, but that's not the only assumption in your argument.
Edit: I think "enterprise customers pay more" was poorly phrased. I mean that enterprise customers are charged per token, presumably with a profit margin, and thus are not subsidized. While personal accounts are (thought to be) highly subsidized if you consistently max out the quotas. We also don't know what proportion of personal accounts do that though, which is another big question mark.
I think you're also missing a quirk and that is, is everyone on a $200 plan using $10,000 worth of equivalent API spend?
I know people that have the most expensive plan on all the platforms... because
The other side to that is, what is 'cost'? Is cost just inference or are expenses also being taken into account? Because the expenses of these companies are huge to build the models.
> Inference is not being subsidized and in fact has pretty high margins.
1. Companies are trying to decrease costs, not increase it, and are looking at alternatives
2. Competitors are catching up, and even if the frontier labs are "better" at some things (like writing plans or complicated analysis), the competitors can take a lot of the inference on routine tasks like implementing a well-defined plan
3. The frontier labs don't just need to have high margins right now. They have to pay back their massive liabilities.
> Inference is not being subsidized and in fact has pretty high margins.
I was referring to the "AI labs" here. Sam Altman himself conceded that OpenAI is losing money on the $200 subscription. Using open-weight/open-source models is indeed cheaper (and no reason for inference to be subsidized).
That's not what I mean. If competitors can offer tokens 15x cheaper, the big labs must have high margins per token. (which they can use to amortize training costs)
>Sam Altman himself conceded that OpenAI is losing money on the $200 subscription.
They have since stopped offering the $200 subscription, probably for this reason.
Subscription margins are harder to judge because it depends on usage; token costs are a better comparison.
I think this is very possible. Plus, something I don't see talked about enough here. The VERY fragile supply chain that keeps it all going. Look at what is happening in the Middle East.
The US can no longer keep global trade secure on the high seas. What if the supply chains for GPUs get disrupted for months, a year? Then what?
idk where this comes from but it's laughably false.
the only place where actual subsidization (below cost) might be happening are the subscriptions. even that is unlikely because to be truly below cost you either need to offer below cost of electricity which isn't happening, have potential API users using multiple subscriptions or have opportunity cost loss due to saturation.
If harness is all that matters, a co-developed harness + model stack + large compute availability advantage + massive distribution advantage with data for post training will win the market.
I agree with this statement in general, but it “hurts less” to spend money when you are running things yourself. Hard for me to give a specific definition as to why, but it may be more palatable to companies to burn their own cash on their own hardware.
Maybe it is “sunked cost” or maybe it is “I will do it myself dammit”.
Obligatory "what's Lygma?" But in all seriousness, this AI doomerism has reached a comedic inflection point. A few years ago, GPT-2 was too dangerous to release, then some Google weirdo said Gemini had a soul or something, then Mythos-tier became a meme, then some kid quit and went on FOX News talking about Skynet, and now Amodei and Altman are both on the "we need to slow down" train again; this is after we've already been down that road and Claude was banned overseas; wait, actually is that ban still around? Honestly who gives a fuck at this point.
It's all theatre. OpenAI and Anthropic will most likely go bust—or, more realistically sold for parts—, and they absolutely should for stealing my (books I wrote, blog posts, etc.) and many others' intellectual property. We're reaching a point where models are becoming commodetized and I'm 100% convinced the next move will be a sort of "software layer" on top of these reasoning systems which will be the actual revolution. The model itself won't be that interesting anymore, it's all the work that goes around it that makes it worthwhile (kind of what computers and phones are today; chips are amazing, but the software is really the magic).
The only scary part is that the boomers in Congress might actually believe these nerds, but seeing how Big Tech approval ratings are grazing the levels of Big Tobacco in the 90s, I don't think we have much to worry about.
I've been saying this since using Databricks at a company almost a decade ago. Most folks do not need big data tools, and it's just so entrenched because everyone wanted to be a "big data" company and pandas was how you handled big data.
I'm with @nilesh on this one, and not exactly sure how merely the existence of a solution precludes the advancement of human knowledge. If a problem is "solved" (say, symbolically verified) without any insights gained, it doesn't seem very interesting to the profession.
Navier-Stokes is a bit different (because there's a prize attached, so "scooping" matters), but almost all interesting problems don't have any prizes attached.
Humanity is very biased for the culmination of work, considering everything that comes before and after busywork for the lower masses.
Replicating a paper is just as valuable scientifically as publishing it, but how many careers advance through replication?
If we move the goal from "find the solution" to "clear up the LLMs work" that doesn't bode well neither for the attractiveness of the problem nor for the career of the professional that takes the challenge.
> Replicating a paper is just as valuable scientifically as publishing it, but how many careers advance through replication?
A lot. In fields where knowledge is incrementally building on previous work the reason the whole field hasn't collapsed from the replication crisis is that usually the results that are really high impact are replicated in as an initial step in new research building on it. It's almost never the focus of the paper but you'll often find a quick mention in methods/supplemental of some previous work that was verified to be valid by a replication of a key technique etc. you'll have crisis where old tools are found to be problematic and findings end up revisited etc. Plus fields like clinical research where there's an awful lot of focus on replicating findings using staged clinical trials with increasing statistical power to determine if new interventions work - that's driven by regulatory requirements grounded in good science and a lot of people make careers in just that.
In mathematics, finding novel proofs of a given result is often valuable; it may be a shorter proof (demonstrating better/expanded understanding of the problem) or a translation of the problem into a new domain, setting up more cross-domain advances.
>Replicating a paper is just as valuable scientifically as publishing it, but how many careers advance through replication?
I don’t think this is true, especially for novel or unexpected results. I suppose it depends on what you mean by scientifically, and there is a debate in the philosophy of science about what the value of research even is, but a successful replication does not result in substantial updates to one’s beliefs in the way new research does. And if the goal of science is to change our beliefs and bring them closer to what is “real”, successful replications can’t be as valuable as the initial research almost by definition.
From a pure statistical perspective the first scientific paper shouldn’t update your beliefs as much as the independent replication study.
People don’t behave this way, but a high percentage of all papers have known flaws and that goes up even higher when you consider unknown flaws. Replication doesn’t own its own solve the underlying issue, but independent replication removes a huge range of potential issues on top of providing more information.
I agree, with the caveat that it probably matters how novel the paper is, the effect size, and the confidence interval. A reputed new psychological phenomenon I'd be more skeptical of than, idk, a newly detected exoplanet.
That alleged superconductor from a few years ago - everybody kind of held their breath and waited for the reproduction.
You have created a fraud machine. Why? With no answer checking then why not make up the most fraudulent crap you can get away with?
Examples: A huge portion of recent non-reproducable science papers.
---
Your thinking, along with everybody that's doing this rat race is causing the pumping out of papers with questionable data, but very little to ensure we are actually making correct science.
I find that to be an issue of maturity (focusing only on the climax and not the process). In Japan, where I live, the culture has a greater appreciation for the context & process, not just the moment of victory.
If you examine the consequences of the inversion of the peak, you realise the need for a balanced perspective.
> An AI-generated solution always provides ... proof that there is a solution
This is only true in the most trivial sense. A solution is a solution, sure... but how do you know it's a solution, and not an incoherent jumble of words? A human has to review and vouch for it.
Just because the AI gives you an arxiv-worthy PDF, or a Lean proof which compiles, doesn't mean it proves what the AI says it does. The AI could give you the same PDF/Lean code and says it proves the opposite, how would anyone know the difference?
You can't advance human understanding unless you produce things that humans can understand.
Not an expert by any means but the assumption here as I understand it is that the arxiv worthy PDF would not be acceptable or meaningful for impossible to understand proofs. And the lean proof would be meaningless unless the specific expression being proven is human understandable as the direct translation of the question the human is asking in formal form. So proving the negation is not a thing but if you make a subtle mistake in translating the statement you want to prove then obviously the QI is going to be proving the wrong thing. And otherwise you're relying on the correctness of lean as a system and on identifying/preventing if the proof is adversarially exploiting bugs in lean to falsely prove things.
> Just because the AI gives you an arxiv-worthy PDF, or a Lean proof which compiles, doesn't mean it proves what the AI says it does. The AI could give you the same PDF/Lean code and says it proves the opposite, how would anyone know the difference?
> You can't advance human understanding unless you produce things that humans can understand.
And you can't advance human understating unless you maintain that understanding.
I can see a version of the junior software engineer problem here: AI wrecks the problems that could train and motivate the next generation mathematicians, so students abandon the field because there's no place for them. The senior mathematicians who can review/vouch/prompt for AI output like Tao retire and die. Then there's no more math that anyone can understand and no more open problems for it to solve.
And that's probably happening already. I've read articles about AI performing the journeyman work that mathematicians cut their teeth on, rendering years of work obsolete, and derailing the careers that work was meant to start.
I might be wrong, but making an assumption that you could learn to read the mathematical output of the AI long before you could write a solution yourself. But hey, what do I know, I'm not a mathemagition.
This is definitely true in an information theory sense: having more knowledge is always better than less knowledge. However, it may not be true in math as a social human endeavor, and having answers without interesting paths to get there may not expand human mathematics in the same way.
If Fermat had a book with larger margins, would Weil have devoted so much time to proving the Taniyama-Shimura conjecture? No one can say.
Were early in this cycle, we will learn to do more, and exercise our new capabilities more fluently, which in turn will create more skilled practitioners
Consider the abacus, calculator, computer, etc, each of these enhanced mathematicians’ capabilities and thus outputs.
> Mathematicians will be less likely to work on a problem if there is a solution
Yes, that is Tao's premise, I'm just not sure I buy it. Suppose an oracle existed which could answer any question truthfully. Let's ignore the mechanics of this for now, but it could say things like "the Riemann hypothesis is False" or whatever and we would take it as gospel.
Does this mean that we wouldn't have mathematicians or physicists or computer scientists or biologists anymore? I genuinely don't think so.
I think his point is that AI is not creating new problems. It may solve "the Riemann hypothesis" but may completely fail to posit a "Mythos hypothesis" which is vital to advance the field. In fact, achieving the former may make the latter even harder because it will disincentivize production of human mathematics which has till now been the only source of "interesting" problems.
FWIW this is my understanding of his argument and I am not a mathematician.
As Tao points out, merely suggesting new open questions isn't really sufficient. Part of what gives these problems their fame is their notoriety, their difficulty, the fact that many prodigious mathematicians have spent an evening or week or month or several years studying it.
It wouldn't be as interesting if it had just been solved by the fifth random mathematician who considered it
Notably, gardening a new field of study in math is somewhat nontrivial. You have to introduce the field, illustrate some relevance or connections, and then - and this is key - not solve all of the low-hanging fruit yourself! Because you need somebody else to become an expert in that particular field.
The analog in programming is: if a large company merely open sources a product that's decent but not great and in a language nobody wants to maintain, but they don't commit to maintaining it themselves.
Suddenly there's a bit of a vacuum because in order to provide something of value, you either need to:
1. Implement something more complete than was initially open sourced
2. Or maintain something in a horrendous language while incrementally improving it and keeping it relevant
3. Or rewrite it into a tolerable and maintainable modern language.
What the large company has done is create a vacuum in the tool space where you now require extreme motivation to get someone else to step in.
Note that in this scenario, in 2026, it's actually not such a big deal. I think several recent models could happily translate it into a more maintainable language themselves or happily maintain it in the original crufty one. And so the question is: which parts of this analogy are true in math, too?
Yes, these things are definitely being conflated and there's a fourth conflated thing I think Tao is particular getting at:
(4) Mathematics, the community which is a living system that decides what is interesting, constructs shared frameworks, transfers ideas between domains, develops taste, teaches new mathematicians, and continually emphasizes what counts as important mathematics, especially in it's overall value to humanity.
This is the level in which theory building, simplification, integration and applications are built on. Contributing meaningfully here requires much more than generating solutions - it requires direction and restraint.
Tao seems to be saying that this direction and restraint is the scarce resource that drives mathematical progress, not problem-solving ability. And AI is not only insufficient at it, but results generated by AI are destroying human ability to excercise this resource.
This seems similar to another problem AI sucks at - drafting legal agreements. Despite being great at evaluating, interpreting, and comparing legal agreements it fails amazingly at drafting them. This is because what you don't say/do is vastly more important than what you do say/do.
And AI is great saying and doing things, it's the selection based on implied values that it struggles with.
You pick your axioms and you see what must be true. With math and logic, we can reason about any possible universe even though we live in only one of them.
The sphere of human comprehensible mathematics is finite. Once everything is solve it is not necessary to advance the field. The recurring error her is to say ai is not the product of human effort but another agent. Ai is human. Ai may well be speeding up human comprehension of math to its limits in which case there is no further need to advance the field and mathematicians might need to get a job. Why is this a bad thing?
But this oracle doesn't just say true / false. It also gives a proof. That makes it much less exciting (not to mention beneficial for your career) to find another one (or even worse, the same one).
What do u think lean is? That's like saying a program that works, is inscrutable because it appeals to the oracle of "code test cases" to prove itself correct.
You're either being intentionally obtuse, or unintentionally ignorant.
Have you tried to read the Lean proofs produced for any of the recent high-profile results? They're extremely long, terribly structured, and don't indicate which parts are restating known results from literature and which are unique to the proof at hand. That's what makes them inscrutable.
It's similar to Mochizuki claiming to have proved the ABC conjecture, with a proof depending on ideas developed over a large number of obscure papers, that required mathematicians to spend a lot of time before they felt they understood it well enough to point out flaws.
If AI solves all famous open problems and the non-famous ones, too, without advances in the readability of their output, there'll still be some work to do to digest and rearrange the proofs for human consumption. During that process, the mathematician may well get some new ideas...
For all your bombast, do take a moment to note that no mathematician has actually said openai has not produced a correct solution. Should make you think.
> Does this mean that we wouldn't have mathematicians or physicists or computer scientists or biologists anymore?
In the case of mathematicians, I think not as researchers. What would a research mathematician do? I don't think there would be any reason to try to gain insight from proofs that AI made for the sake of understanding. I don't see what that would achieve besides just retaining extremely niche knowledge (which AI or the oracle already does). The whole point of having that knowledge was to build toward novel work which the AI/oracle does. Also, the time spent and difficulty understanding them could be very high but with no payoff besides just understanding them because the AI/oracle would be used to solve all the problems anyway.
I mean, the oracle doesn't really seem so hypothetical right now. And clearly it's going to drastically change these fields, and mathematics, particularly pure mathematics, must change most of all in order to adapt to the existance of a math oracle (or something close to it).
More of a 'its the journey' rather than the destination type of thing.Since the insights , quirks, tricks and procedures gained along the way allows insights intoother at that moment unknown problem/domains in the future.
As far as researchers sharing their data/notes with the AI hyperscalars looks like that would be coming to an end wihth a mor guild-like structure going forward to prevent their progress being fron-run by the AI labs.
I wonder if it would be possible for researchers and scientists to submit their papers to an organization which would then collect them, submit them for peer review by other experts in the field, and then release them in periodical form ONLY to individuals and organizations who pay a subscription fee in order to read them while suing those who try to redistribute them without permission?
Why would society fund mathematicians if they decided to become a guild that hides secrets? They could pursue that as a hobby, but they’d end up like the coders who refuse to use LLMs - rapidly becoming irrelevant and a bit sad from an outsider’s perspective.
Ah but heres the thing , society/gov expects mathematicians to be productive and tries to measure that by awards/publications/citations gained. Within a guild ope or secret they could possibly use a local LLM (even if slow) to accelerate their collective output.While ensuring their credit/publication/citations remain intact rather than with the AI labs taking a lions share of that.
Think along the lines of the Nicolas Bourbaki persona/collective :
" was a collective pseudonym chosen in 1934 by a group of young French mathematicians. None of them carried the name alone; all of them carried it together. And under that name, they launched the most ambitious mathematical publishing project of the twentieth century: a series of texts rebuilding modern mathematics from scratch, on entirely axiomatic foundations."[1]
Current career structure of mathematicians works partially by looking at whether they have solved novel and interesting problems, or at least done theory-building that can help solve such problems. Many mathematicians are also motivated by being the world's first to solve such problems
Removing this measure suddenly means that academic mathematic norms need to adapt rapidly, and, even more importantly, intrinsic motivation for many mathematicians needs to change rapidly. That is understandably a sea change for the current mathematics community.
Rephrasing the argument made in the post, math was like a "take a book, leave a book" exchange where you solved a problem, and in the process you discover more problems so the pool of interesting open problems is renewed. What OpenAI and the other labs are doing is akin to
- taking all the books in the exchange (which is technically allowed, no rules about how many books you can take)
- destructively scanning them (still technically allowed, no rules about what you do with the book)
- and not leaving any new books in their place (which is frowned upon, but there's no rule that says you have to replace books you take)
This is my question too. If we are all just going “well that sucks” after AI solves this problem, why did anyone care about the problem being solved in the first place?
Is the bummer that we got a solution we didn’t want - that navier-stokes is not always applicable or something, but we hoped it was?
I think the Navier Stokes problem kind of illustrates what he’s highlighting. I think most people even before AI expected that this would resolve in the negative and that you could get finite time blow up. There wasn’t really ever going to be a situation where the resolution to this question, or really any of the other Millenium Prize problems as far as I know, gives some kind of immediate massive practical feedback.
The hope with many of these problems in math is that in trying to prove that, we get some additional insight into why it blew up that could be applied elsewhere to more general PDEs that cannot be easily controlled.
I think the observation from Tao and many others is that when humans solved these problems, the additional insights into intuition and theory building came for free since humans can give expository on what they found hard or what was their own intuition. This is much more difficult or tedious to extract from an AI model. Even when people did have access to the chain of thought, it wasn’t always very helpful to figure out what was the exact thing that made it all click. This is even more difficult how that the CoT are hidden but I would think the sort of difficulty of extracting the key ideas for a human might be worse now with more advanced models.
There’s a long term aspect to this too where we have historically used these problems as markers for the other parts of mathematics but if AI can solve it all, then suddenly this signal is not very meaningful.
Maybe to bring it closer to home. If an oracle just gave you P \neq NP, then this would be generally uninteresting since this was already expected. There’s a deeper question of why that needs to be answered. However, one would hope that creating such a separation would give us tools that allow us to create lower bounds on a lot more problems we do care about and perhaps some bigger insight onto what makes a problem intrinsically hard or easy. These long term considerations are helpful but are definitely more vague. The remarkable part is that AI is separating the part about proving theorems and the “free” insight you get.
Because the mathematicians consulted 25 or so years ago believed their solutions would lead to the greatest amount of interesting new maths to explore, and because they had been validated as being hard by being attempted and not solved for a long time.
Honestly, the attitude of the math community is a bit cringe and increasingly I think some of the elite/mystical aura is fading. Rather than a rich fertile jungle where AI can barely chomp through a fraction of the luscious terrain, one gets the sense it's a desert and all the oases are running dry.
Let's say for the sake of argument that an LLM finds a counter example to the Reimann hypothesis. Wouldn't that just create a bunch of opportunities for understanding and explaining _why_ there is a counter example?
AI companies don't share the dead ends and only sometimes a bit of the process toward success so people don't understand what was curious along the way.
> and not exactly sure how merely the existence of a solution precludes the advancement of human knowledge.
You'd be more sure if you read the tweets.
Tao's point is very simple.
1. Working on problems that AI solvers can solve is a waste of human time.
2. We have no idea which problems can be solved by AI solvers...
3. ...Because the AI labs are keeping their negative results secret, and don't tell us which problems they've tried and failed to solve, and why they've failed to solve them (or succeeded at solving others).
There are additional points surrounding it, but that is the thrust of his argument. His issue is not the existence of AI, but the anti-scientific secrecy in how it is used to solve problems. All the incentives around its current use result in closed, uncollaborative work - which while very attractive to a vulture capitalist, is anathema to scientists.
---
He also posits that having a solution to a problem is a small part of the value of solving a problem. What the AI labs are doing is the equivalent of a student turning in their homework, which has 100% of the right answers, but with none of the 'show your work' steps. Those steps are a critical artifact for doing mathematics, because the process of solving a difficult problem teaches us things about other problems.
> perhaps most importantly on what makes humans, human
What makes humans human is love, laughter, community, children, art, beauty. How are LLMs even remotely a threat to this? The piece is so hyperbolic, it's just hard to take seriously
I really feel that the anti-AI crowd is becoming a weird religion, kind of like the crypto NFT crowd was a few years ago. Most people that use AI are just like "yeah whatever, it does X, Y or Z, sometimes it sucks and I have to re-prompt it, it's pretty neat."
While the anti-AI crowd is like "I PLEDGE TO NEVER USE AI, HERE IS MY BLOOD OATH." Like, calm down. It's not that big of a deal. Some of these bullet points are just straight-up nonsense.
> I won’t read AI summaries as a substitute for reading the source material with my own damn eyes.
Author is... just defining what a summary is. Yeah, no one looks at summaries as if they're the original text. Ever. What is he even saying here?
Absolutely opposite experience. People do not shut up about it once they’ve tried it, and come up with all sorts of whacky use cases for it. Best friend suggested I ask ChatGPT whether it’s cheaper to renovate or knock down a house I bought specifically to renovate it, which I’ve told him on multiple occasions.
AI is the successor to all things crypto, there’s definitely religious fervour to it like there was around blockchain/NFT’s changing the world.
> People do not shut up about it once they’ve tried it, and come up with all sorts of whacky use cases for it.
Other than Twitter engagement baiting (which people have done w.r.t. every other new hyped up thing, whether it's drop shipping, SaaS, crypto, whatever), I really don't see this day-to-day.
> Best friend suggested I ask ChatGPT whether it’s cheaper to renovate or knock down a house I bought specifically to renovate it, which I’ve told him on multiple occasions.
I don't understand what this anecdote has to do with AI, it sounds like your friend is just a bad listener.
> What cool problems did you (not the robot) solve?
Building products has nothing to do with technical problems, just the end result. This is in contrast with things like writing a library or coming up with a new algorithm, or doing research, or even writing a technical blog post, etc.
> or is it just pretty?
People have been sharing "just pretty" things on HN for decades. It might be interesting, thought-provoking, discussion-worthy, or whatever. Even if this wasn't vibe-coded, it wouldn't be some monumental technical achievement. Your point is absolutely moot.
> Computer folders are still organized in structures set up some 50 years ago. Operating systems change. Computers change. Search changes. But this has never changed. I can't change that.
And we've been using the Dewey Decimal System for like 150 years. Both are good systems. Kepter (a visual/gallery system) would break down when looking at more than a dozen files (or when files are very similar). I don't really think it's a good idea.
What we desperately need is a system-wide (or at least documents-wide) semantic/vector search, which is basically a weekend project and a $4.99 one-time purchase.
Was trying to combine AI with generative storytelling with a card game. It was a fun experiment. To play, you'll have to get a friend to queue up at the same time.
reply