We're gonna need a lot more mathematicians
terrytao.wordpress.com272 points by srcreigh 15 hours ago
272 points by srcreigh 15 hours ago
> Before approving construction, I would want communities of humans to understand why the design works and what justifies confidence in its safety. I would hope that we all would.
Until very recently, I pored over every single line of code Claude generated with razor sharp scrutiny. I would usually catch issues with every response. I'm catching fewer problems these days. Maybe the model is just getting better, and maybe I'm being less careful while under pressure to ship more and more often. But model capability is obviously growing. Even back in March, you could tell it "give me a function that adds two numbers" and you could be 100% confident that it would write the correct function. There was almost no point in looking at the code. Since then, the complexity floor of problems in the category "this is so simple that the model couldn't possibly get it wrong" is rising, and with it, my cognitive surrender to the model is increasing too. Why check it? It's obviously going to be correct.
If AI designs a terawatt fusion plant, then of course we're going to meticulously pore over every detail to ensure safety, reliability, efficiency, whatever. If we find no flaws in the design whatsoever, will we be less careful about the second one? The third one? What about the ten thousandth one? Will "a nuclear fusion plant" become something that models couldn't possibly get wrong?
Terence Tao is arguing that the human involvement in research is crucial, but doesn't convincingly justify why, in my opinion. He says that "human agency is a value of fundamental importance" and that we will need to build "thriving human communities that can understand [AI ideas] together" - not for the sake of correctness, which AI may surpass us on, but for, I guess, the possibility of reclaiming human meaning and purpose. I don't disagree with this at all, but it's not an argument, it's a statement of values. Unfortunately, the stark reality is that if AI does surpass humans, it will become the economically dominant strategy to not verify them and not double check them, but to just do whatever they say. This seems like a great way to raise p(doom). But as the models get better and better, and as I'm scrutinizing Claude's output less and less... I just hope that there are more Terence Taos out there than people like me.
A metaphor I'm constantly drawn to is the transition from agrarian to urban societies following the Industrial Revolution. Somebody who somehow saw the Industrial Revolution coming from the perspective of somebody living in an agrarian society might have envisioned it leading to 'super farms.' And it did.
But the biggest change wasn't what it did to farming, but enabling people and societies to start doing much more than just farming, as well as enabling some great social change as well by simply economically obsoleting slave labor. And trying to imagine all of the implications of this, as well as much society might look like, from the perspective of somebody living in an agrarian society would probably have been simply impossible.
I think people keep ignoring this possibility for things that LLMs will change. There's a vast amount of the 'cognitive economy' that LLMs stand to be able to automate. And I think that will open up a vacuum in society for people to build on top of what LLMs will do (and already are doing). I don't know what that means exactly, but that's because we still live in that 'agrarian society' and trying to imagine what things will look like after the 'Industrial Revolution' is probably just impossible.
> I think that will open up a vacuum in society for people to build on top of what LLMs will do (and already are doing).
If an AI can replace me on the mental aspects of work, and robotics are on their way to replacing humanity on the physical aspects of work... then what's left? When there was agrarian societies, there were writers, priests, bankers, merchants, and laborers before and after - I really don't think things were that unclear even at the time. Now that we have machines that are close to exceeding humans in every way, what good are humans?
I've struggled with this question as well, and this is why I reject the premise of the historical pattern of technology enabling us to "move up" to something else.
At the same time, I agree with the original comment as well. I don't think this necessarily leads to some doomsday scenario. Whatever happens it'll likely be better for us and imo we will merge with the AIs at some point, so it won't be a question of us vs them.
People enjoy a lot of the jobs in the cognitive economy though. They are fulfilling. Coding and making art and media is a passion for a lot of people, the actual act, not just the outcome. So unless people were passionate about doing back breaking farm labor its not the same thing. I'm mostly embracing AI because of what it lets me explore and learn beyond what I could before, but I'm not convinced the outcome is going to be a better world at this point.
I think coding is done, as a trade practice. But a lot of jobs have sucked for a while anyway - endless api glue, js tedium.
There’s now a chance to do more with software.
I actually think art is safe. The machines don’t value it but we do. There’s something in that.
True art yeah, and people are rebelling a bit on entertainment (I don't think that will last once you can't tell). But already crushing the commercial art jobs that supported a lot of people while they did their unprofitable real art.
What is the analog of the superfarm in this case?
Or will it be the cumulative total of various advances?
I've equated Claude Code, or Codex, to the looms that made fine fabric more affordable during the Industrial Revolution; life-changing, but not society-changing. Neither the steam engine nor the automobile.
Perhaps I've answered my own question in that it's the LLM technology itself that equates to the steam engine, and it will power superfarm analogs that have yet to emerge. I'm still curious what you think they will be.
From my perspective, the superfarms are the things that we get when we try to predict the future by just pushing the present forward, but without thinking about newly emergent industries, societal shifts, and so on. So in other words, just seeing the same stuff as the present, but bigger, better, and more efficient.
So an obvious example there would be software. It's certainly true (if we assume LLMs reach their 'potential') that software will be able to reach new heights, and with a far smaller headcount driving the development. So some people see this as economically catastrophic for software developers, or an economic boon for certain large software companies.
But I think that when software can be built at the drop of a hat, software itself will no eventually no longer really matter in economic terms. Yet things you can build on top of it will matter more than ever. Those things are difficult to see from here, but I expect they will be the giants of the economy of tomorrow.
Many goods/products have seen a transition from scarcity to abundance. The analogy of fiber looms was already made. Humans used to spend a huge amount of time producing clothe. It was even used as currency in many cultures. No cheap clothing is so abundant we don't even think about where it comes from.
Salt is another one. Used to be payment for Roman soldier, now you can just take it from a McDonalds if you want.
Seems like we can have an abundance of software.
I agree we can't predict, and I explain it this way - it is like a football game, where the ball will be 10 seconds later depends on what every player is doing. It is the same with AI, what we will do with it depends on what others will do with it, the reason we can't foresee it.
But what have we built on top of software so far? My first thought is that we have made new communities. But most of them are of much poorer quality than what we had in person.
I hope we start to rebuild in person connections again with this technology.
Well it has certainly transformed bureaucratic operations. And I don't mean that in a bad way; superior bureaucracy and logistics was a key to the British Empire.
Software allows us to push computation (intelligence) into our environment.
Telecommunications did that, not really software. Twitter started out as an SMS compatible service remember.
I do think there's something there though: I've spent the last 2 days building an app I've always wanted to build for myself with my computer in the corner running Claude and Claude Remote. Prototypes land on my phone and I don't even look at it for more then a few minutes before doing something else.
It's software, actually useful software, which doesn't take 100s of hours to build.
So I didn't even really spend two days on it: I mostly didn't look at all. I'll spend more time setting the result up on my home server.
> If AI designs a terawatt fusion plant, then of course we're going to meticulously pore over every detail
But suppose some future holy grail AI can do much more than that.
Suppose it could find a cure for cancer, fix the climate, build fusion plants, Dyson spheres and so on.
But nobody can understand anymore how any of it works. We just ask and then trust the AI to deliver (as it always has).
Isn't it fun to imagine how life would look like in that scenario?
We would probably no longer care about code, engineering or even physics and mathematics among other things. We would probably mainly care about
This is how most people already live.
The average person doesn't know how the medication they take works, the mechanics of climate and climate change, how the energy they consume is generated, etc.
In the case of medicine it goes deeper than that, sometimes nobody really understands why a medication works.
I feel like we just don't understand a lot about the human body still. Only a century and a half ago we had a US president die because a doctor rummaged around and damaged internal organs (with unwashed hands, because germ theory was still not uniformly accepted) to try to find a bullet on the wrong side of his body. We've come a long way since then, but it's not like we've been doing the field of medicine rigorously for all that long. The best strategy we have seems like pretty much the same as any other scientific field; come up with ideas based on what we know, try them out, and see if they work. That doesn't necessarily give us any actual insight into why it works though.
Yet it takes rigorous years long tests to understand effects of a medication.
Yes, and understanding the effects of a medication is different to understanding how the medication has that effect.
The medical field as a whole isn’t generally interested in understanding how medication, only in empirical measuring and qualify the effects.
> The medical field as a whole isn’t generally interested in understanding how medication, only in empirical measuring and qualify the effects.
I don't think this is quite correct. I mean many practitioners of medicine will have the attitude of ... "if it works, it works". And that's perfectly reasonable.
But if you understand the mechanism of action of a drug (or other treatment), it (often) makes it easier to improve a drug.
So ... some sectors of the "medical field" understandably care only about empirical results. But other sectors would prefer to understand what's going on.
>The medical field as a whole isn’t generally interested in understanding how medication, only in empirical measuring and qualify the effects.
This is nonsense. Most, if not all professionals are interested in mechanism of action, but without Ms Frizzle, it is extremely difficult and expensive (time and money wise) to figure that out. So while the labs run the experiments with the very limited funding they have, we make do with using the second best thing we have, which are statistics.
Yes and no: understanding means to know the relationships to the underlying system.
When you already know that system well, those effects are often just a matter of simple inference.
Just like here: most people are actually perfectly capable to foresee the detrimental effects of abandoning understanding.
Living in a fantasy world of "magic" makes you dependent upon your caretakers, who provide the ingredients.
That cant be true, theres a paper trail of sign offs of the most likely people to understand the probabilities.
Please define proof of "really understands why a medication works"
It reminds me of fynemans why do magnets work. Yeah sure does anyone really understand anything? Its metaphors all the way down
We don’t really understand how NSAIDs reduce pain, as a concrete example. In contrast we have a pretty good understanding of the pharmacology of caffeine.
It’s not like magnets, some things really are gaps.
Some medications are designed for one purpose, then other effects are discovered in practice. Gabapentin, for example, was designed as an anti-seizure medication structurally similar to the inhibitory neurotransmitter GABA.
Now it is primarily used to treat neuropathic pain, and the mechanism for that is not well understood. The GABA receptor is not involved. This effect is just a happy accident, and nobody really understands why it works.
its a good thing that our lack of understanding of medicines has never killed anybody or have had any intended consequences
oh wait
A major school of German sociology places this at the core of its theory. According to this theory, modern societies are characterised in particular by the fact that individual subsystems of society reduce the complexity of their own (sub)system to the other systems to enable them to act at all; they fulfil what is called an "Entlastungsfunktion" (relief function). Prominent representatives of this school of thought are Max Weber, Arnold Gehlen and Niklas Luhmann.
In their view, it is modern institutions (public and private) which, as supra-individual entities, have long since become autonomous systems. The fact that the individual office-holders are human beings, meanwhile, is of little significance.
Hannah Arendt, in her theory of totalitarianism, attributed the effectiveness of both Nazi and Stalinist policies of extermination to the largly moral indifference of bureaucracy as a system.
In this sense, the task of controlling AI is a variation on the problem of harnessing a complex society consisting mainly of autonomous subsystems. This is a problem which has increasingly challenged humanity already for quite a long time. And it has been very difficult so far, even without AI ...
I'm going to say it's how everybody already lives, because while somebody may know some of those things there is nobody who knows them all.
Very interesting point. I think the counter-argument to that is that the complex modern society is based on somewhat “deterministic” systems, in that, even if a single decision or event isn’t rationally explainable in the moment, at least in the aftermath, it typically becomes understandable, maybe even reproducible. There is someone, somewhere, capable of explaining, maybe even multiple someones.
We don’t generally have that insurance with LLMs/AI, yet?
This seems to me like asserting everything in the universe is explicable by physics. It may be technically true, but still not relevant to understanding earthquakes.
(Don’t bother to argue this not true unless you disagree with the essence of the argument.)
Anyway, post-hoc explicability isn’t a counter-argument to the assertion that almost everyone takes almost all technology as magic, from medicine to computers.
I’m still trying to understand your argument. Are you saying that after the fact we understand AlphaGo move 37? But somehow we are never going to understand an LLM’s decision afterwards? Seems like a disconnected take to me.
Imo the counter-argument is that in principle you can research and understand the mechanics of climate change or medicine. You just don't have the free time and/or the motivation. With AI that changes - you could spend your entire life trying in vain to understand its reasoning.
It reminds me of how some religious people say that science is effectively no different to religion because we all take expert opinions on faith. But the difference is that there is a well-defined pathway to understanding, if you wish to do so.
There are huge systems supporting that ignorance though. And where those systems are breaking down people do care.
This is a very Western take though.
The problem with "ignorance" in Western countries (particularly the US right now) is that it's very common for people who don't know to believe they know and form ignorant opinions that they often want to be applied society-wide in some way. You can see this with everything from climate change to vaccines.
In much of the world, even in middle income countries, people are comparatively poor and, in my experience living abroad in such countries for many years, much less concerned with "understanding" and forming opinions about everything under the sun. It doesn't mean they don't value education and are opposed to development/progress, but it does mean that they don't question whether the vaccine they're taking is the product of a conspiracy, think too deeply about why the river is flooding more often, etc.
They just deal with life the best they can and are more focused on supporting their families, enjoying what they can, etc.
Culture and religion play into this. The way secular and Judeo-Christian people look at the world is very different than, say, Buddhists, Muslims, Fulani tribespeople, and so on.
the average person doesn’t have an internal monologue
can’t use a computer (they’ve had like 30 years now in first world developed countries)
many can’t even use their smart phone beyond calling, texting (many can’t type well), and doom scrolling (they get addicted to drugs, gambling, and other LCD activities)
many read at a 6th grade level. most can’t even calculate tip in their head.
meanwhile, the same smartphone can give them access to literally any information and knowledge the world in seconds. and now gemini can explain stuff since most ppl can barely read or think.
it’s sad out there.
but more importantly. it’s not my problem.
> but more importantly. it’s not my problem.
But it absolutely is! Those people can and do vote.
> Isn't it fun to imagine how life would look like in that scenario?
I think there's a lot of sci-fi out there that already did. Maybe it's not utopian because a pure utopia would not be likely to have an interesting story, but on the other hand, most huge technological advancements end up having just as much potential to reinforce existing power imbalances in society rather than solve them. It's not obvious to me that if we got magic super AI that can solve every scientific problem in society that it gets used in pretty much the same way as anything else: making the people who control it a lot of money rather than sharing the power with everyone without charging them.
> But nobody can understand anymore how any of it works. We just ask and then trust the AI to deliver (as it always has).
> Isn't it fun to imagine how life would look like in that scenario?
This is horrifying to me.
Rightfully so.
People succumb to defeatism and acquiesce to regressing to zoo animals, with AI as their caretakers.
They simply cannot help but to apply the economic gauge of short term profits to value the alternatives.
Even though, obviously, here long term human survival and living conditions are at stake, necessitating an entirely different set of considerations.
Some of it has to do with the fact that many people pushed forward training themselves in a way they understood would better themselves for the future. They put in time, and now their sense of value is in question. They're forced to think about things like value (of life, themselves, and hopefully others), and the things they want to remain valuable in the future.
It's interesting to me that you only mentioned the people using 'AI' in the 'short-term' ways, and not the ones that use it to better themselves in the 'long-term' ways. You can spend your own time focusing on either usage, it's really up to you and your concerns. Either group's sense of value is what determines their behavior. Where they spend their time and thinking must be elsewhere, and you disagree with it. Who judges the quality of time spent? You do.
Is it more useful to think about self-improvement, and how to navigate the future in ways that might help you re-establish value of yourself, life, and others? Acquiring knowledge is a struggle, the author mentioned this. There is also Plato's Allegory of the Cave, which highlights some of that struggle, a resistance to change. And we're all limited by time, our genes, our station in life.
The only way to help anyone out of the cave, is to help them believe something different about themselves. To help them believe there is good reason to spend time going deeper into knowledge, or at the very least, allow others with the passion and station for it to do so.
If not for the very least reason that it keeps us 'in the loop' of some central idea behind intelligence (prediction?). Or because we feel it keeps us safer, as a fallback measure because we acknowledge we have to trust other's knowledge to exist.
Here's the thing: nobody is stopping you from putting in the time to understand all that. The problem is, nobody has that much time, and we get hungry. And so we want the community to move with us, spend the time the same way as we do, to ensure value. We are all saying: we want someone else, others, to put in that time for us. The truth is, we all want quality, and value is closely related.
I can almost guarantee you the first time you show cancer symptoms, you won't care whether the cure came from an AI or human's understanding. But we haven't seen that, so we can't make the judgement call.
We would care about having fun.
For some people, fun is doing physics and mathematics. So they are going to keep doing that.
> We would care about having fun.
For even more people fun is TikTok, Snap, Instagram -> sounds like a collapse of a civilization to me if you increase the ratio even more towards dancing kids sharing their content non-stop with no added value to the society
It's not fun, it's a physiological addiction, because we learned enough about our brains to hack dopamine and reward cycles for profit. It's literal abuse.
> We would probably no longer care about code, engineering or even physics and mathematics among other things
This sounds like a boring existence. I take your meaning, but want to point out that not everyone learns about things because of practical utility, some of us find it incredibly satisfying to learn how things work just for the sake of learning.
Sure but no one's stopping you? I know a great many things of no direct practical value to me.
There's also things I don't know and don't have the time to learn which are very helpful to have AI do for me: web interfaces are really useful and I look forward to them now working exactly how I want. I'm not ever going to regret not spending more time trying to figure out how to center divs or which framework I should use because they're all deprecated.
Well there is this one framework, CSS, that is quite alive and well :)
But point taken.
Before extrapolating that far, take a look at the frontier labs' own job boards (https://openai.com/careers/search/?). Isn't it curious that they are still recruiting human "Android Engineers", "Account Associates", "Consumer Marketing Leads" instead of automating them with their world-beating models?
You’re describing a world where no human being has any agency. Curing cancer etc. sounds great but we’d be losing something priceless in the exchange
Ian M Banks covers this in great detail in his science fiction books. Highly recommended.
I’m reminded of themes from the Hyperion Cantos! Maybe my mind is over-connecting, but it’s not the first time I’ve drawn similarities in the last few years.
It’s terrifying to me to think we’d let AI make things for us we never understand. Like livestock not knowing how auto-feeders dispense their daily food were built and appeared, they just gladly eat until…
I think the weirdest part of that scenario is that science might become something closer to archaeology
Our best philosophers have already pondered this question, and showed us the answer in the form of humans on the Axiom starship in WALL-E.
> Isn't it fun to imagine how life would look like in that scenario?
Look at the financially desolate subcultures with no option for advancement or dignified life.
That is the goal and that is how it will lool like, if the tech CEO managed to gain the power they want.
Those subcultures are like that because they are small. If that became the lot of most people, society would look very different. The choice would be between Fully Automated Luxury Communism and Oligarchic Hellscape.
But where are the “cure for cancer” papers for it to ingest and then give to us?
> We would probably no longer care about code, engineering or even physics and mathematics among other things. We would probably mainly care about
… about what?
It's about as fun, and as realistic, as imagining magical ponies and people with superpowers?
This is such an incredibly naive and absurd vision; we've already proven that humans are very often very bad at implementing other humans' good ideas. There's nothing that AI is likely to bring that will improve this discernment.
>Terence Tao is arguing that the human involvement in research is crucial, but doesn't convincingly justify why, in my opinion. He says that "human agency is a value of fundamental importance" and that we will need to build "thriving human communities that can understand [AI ideas] together" - not for the sake of correctness, which AI may surpass us on, but for, I guess, the possibility of reclaiming human meaning and purpose. I don't disagree with this at all, but it's not an argument, it's a statement of values. Unfortunately, the stark reality is that if AI does surpass humans, it will become the economically dominant strategy to not verify them and not double check them, but to just do whatever they say.
I think you have a fundamental misunderstanding here, and it's not really explained because I think it seems self-evident from within the field. In short: writing code is a means to an end; doing mathematics research is not, but is the end in itself.
The human involvement is crucial because the entire purpose of mathematics research is to increase human understanding of mathematics. It is pursued because it is interesting, not because it is economically useful. In this sense it's a lot closer to the humanities.
A black box oracle that just tells you whether statements are true or false is not the goal of mathematics and would not be particularly interesting to the field (except insofar as it could be harnessed to improve human understanding).
Coding is totally different from this, where it is essentially always done as a means to an end. Likewise with many other fields, like pharmaceutical research or materials science or what have you, that are oriented around solving problems for some practical purpose. Pure math isn't really like that for the most part.
I think you have misunderstood the OP's point here. You're arguing that deepening human understanding is an end in itself, and you are right. The OP is arguing that advances don't need to be pegged to human understanding, and they are right too. The two can coexist, superintelligence far ahead of us, pioneering discoveries - and mathematicians catching up at a pace suited to biological minds. I don't see the issue here. Of course, it does mean mathematicians adopt a new role as hobbyists.
> A black box oracle that just tells you whether statements are true or false is not the goal of mathematics and would not be particularly interesting to the field
This is a crude distortion. The recent breakthroughs have come with proofs, reasoning and verification, and there is no proposal that I'm aware of that would do away with these foundations. There's also some rather ugly solipsism in the idea of keeping what interests the field as a limit. Mathematics has broader relevance to humanity than merely to please and support mathematicians, and if other fields can make practical use of profound well-proven future math, mathematicians will have a hard time making a case that their comprehension must come first.
Superintelligence is old hat already: we're now racing ahead at full speed towards Super Duper Intelligence!
> > A black box oracle that just tells you whether statements are true or false is not the goal of mathematics and would not be particularly interesting to the field
> This is a crude distortion. The recent breakthroughs have come with proofs, reasoning and verification, and there is no proposal that I'm aware of that would do away with these foundations.
I believe that we are still at the point where these proofs serve as verifiable certificates of correctness, so that it's not a "trust me bro" situation, but where humans mostly still don't find them understandable, so that they are still just a highly reliable black box.
Sure, but I don’t think most of the money that goes into funding math is for the purposes of pure understanding. The reason governments fund mathematics research grants is generally for a more instrumental purpose; taking the US congress as an example, the mission of the NSF is to, “Promote the progress of science; advance national health, prosperity, and welfare; and secure national defense.” Most federal math grants come from the NSF.
Of course, math research is cheap and most academics don’t rely upon grants, their salary covers most of their expenses. But here too, the mathematics professor spends a substantial amount of their time teaching future engineers/quants/other applied mathematicians, who need to understand math for instrumental purposes, not as an end in and of itself. Without the tuitions of these students, I can’t imagine universities maintaining the size of their math departments, let alone expanding them as Dr. Sahai advocates for.
So who or what funds the community of pure mathematics going forward?
Research in pure mathematics is part of what we call "basic research". There are no applications in mind a priori. People instead focus on understanding, because history has taught us that understanding tough problems in mathematics finds natural applications elsewhere. It's the same as theoretical physics or theoretical computer science.
> People instead focus on understanding, because history has taught us that understanding tough problems in mathematics finds natural applications elsewhere.
If the goal is still eventually the applications elsewhere, we're back to what happens if the AI is simply better at this.
You can probably make an argument that human understanding is better as humans are better at finding new patterns or fundamental new ways of thinking and also applying them to new applications.
However, what if AI becomes better at humans for that as well?
No reason you couldn't have an AI be optimised for advancing basic research and understanding and a second AI to take these results and optimise for finding new applications for these discoveries.
Basic research is funded with the understanding that applications are not imminent, yes, but also with the expectation that some of the knowledge gained will eventually result in advancements to the public welfare
Another day, another HN thread full of programmers who think mathematics is just like programming.
Thanks for providing a (much needed!) correction.
Nit, but Terrence Tao did not write this article, it’s a guest post.
I’m not sure it’s really a nit. The article says that explicitly at the very beginning, and implies that again at the very end.
I’m not sure what to think about an analysis written by someone who didn’t catch THAT.
> Maybe the model is just getting better, and maybe I'm being less careful while under pressure to ship more and more often. But model capability is obviously growing.
If you are a normal person research (e.g. https://arxiv.org/html/2606.22721v1 but there are a lot more, not necessarily on coding) has shown that you indeed are being less careful. It most likely also works better simply because more resources are being poured in.
Having a big spike on your steering wheel would make you drive more carefully. More broadly, I think if people put in less effort without massively increasing mistakes, it should come out fine. Formal methods might improve AI product safety as well, especially in some safety critical applications.
LLMs dont create anything new, if programmers stop reading the code technology will be forever frozen to 2022, no new programming languages, operating systems, concurrency primitives, databases, networking protocols, UI frameworks everything will be based on the training data and future generations will forget about all the primitives we now take for granted.
If someone creates a new programming language/ framework or new better way to do async or whatever, no one will use it because it is not in the training data and it wont take off because everyone is using LLMs. It will be like using the same Lego pieces over and over.
LLMs receive new data via input context, not just training data.
Thought experiment: How effective will 2026 LLMs be for humans in 2526?
It's not game over just because 500 years are missing from the training data. The important question is how well can 2526 humans make culture and knowledge navigable to LLMs via tool calls.
Today's LLMs might need for example sub agents to translate to 2526 English, sub agents to read 2526 docs.
It's _really not clear_ whether 2026 LLMs will be useless. To believe that reflects an enormous misunderstanding.
> LLMs receive new data via input context, not just training data.
Be more specific about the "new data". If everyone is using LLMs for work (generating code), especially the juniors who won't get the chance to learn from first principles, LLMs will be training on the data they generated. How will new code enter the system at large enough quantity that it can be used for training?
> It's _really not clear_ whether 2026 LLMs will be useless. To believe that reflects an enormous misunderstanding.
They won't be useless, they will just be frozen knowing only whats in their training data. No new programming languages will emerge, in 2526 they'll still be using Rust and javascript, same exact code from 2022 which dominates the training data.
What if programming languages, operating systems, concurrency primitives, databases, networking protocols, UI frameworks are already good enough, and the innovation lies elsewhere?
You can do a lot of cool stuff with the same lego pieces.
That is like saying what if music is already good enough.
Totally OT. Any advance in music in the last 100 years?
What does an advance in music even look like? Shifting tastes for pop music? Or new techniques? New music theory? Or just experimentation?
Considering all music is subjectively influenced by the culture in which it's born (see the difference between Asian traditions of music, European traditions of music, African traditions, and traditions of the Americas) not even all of those have a given structure that is present today like the typical 4/4 and have polyrhythmic and multitonal structures by design. The fact that everything on the radio has converged towards 4/4 165bpm major chord progressions is evidence of that cultural phenomenon.
Metal vocals are still advancing today. Check out Will Ramos doing harsh overtone screaming.
But what if the fundaments of all these, in the human produced literature, actually contain hidden circularities and holes which make very hard the progress?
IMO for the moment the greatest value from these AI tools is that we can start an audit and hopefully proceed on a saner foundation, after we use the tools and think about it.
This is different than too many AI generated proofs or panic reactions from the academic system with its stupid incentives.
This is obviously false, and the same silly arguments were made back in the day with Deep Blue and AlphaZero.
False dichotomy. Chess/Go can still be played between two humans and there is allot of value in that because humans compare each other to other humans, when you see a skillful Grandmaster play you know they are good compared to yourself or the average human, that is why people still watch, play chess/go and train hard to get good. Programming is different because you are creating something not necessarily trying to win a game.
Most programming tasks are exactly like that. Is this agent able to complete this task? Is this agent able to optimize a kernel beyond previous attempts?
Of course some are subjective and that's where progress is harder, like "Is this website pretty?". But for tasks that can be objectively measured, LLMs will go beyond human level, just like with Chess and Go.
That's why RL is so important when training LLMs.
My point is that LLMs depend on training data so the code they produce will be stuck in 2022, no new languages, techniques beyond that because new techniques are not in the training data (at least not enough of it for training because most coders are now using LLMs).
Chess/Go continues to progress because it is primarily a human vs human activity, people will always be learning to play chess and chess will continue to develop.
> Chess/Go continues to progress because it is primarily a human vs human activity, people will always be learning to play chess and chess will continue to develop.
AIs are not continuing to get better at chess/go because humans continue to play at levels far below themselves who discover new techniques. They get better because they play against other AIs and discover new techniques that have a higher win rate that way.
I would bet that even if humans stopped playing chess/go and people were still willing to run these AI models against each other they would continue to get better.
I am not talking about the advancement of AI, I am talking about the advancement of chess.
Two things can be true AI drastically contribute to the advancement of chess and humans playing against each other also contribute (even if slowly) to the advancement of chess as it has always been since the invention of the game. The point is that because chess is primarily a human vs human game humans will always have the knowledge of chess, unlike with programmers who are giving it up to prompting, and programming being much more complex than chess (checkmate and win) will be stuck in 2022 because of the training data.
Pre training data is in large part synthetic these days, and RL data is almost all synthetic.
Computer Chess progress has nothing to do with human vs human activity. AlphaGo Zero used no human game data at all.
> Pre training data is in large part synthetic these days
How much of that data can lead to innovation? Can you predict all innovation map it out on paper.
> Computer Chess progress has nothing to do with human vs human activity.
The point is that humans will always be learning chess because it primarily a human vs human activity they will be contributing games to the chess database, unlike with programmers who are stopping to code and only prompting, generating code stuck in 2022.
> AlphaGo Zero used no human game data at all.
Sure, but that instance of AlphaGo is still dependent on its training, its intelligence, so it is a question of is that the best and only way to win a game of Go. Just a few weeks ago, a Go Grandmaster found a way to beat one of the strongest Go AIs.
So a specific instance of an LLM might be the smartest based on what we know and need today but that is not the limit of how far we can go, this is why it is important for humans to always have an intimate connection with the code, math, science, chess etc for progress to continue.
> Sure, but that instance of AlphaGo is still dependent on its training, its intelligence, so it is a question of is that the best and only way to win a game of Go.
If this were true then it would be impossible for these models to ever exceed the top human level as there would exist no training data that allows them to exceed the top human level.
However, despite there being no training data on ability to beat the top humans these models have achieved it.
> this is why it is important for humans to always have an intimate connection with the code, math, science, chess etc for progress to continue.
This is just you wanting to remain relevant rather than actually based on evidence.