This is Fisher/Box feedback loop (https://www-sop.inria.fr/members/Ian.Jermyn/philosophy/writi...) implemented on a modern computational system. LLM is just a component. I wish Sutton had commented on this fuller picture of what we have now instead of commenting just on the LLM/Backprop side of things. I am honestly curious of whether such a loop can at least partially automate discovery.
There are more elements to discovery though. It is still not clear where the initial working model/hypothesis comes from or how the updates are selected (unless it is just parameter induction). I recently read about Hanson's Patterns of Discovery which aims in that direction. I have still not read it, but I am curious if it has any mechanistic clues.
That is a problem in RL, so we usually do supervised training first, teach it to imitate some trajectories, then do RL to refine the model. RL alone has a huge problem because it might be hard to reach a reward, hence hard to learn the task by pure reinforcement. Humans also combine supervision (learn from books) with search (solving problems) to break the discovery problem. For example, a human with no initial instruction in math would not produce great results no matter how smart they are. The bootstrap was exploration paid for in the past.
If I get your meaning right, SFT creates the right inductive bias so that the RL search + reward guidance does the trick.
For novel discovery, the question might then be whether the inductive bias builds a strong enough prison so no new discovery is possible by RL or if the search can escape the boundaries set by SFT given enough randomization and the right reward function.
I know that RL is usually not performed at inference time, but in-context learning mechanisms might be developed by RL to discover at test time. Edit: I would love to hear if that actually happens or not, like new induction heads (https://transformer-circuits.pub/2022/in-context-learning-an...) forming during RL. I really have no idea.
The problem I see is the same problem Evolutionary Algorithms had: you can generate potential solutions until you run out of cash, but you still need to evalulate those solutions. You need a fitness function, and that means you need to at least know the general shape of the solution. If anyone knows of any work towards more open-ended fitness functions, I'd love to read it.
So all that is to say, I'm not sure it is even theoretically possible to create a single algorithm to do open ended search and evaluation. Biology has billions of years of evolution and accumulation, whereas a simple algorithm in a computer, even if smart and connected to the real world, has no such accumulation.
I think humans hit the perfect sweet spot where we have the simplicity of the self preservation instinct, but we have the complexity of the cortex and lots of degrees of freedom because of it, plus on top of that we have a lot of accumulated degrees of freedom in the society and technology and knowledge that have we, which has been built up for thousands of years, all of which we can't just create an algorithm to encapsulate without going through the actual evolution.
And just to make it explicit - a large percentage of what humans think derives from an instinct to preserve the self, the mind, the future and the environment, even if it is very abstract at times. Not absolutely all, but I think a good chunk. And the complexity and degrees of freedom comes from that we have so many neurons in the brain, and a complex body with hands and whatever else that allows a lot of behaviors, as well as a complex environment that is constantly challenging us.
There is research in open-ended learning, see "Why Greatness Cannot Be Planned" by Kenneth O. Stanley. The core idea is that in open-ended scenarios you don't know what action was good except in hindsight because your path is deceptive. So the idea is to replace fitness with novelty search which provides more stepping stones towards the goal.
Those mathematical shortcomings very well might mean they arent a path to true AGI, but that honestly seems fairly irrelevant at this point tbh.
Can you expand on this? Because when I think of it, it calls to mind p-hacking and selective publishing more than anything.
Like, when you try prompts over and over and over until you get code that finally works. At which point you stop and pronounce the dubious claim, "AI is amazing, look what it does!"
i.e, evaluation, retention yes. variation or "planning" no.
That is not to say you cannot use LLMs. Alpha evolve does exactly that. It uses an external simple evolutionary planner. The overarching point he's making is that our planner is still "dumb" and we need to work on it.
When you iteratively guide an LLM in claude code, you are the external planner. That also works.
I really like the way he frames this here. I think a lot of people in the twitter comments (and maybe a few here) aren't reading past the introduction. He isn't saying AI systems are incapable of creativity and discovery. He is claiming generative AI without a harness is not capable of creativity and discovery. There needs to be some other system that "recognizes the value" of the novel idea and remembers it. He gives examples of where this value recognition step is automated and thus by his definition achieve creativity and discovery in a fully automated system.
i.e, evaluation, retention yes. variation or "planning" no.
That is not to say you cannot use LLMs. Alpha evolve does exactly that. It uses an external simple evolutionary planner though. The overarching point he's making is that our planner is still "dumb" and we need to work on it.
When you iteratively guide an LLM in claude code, you are the external planner. That also works.
Seems clearly false. Pretraining finds the mean/mode of the data distribution. RL can easily generate many samples around that mode, evaluate them on an external source of truth (eg compile the code and run it) and then selectively train on the good samples. This clearly can go beyond the initial data distribution.
I'm not saying this makes it useless - it clearly helps for math and coding tasks. But the ceiling exists and that's what the original tweet was referring to. Alpha evolve also shows what lies beyond the ceiling, altho their planner was rudimentary.
Would you agree that it is a matter of degrees rather than a qualitative distinction? There seems to be a broad misconception in Sutton and others that output quality cannot exceed that of the base internet distribution; my point is that RL allows you to easily produce an output distribution that is better than whatever data you trained on according to some evaluation criteria. There are no clear theoretical limits on how much better it can get, rather there are many people asserting guesses that there is an upper bound and it lives below "human creativity." I just haven't seen any solid theoretical argument, and the empirical evidence has so far shown continual improvement.
Also, I would be keen to look at any sources you have of pass@k not improving much during GRPO.
Again, no one is saying models can't improve beyond the internet i.e data distribution! They clearly can. The claim is that RL without real exploration cannot exceed the base models distribution, which by virtue of SGD _does_ generalize.
And also, it doesn't mean it's not useful. Improving sample efficiency and making something that happens 1 in 15 times happen 1 in 1.2 times is insanely useful and is what has enabled the kind of coding agents we have today.
Sutton, especially, I doubt has a misconception about this :)
> pass@k
Yeah, AFK now. But it's a well researched thing. You can look for more, but here's one off the top of my head: https://openreview.net/forum?id=4OsgYD7em5 The original deepseek paper also had the result, i.e the paper that first got famous for using grpo as a method that works for LLMs. A side result in one of these papers I forget which one, is that the base model converges in performance with the RLd one at high k.
So perhaps the right framing of your/Sutton's claim is: RL can upweight low-probability (p) but correct outputs, but there is a limit to how small p can be, and it is on the order of 1 in a 100 or 1 in a 1000. Implicitly there must be some crossover point where you would call this discovery/creativity if it works for sufficiently small p right? Eg if RL can upweight a correct but 1 in a trillion output to 1 in 5, that's got to count as discovery given that all possible sequences are technically "in the distribution"?
In practice, it does seem like that kind of progress is happening. For example with the recent Erdos solution [0], I would wager that GPT 4's hit rate on this would have been functionally 0 (certainly less than 1 in a thousand). Curious to hear whether you'd still say this is mode-seeking within a base distribution, or if not then what is the right explanation if not iterative RL.
I'd also highlight that the paper you linked with the pass@k equivalence doesn't technically address the question of how small p can be before RL upweighting breaks down - all of the example problems were easy enough that the base model had decent hit rate with 128 tries.
[0] https://openai.com/index/model-disproves-discrete-geometry-c...
> Discovery / creativity
I'm absolutely uninterested in the semantic discussions of what is a real discovery, what is creativity, what is intelligence, etc. I simply don't care. If it's useful great use it. If it's not great don't.
> How small p can be
All that depends on your sampling procedure. If you intentionally smooth the distribution out you can sample the smallest thing, but you pay for it with noise. Taken to an extreme, this is the monkeys typing on the keyboard argument.
It's a mathematical fact that RL cannot improve things it doesn't sample. In any learned distribution you pay a heavy cost by sampling far away from the mode. Most RL algos sample rollouts maybe with some smoothing but that's it. This is why external planners are necessary in order to sample something effectively un-sampleable in the base distribution. Simple example: tool use!
Sutton and everyone are simply calling for a focus on improving these external planners in the same way, as they also enable much better "continual" learning and so on.
> Erdos solution
The RL was what enabled such a huge trajectory to ever become efficiently sampleable in our lifetimes probably. You can do many useful things like this and more purely with the base model distribution.
In fact. Doing RL on user chats and so on especially from pair coding sessions are improving these models coding abilities by a lot making them even more reliable for SWE. In this regard, mode-seeking is a win.
> All sequences are technically in distribution
If it was truly improving 1 in million things systemically, then you wouldn't see base getting the same results given many samples. Albeit they are not erdos problems.
Could it be that at 1T scale, and for difficult problems specifically, grpo somehow filters through the noise and picks out the 1 in trillion? Extremely unlikely (you have your expected rollouts required to sample that, and then you have your sparse reward signal and no credit assignment on top of that...). But of course, only 2 companies in the world can do experiments with it, so there could be some unknown effect the rest of the world has not seen. Barring that, no.
I wonder if this is a precursor to Keen Tech leaning into David Silver's Ineffable Intelligence approach.
Its the secret sauce behind why the current models are so great at coding and soon to be unbeatable at math.
LLMs can pose many questions and if they are easily verifiable, fine tune very heavily. A lot of the world models discussion will inevitable lean into simulations as verification.
It's RLVR tuned, but not to the ChatGPT level of brain damage, and it's still backed by a fuck off huge pool of model weights - which matters for what you call "touchy feely stuff".
AlphaGO is given a hard evaluation externally. It did not itself come up with it.
When GAI models are given an external hard evaluation, they can also succeed in many different domains (that is one of the remarkable features, succeeding in many domains) ranging from simple programming tasks to frontier mathematics (disproving conjectures recently) to writing more optimized kernel code than before.
And there is plenty of RL especially in these fields where the solution may be extremely complex but eval is rather less complex. And even the discovery and the "evolution-like" trace-selection is also happening.
For this reason it seems strange to compare it to AlphaGO as alphago is given a hard eval independent of itself, from an external source (humans) in a narrow domain. If GAI is given such, it can also show some remarkable results.
But what I find more strange is that innovation and moving forward in many many many cases does not require truly novel ideas but instead a high-quality execution of layering different methods, tactics, ideas on top of each other. Because in many domains our collective knowledge is incredibly sparse and complex, something being able to recombine tools, models, ideas in a high quality way (as he mentions being selective) I think is extraordinarily powerful. And in such cases, with a finite exploration horizon (time, resource available) with 1% "good choices" vs 3% "good choices" are worlds apart, incomparable.
Most importantly: none of the above is about intelligence, it's barren solution-farming to important, valuable problems we have. Most of the AGI and intelligence-related debate seems to miss out on this simple fact. (Insert the usual stuff like a plane being unable to fly like a bird or a submarine not swimming is totally irrelevant to it being useful).
And then a final point: do we really think this thing is incapable of doing better on average on problems we average people face in our lifetime? What should we think, how should we define human intelligence when we give out degrees in science or medicine for 60-70% exam results on problems considered to be generic in the field?
Just a brief reminder that planes have wings with airfoils just like birds and submarines have air tanks just like fish have swimming bladders.
Some birds fly without flapping their wings much, too, e.g. albatrosses.
Leonardo tried to create flapping wings.
If it's a), he doesn't propose such an algorithm, and I don't know how you'd do it at such a low level because how do you quantify abstract goals? Did he suggest such an algorithm and I misread? If it's b), that already exists, see AlphaEvolve or any number of things he said. Or, to be a bit of a smart-ass, just type /goal and let it rip ...
I also think he's just categorically wrong that LLMs cannot do good and novel things. And if it can, then you could just say "well that's not novel, that's derivative". A simple example, if I make up a programming language with an LLM and it works well for my purposes, then is that not novel and good? I mean, is any language other than FORTRAN not novel?
Everything is derivative and you can put an LLM in a loop to evaluate LLMs trying things. I must be misunderstanding because he's too smart to be this wrong.
AlphaGo uses discovery when it evaluates potential moves and iterates.
Claude Code uses discovery when it generates a script and the evaluates whether it works or not.
He’s saying we need to allow ai systems to do the evaluation and iteration themselves for science and engineering the same way we do for code.
Basically, harness engineering for engineering.
https://youtu.be/ThFq87Rp21s?si=SrKj72_X8bjnB6ED
Around 35min mark
But I think humans are better at it, while ML is better at algorithmic thinking. “Better” being more efficient and something we more enjoy doing; we can also more accurately rank what subjectively appeals to humans (i.e. taste), especially ourselves.
I think ML should be optimized for tasks that require more generalization than programming, but are still mostly logic. Like software development, translation, and tools for art and discovery.
Can A.I create art. Well it can create something that's pleasing to our senses but art is ultimately about conveying human feelings and emotions. Even as humans, understanding art is not universal. "feelings and emotions" and therefore art, can be deeply tied to a particular groups shared beliefs and experiences.
Can it be creative in non-subjective fields such as math or sciences. Einstein derived GR from his creative thought experiments. If an A.I poped out GR's field equations simply by testing different mathematical frameworks that resolve the issues discovered by experiments, is that creative? Perhaps but certainly not in the same way.
Now if the question is, can a machine make art, well ultimately someone needed to turn the machine on and design the machine to make art, so arguably that person/people are the ones making the art.
Historically, every question of "is x art" ends up having the answer "yes". I don't know why people fall for the same thing over and over.
you made a small error, art is mostly about generating an emotion in the viewer/listener/.... not about transmitting an emotion of the creator
the Wikipedia page on art starts with:
> Art is a diverse range of cultural activity centered around works utilizing creative or imaginative talents, which are expected to evoke a worthwhile experience
https://en.wikipedia.org/wiki/Art
so AIs can do art, because they are only required to generate an emotional response in the receiver
Also lol @ using Wikipedia as some source of absolute truth.
The practical problem is that models have very limited prompt adherence. The level of detail you can specify in scene design is very crude. So you can get the slop effect where there's a lot of in-fill pastiche detail, but you could never create something like this, where all of the incidental objects are specifically included to enforce the message.
https://en.wikipedia.org/wiki/The_Awakening_Conscience
It's basically the professional version of the "Draw me a pelican on a bicycle" problem.
There are situations where you want that level of creative control, and current image generators don't get close to it.
And without it you can't get to the meta-creativity level where you're creating a new aesthetic that's a cultural landmark - which is what the famous artists did, and still do.
I gave this approach a shot over the first few months of this year[1] (although my director didn't have any custom training). The results were interesting, but I'd not call them "art", since they're low-quality derivative pieces. With reasoning traces enabled, you can see that there's not much intent going on. Though they do attempt to include "incidental objects" to reinforce meaning, like in this jungle scene[2].
[1] https://news.ycombinator.com/item?id=48105385
[2]https://www.liamlaverty.com/paint-by-language-model/inspect/...
Recent image models are advancing rapidly at prompt adherence specifically, and being able to iterate on the same image is propelling them even further. Images 2.0 being the poster child of this "agentic iterative image composition" approach.
It's the opposite of a skill issue. No image generator is anywhere near the ballpark of pro-level manual Photoshop or Illustrator editing for individual elements in an image.
If you don't understand this, try precisely kerning the text in a generated book cover to handle letter combinations like A and V.
This is one of the big problems with GenAI. You can do new things with it, but it's crude Dunning Kruger good-enough-if-you-don't-ask-for-more creativity.
The pros can see what most people can't, and the flaws and missing features are frustrating and obvious creatively, not just in terms of production values.
We went from "AI can't generate text that isn't at least 20% typos and it always looks like shit" to "some letter combinations aren't kerned to perfection sometimes and adjusting that with prompts is hard". In a couple of generations.
They just want dopamine. No thinking because that hurts.
Even for humans the brains who managed a step change in thinking are so rare that we literally know them by name.
His main point is that discoveries involve
1. Variation,
2. Evaluation, and
3. Selective retention.
He makes a jump saying AI is only capable of 1) and humans are capable of 1) 2) and 3). I don't know what makes humans special enough that they can do 2) and 3)?
In fact, the more you think of this it is kind of strange - in science humans can only do "evaluation" because they have access to the real world. They can evaluate a new drug because they can do it on people so it is not some inherent limitation of AI but rather access to physical realm.
Finally I want to ask a specific thing: how do you mathematically falsify what this person is saying? How can you formally prove that - no AI can not "evaluate"? I ask because I make AI evaluate a lot of people's claims and it works for me.
> This is the weakness of deep learning that is alleviated with a new algorithm that my group presented in Nature a couple of years ago. Our “continual backpropagation” made one small change: every so often a less-used neuron would be re-initialized to small random weights. This allows the variation to continue and plasticity to be retained.
Here's the paper: https://www.nature.com/articles/s41586-024-07711-7
It has a fair number of citations, but I haven't looked into how much it's used.
Kant said something like this: knowledge can’t be obtained by pure thinking, it needs interaction with the world.
This is obvious to me so why is the author making a claim that LLMs can make knowledge without access to environment but purely through thinking in aether
But so do humans? How do humans make discoveries without having formal ways to evaluate? In my pharma drug example, humans could evaluate only because they had access to the physical realm.
I can’t think of an example of humans evaluating a discovery in a way that LLMs can’t. can you?
Access to some forms of evaluation and selective retention is inherent to humans and it's not inherent to LLMS. But it can be somehow bolted on and that's when they work best. It makes sense that more focus on those principles can yield better AI. I think the retention part is the real limitation of LLMs, because it's limited to stuffing things in context window.
I'm not sure I understood - what forms of evaluation is inherent to humans? If you don't give humans tools or access to the physical world, how can they evaluate?
Its not too interesting.. we already know that giving AI access to compilers and tools make them better.
And the core point is not even true. They can definitely output novel things that are good - less so but they can and they do. Plenty of examples.
> Thus, the trajectory is either novel or good—based on randomness or based on data—but never both at the same time.
This assumes no possible unexplored path yields good results, or said another way, that none of the random results can be good, which is not true. The whole text seems to try to prove a point decided a-priori rather than make a case based on reality.
That contradiction kind of says he doesn't know what he's talking about.
He is saying no generative AI is going to produce output that is both good and novel because it is always derivative. And then adds a generative AI (Claude Code) into his list of AI that have produced output that he feels is good and novel, invalidating what he is arguing.
"...no matter how many instances of white swans we may have observed, this does not justify the conclusion that all swans are white."
Although personally I think code doesn’t actually need to be very novel so it’s actually the best example.
If you don’t agree with somebody, nothing else matters
It’s like people (you as an example) have taken the concept of experts and fucked it up so bad that simultaneously everybody thinks they’re an expert while also dismissing everybody else who claims they are an expert
It’s like the whole concept is entirely poisoned. Worse everyone is smugly pointing at the Wikipedia for “appeal to authority.”
Nothing new I suppose, Socrates after all was driven to suicide by the madness of his society accusing him of impiety.
This doesn't seem true? You can be both random and based on training data.
A joke is just an "error" - your brain predicted something, and the butt of the joke goes in another direction, and it's the mismatch that makes it funny.
The same goes with creativity, and intelligence.
The problem is that, by design, while trying to make machines "reliable", we make it impossible for them to be intelligent and creative
The term "surprisal" is used in information theory:
> For a given probability space, the measurement of rarer events are intuitively more "surprising", and yield more information content than more "common" events.
Can a machine surprise us? Given enough complexity, I think so. They can produce unpredictable results, even novelty, something we've never seen before. But does that mean a machine can be creative? Or funny? Maybe there's a threshold of acceptance, where eventually its output will become surprising enough that we might as well call it creative.
(Currently returning 502 "Bad Gateway" for me, but should be restored at some point.)
Still about ten million discussions to go.
> Claude-Code, which have brought true advances in ... programming. ... these systems have found things that are both novel and good.
I don't think I would attribute anything in that process that I would consider an AI to be incapable of.
The characterisation of variation like this would seem to rest on the same 'random but directed' crutch that some free will arguments rest upon.
There is no random but directed of course, there is random and there is caused, and there are things that use both as components, but the random remains wholly random, and the caused remains entirely deterministic.
I think there is a good case to say that, in many fields, AI is better than humans at evaluation.
To find avenues to consider, I'm not entirely convinced that human innovation is more than a heuristic that appears more chaotic by virtue of a inconsistent and opaque formulation.
Many aspects of ideas com from noting how some two things are different and then considering that axis of difference when applied to another thing.
The possibilities thrown up by this extremely simple method are vast enough to require multiple layers of evaluation, most could be dismissed out of hand by a quick 'This is nonsense' check that I suspect people do so often and at a rate that it wouldn't even rise to the level of consciousness.
The point seems to be that generative AI just generates stuff, and that real discovery requires variation, evaluation and selective retention.
The call to arms seems based on the assumption that people only every talk about generative AI as discovery machines themselves. I think it's pretty widely accepted that's not the case by everyone apart from cliche out-of-touch CEOs.
But the talk makes me realise that generative AI are incredible tools to do the discovery cycle with, and this is what I imagine professionally successful AI users are doing: variation, evaluation and selective retention of their inputs and outputs to generative AI.
Then it shifts to discovery.
These seems related but not exactly the same thing.
Best thing about nerds is watching them try and build frameworks and formulas for the creative act. Like a metronome trying to compose a symphony.
Should we automate exercise and play as well? How about learning?
The machine didn't have a soul, so we donated ours.
Eureka! My AI found it!
- Fusion (a clean sustainable form): Without this I think we are heading in a very wrong direction, whether it is conflict or climate change does not matter. Everyone is aware of this and instinctively afraid of the implied loss of quality+quantity of life.
- Cure for Cancer: It is a world wonder even in Civ. I and for good reason. As a father of a teenager, every time I hear a story of someone losing a parent/child I cringe. We have to accept this as a reality of life until a proper/generic cure is found that eliminates the most common offenders.
I am skeptical that we will have AGI anytime soon and I think the social aspects will help balance the technical developments even it becomes a reality (Three laws, A Butlerian uprising, you name it).
Chess bots can beat grandmasters, but I have a friend who takes his son to tournaments. Humans are still playing chess, kids in the same tournament with grand masters. We have to have faith in the humanity, or all else will not matter.
And I will definitely keep playing Factorio even if AGI comes to pass ;-)
But now I see that my interpretation is almost opposite of what the words mean... Huh...
Like how during Covid the beaches were closed while the billionaires partied on yachts.
Same with learning, humans historically where generalists without that deep of knowledge when compared to one another. Now we study a quarter of our short lives just to get to the point where we can specialize for the rest of our lives. This situation doesn't seem exactly tenable as complexity increases in the future.
I don't know what the future looks like, but I can tell you that entropy and complexity will only increase.
Obviously. Exercise, play and learning is in large part what makes us smart. If we want AI to be smart we should definitely automate those.
But that's not what you had in mind. To address that I can only say that you are allowed to still do the automated things manually.
Add this to the long list of names like Terence Tao, and others who seem to be intellectually incontinent lately in the sense that one cannot navigate this space anymore without encountering their thoughts
I tell jokes and the group of friends get them, for family they don't get them anymore.
I do not like 'basic jokes' and despite that, german television is full of it.
Most things in our world are more of a challange of finding the answers and not 'creating' the answers.
Math: the answer is already there, you only need to find it an fast space of posibilities. This is perfect for LLMs.
Creativity: a LLM can iterate over things a lot faster than a person. So we can iterate over this space too. We can also get feedback from people, tiktok, instagram.
I'd link to the HN guidelines here but I'm on my phone!
"Evaluation" means environments or datasets, the model is supposed to discover its representations from scaled up experience. That was the bitter lesson - more data and compute beat heuristics.
Stochastic: describes a process, model, or system that involves inherent randomness, meaning its future states or outcomes cannot be predicted with absolute certainty but can be described using probability distributions.
> "There's a tradition of scientists approaching senility to come up with grand, improbable theories. Wolfram is unusual in that he's doing this in his forties."
I always felt it was an unfair dismissal of someone's life's work. Maybe it was true but it didn't enrich the discussion or our understanding. I suppose it means even a respected thinker can be guilty of shallow dismissals and saying hurtful things in public about others.
It's similar to a "thought-terminating cliché", in that it just reinforces an existing opinion without adding anything, making us think deeper, or furthering the conversation.