What John Carmack is exploring is pretty revealing.
Train models to play 2D video games to a superhuman level, then ask them to play a level they have not seen before or another 2D video game they have not seen before. The transfer function is negative. So, in my definition, no intelligence has been developed, only expertise in a narrow set of tasks.
It’s apparently much easier to scare the masses with visions of ASI, than to build a general intelligence that can pick up a new 2D video game faster than a human being.
Seeing comments here saying “this problem is already solved”, “he is just bad at this” etc. feels bad. He has given a long time to this problem by now. He is trying to solve this to advance the field. And needless to say, he is a legend in computer engineering or w/e you call it.
It should be required to point to the “solution” and maybe how it works to say “he just sucks” or “this was solved before”.
IMO the problem with current models is that they don’t learn categorically like: lions are animals, animals are alive. goats are animals, goats are alive too. So if lions have some property like breathing and goats also have it, it is likely that other similar things have the same property.
Or when playing a game, a human can come up with a strategy like: I’ll level this ability and lean on it for starting, then I’ll level this other ability that takes more time to ramp up while using the first one, then change to this play style after I have the new ability ready. This might be formulated completely based on theoretical ideas about the game, and modified as the player gets more experience.
With current AI models as far as I can understand, it will see the whole game as an optimization problem and try to find something at random that makes it win more. This is not as scalable as combining theory and experience in the way that humans do. For example a human is innately capable of understanding there is a concept of early game, and the gains made in early game can compound and generate a large lead. This is pattern matching as well but it is on a higher level .
Theory makes learning more scalable compared to just trying everything and seeing what works
He is not using appropriate models for this conclusion and neither is he using state of the art models in this research and moreover he doesn't have an expensive foundational model to build upon for 2d games. It's just a fun project.
A serious attempt at video/vision would involve some probabilistic latent space that can be noised in ways that make sense for games in general. I think veo3 proves that ai can generalize 2d and even 3d games, generating a video under prompt constraints is basically playing a game. I think you could prompt veo3 to play any game for a few seconds and it will generally make sense even though it is not fine tuned.
Veo3's world model is still pretty limited. That becomes obvious very fast once you prompt out of distribution video content (i.e. stuff that you are unlikely to find on youtube). It's extremely good at creating photorealistic surfaces and lighting. It even has some reasonably solid understanding of fluid dynamics for simulating water. But for complex human behaviour (in particular certain motions) it simply lacks the training data. Although that's not really a fault of the model and I'm pretty sure there will be a way to overcome this as well. Maybe some kind of physics based simulation as supplement training data.
To confirm: hybrid approaches can demonstrate competence at newly-created video games within a short period of exposure, so long as similar game mechanics from other games were incorporated into their training set?
> I think veo3 proves that ai can generalize 2d and even 3d games, generating a video under prompt constraints is basically playing a game.
In the same way that keeping a dream journal is basically doing investigative journalism, or talking to yourself is equivalent to making new friends, maybe.
The difference is that while they may both produce similar, "plausible" output, one does so as a result of processes that exist in relation to an external reality.
> generating a video under prompt constraints is basically playing a game
Besides static puzzles (like a maze or jigsaw) I don't believe this analogy holds? A model working with prompt constraints that aren't evolving or being added over the course of "navigating" the generation of the model's output means it needs to process 0 new information that it didn't come up with itself — playing a game is different from other generation because it's primarily about reacting to input you didn't know the precise timing/spatial details of, but can learn that they come within a known set of higher order rules. Obviously the more finite/deterministic/predictably probabilistic the video game's solution space, the more it can be inferred from the initial state, aka reduce to the same type of problem as generating a video from a prompt), which is why models are still able to play video games. But as GP pointed out, transfer function negative in such cases — the overarching rules are not predictable enough across disparate genres.
> I think you could prompt veo3 to play any game for a few seconds
I'm curious what your threshold for what constitutes "play any game" is in this claim? If I wrote a script that maps button combinations to average pixel color of a portion of the screen buffer, by what metric(s) would veo3 be "playing" the game more or better than that script "for a few seconds"?
It's not ideal, but you can prompt it with an image of a game frame, explain the objects and physics in text and let it generate a few frames of gameplay as a substitute for controller input as well as what it expects as an outcome. I am not talking about real interactive gameplay.
I am just saying we have proof that it can understand complex worlds and sets of rules, and then abide by them. It doesn't know how to use a controller and it doesn't know how to explore the game physics on its own, but those steps are much easier to implement based on how coding agents are able to iterate and explore solutions.
fair, and I edited my choice of words, but if you're reading that much aggression from my initial comment (which contains topical discussion) to say what you did, you must find the internet a far more savage place than it really is :/
> I think veo3 proves that ai can generalize 2d and even 3d games
It doesn't. And you said it yourself:
> generating a video under prompt constraints is basically playing a game.
No. It's neither generating a game (that people can play) nor is it playing a game (it's generating a video).
Since it's not a model of the world in any sense of the word, there are issues with even the most basic object permanenece. E.g. here's veo3 generating a GTA-style video. Oh look, the car spins 360 and ends up on a completely different street than the one it was driving down previously: https://www.youtube.com/watch?v=ja2PVllZcsI
It is still doing a great job for a few frames, you could keep it more anchored to the state of the game if you prompt it. Much like you can prompt coding agents to keep a log of all decisions previously made. Permanenece is excellent, it slips often but it mostly because it is not grounded to specific game state by the prompt or by the decision log.
I wonder if this is a case of overfitting from allowing the model to grow too large, and if you might cajole it into learning more generic heuristics by putting some constraints on it.
It sounds like the "best" AI without constraint would just be something like a replay of a record speedrun rather than a smaller set of heuristics of getting through a game, though the latter is clearly much more important with unseen content.
The subject you are referring to is most likely Meta-Reinforcement Learning [1]. It is great that John Carmack is looking into this, but it is not a new field of research.
I don't get why people are so invested in framing it this way. I'm sure there are ways to do the stated objective. John Carmack isn't even an AI guy why is he suddenly the standard.
Who is an "AI guy"? The field as we know it is fairly new. Sure, neural nets are old hat, but a lot has happened in the last few years.
John Carmack founded Keen technology in 2022 and has been working seriously on AI since 2019. From his experience in the video game industry, he knows a thing or two about linear algebra and GPUs, that is the underlying maths and the underlying hardware.
So, for all intent and purposes, he is an "AI guy" now.
That does not mean there isn't an AI system that can do X. Especially considering that a lot is happening in AI, as you say.
Anyway, Carmack knows a lot about optimizing computations on modern hardware. In practice, that happens to be also necessary for AI. However, it is not __sufficient__ for AI.
Keen includes researchers like Richard Sutton, Joseph Modayil etc. Also John has being doing it full time for almost 5 years now so given his background and aptitude for learning I would imaging by this time he is more of an AI guy then a fairly large percentage of AI PhDs.
Not sure why justanotherjoe is a credible resource on who is and isn’t expert in some new dialectic and euphemism for machine state management. You’re that nobody to me :shrug:
Yann LeCun is an AI guy and has simplified it as “not much more than physical statistics.”
WWhole lot of AI is decades old info theory books applied to modern computer.
Either a mem value is or isn’t what’s expected. Either an entire matrix of values is or isn’t what’s expected. Store the results of some such rules. There’s your model.
The words are made up and arbitrary because human existence is arbitrary. You’re being sold on a bridge to nowhere.
One phenomena that bared this to me, in a substantive way, was noticing an increasing # of reverent comments re: Geohot in odd places here, that are just as quickly replied to by people with a sense of how he works, as opposed to the keywords he associates himself with. But that only happens here AFAIK.
Yapping, or, inducing people to yap about me, unfortunately, is much more salient to my expected mindshare than the work I do.
It's getting claustrophobic intellectually, as a result.
Example from the last week is the phrase "context engineering" - Shopify CEO says he likes it better than prompt engineering, Karpathy QTs to affirm, SimonW writes it up as fait accompli. Now I have to rework my site to not use "prompt engineering" and have a Take™ on "context engineering". Because of a couple tweets + a blog reverberating over 2-3 days.
Nothing against Carmack, or anyone else named, at all. i.e. in the context engineering case, they're just sharing their thoughts in realtime. (i.e. I don't wanna get rolled up into a downvote brigade because it seems like I'm affirming the loose assertion Carmack is "not an AI guy", or, that it seems I'm criticizing anyone's conduct at all)
EDIT: The context engineering example was not in reference to another post at the time of writing, now one is the top of front page.
> Now I have to rework my site to not use "prompt engineering" and have a Take™ on "context engineering". Because of a couple tweets + a blog reverberating over 2-3 days.
The difference here is that your example shows a trivial statement and a change period of 3 days, whereas what Carmack is doing is taking years.
Right. Nothing against Carmack. Grew up on the guy. I haven't looked into, at all, into any of the disputed stuff, and should actively proclaim I'm a yuge Carmack fanboy.
These questions of whether the model is “really intelligent” or whatever might be of interest to academics theorizing about AGI, but to the vast swaths of people getting useful stuff out of LLMs, it doesn’t really matter. We don’t care if the current path leads to AGI. If the line stopped at Claude 4 I’d still keep using it.
And like I get it, it’s fun to complain about the obnoxious and irrational AGI people. But the discussion about how people are using these things in their everyday lives is way more interesting.
According to Carmack's recent talk [0], SOTA models that have been trained on game A don't perform better or train faster on game B. Even worse, training on game B negatively affects performance in game A when returning to it.
In the Physics of Language Models[1] they argue that you must augment your training data by changing sentences and such, in order for the model to be able to learn the knowledge. As I understand their argument, language models don't have a built-in way to detect what is important information and what is not, unlike us. Thus the training data must aid it by presenting important information in many different ways.
Doesn't seem unreasonable that the same holds in a gaming setting, that one should train on many variations of each level. Change the lengths of halls connecting rooms, change the appearance of each room, change power-up locations etc, and maybe even remove passages connecting rooms.
I think this is clearly a case of over fitting and failure to generalize, which are really well understood concepts. We don't have to philosophize about what pattern matching really means.
They learn the value of specific actions in specific contexts based on the rewards they received during their play time. Specific actions and specific contexts are not transferable for various reasons. John quoted that varying frame rates and variable latency between action and effect really confuse the models.
It’s really easy: go to Claude and ask it a novel question. It will generally reason its way to a perfectly good answer even if there is no direct example of it in the training data.
When LLM's come up with answers to questions that aren't directly exampled in the training data, that's not proof at all that it reasoned its way there — it can very much still be pattern matching without insight from the actual code execution of the answer generation.
If we were taking a walk and you asked me for an explanation for a mathematical concept I have not actually studied, I am fully capable of hazarding a casual guess based on the other topics I have studied within seconds. This is the default approach of an LLM, except with much greater breadth and recall of studied topics than I, as a human, have.
This would be very different than if we sat down at a library and I applied the various concepts and theorems I already knew to make inferences, built upon them, and then derived an understanding based on reasoning of the steps I took (often after backtracking from several reasoning dead ends) before providing the explanation.
If you ask an LLM to explain their reasoning, it's unclear whether it just guessed the explanation and reasoning too, or if that was actually the set of steps it took to get to the first answer they gave you. This is why LLMs are able to correct themselves after claiming strawberry has 2 rs, but when providing (guessing again) their explanations they make more "relevant" guesses.
You have probably seen examples of LLMs doing the "mirror test", i.e. identifying themselves in screenshots and referring to the screenshot from the first person. That is a genuinely novel question as an "LLM mirror test" wasn't a concept that existed before about a year ago.
The question isn't about universal novelty, but whether the prompt/context is novel enough such that the LLM answering competently demonstrates understanding. The claim of parroting is that the dataset contains a near exact duplicate of any prompt and so the LLM demonstrating what appears to be competence is really just memorization. But if an LLM can generalize from an elephant mirror test to an LLM mirror test in an entirely new context (showing pictures and being asked to describe it), that demonstrates sufficient generalization to "understand" the concept of a mirror test.
And as any programmer will tell you: they immediately devolve into "hallucinating" answers, not trying to actually reason about the world. Because that's what they do: they create statistically plausible answers even if those answers are complete nonsense.
Just go and ask ChatGPT or Claude something that can't possibly be in its training set. Make something up. If it is only memorising answers then it will be impossible for it to get the correct result.
A simple nonsense programming task would suffice. For example "write a Python function to erase every character from a string unless either of its adjacent characters are also adjacent to it in the alphabet. The string only contains lowercase a-z"
That task isn't anywhere in its training set so they can't memorise the answer. But I bet ChatGPT and Claude can still do it.
Honestly this is sooooo obvious to anyone that has used these tools, it's really insane that people are still parroting (heh) the "it just memorises" line.
LLMs don't "memorize" concepts like humans do. They generate output based on token patterns in their training data. So instead of having to be trained on every possible problem, they can still generate output that solves it by referencing the most probable combination of tokens for the specified input tokens. To humans this seems like they're truly solving novel problems, but it's merely a trick of statistics. These tools can reference and generate patterns that no human ever could. This is what makes them useful and powerful, but I would argue not intelligent.
Well yeah... If you only ever played one game in your life you would probably be pretty shit at other games too. This does not seem very revealing to me.
I don't think thats true. If you'd only ever played Doom, I think you could play, say, counterstrike or half-life and be pretty good at it, and i think Carmack is right that its pretty interesting that this doesn't seem to be the case for ai models
yeahhhh why isnt there a training structure where you play 5000 games, and the reward function is based on doing well in all of them?
I guess its a totaly different level of control: instead of immediately choosing a certain button to press, you need to set longer term goals. "press whatever sequence over this time i need to do to end up closer to this result"
There is some kind of nested multidimensional thing to train on here instead of immediate limited choices
The line is: building an internal world model requires interfacing with the world, not a model of it, and subsequent failing (including death and survivorship over generations) and adaptation. Plus pattern matching.
Current AI only does one of those (pattern matching, not evolution), and the prospects of simulating evolution is kind of bleak, given I don’t think we can simulate a full living cell yet from scratch? Building a world model requires life (or something that has undergone a similar evolutionary survivorship path), not something that mimics life.
I've wondered about the claim that the models played those Atari/2D video games at superhuman levels, because I clearly recall some humans achieving superhuman levels before models were capable of it. Must have been superhuman compared to average human player, not someone who spent an inordinate amount of time mastering the game.
I'm not sure why you think so. AI outperforms humans in many games already. Basically all the games we care to put money to train a model.
AI has beat the best human players in Chess, Go, Mahjong, Texas hold'em, Dota, Starcraft, etc. It would be really, really surprising that some Atari game is the holy grail of human performance that AI cannot beat.
I recall this not being true at all for Dota and Starcraft. I recall AlphaStar performed much better than the top non-pro players, but it couldn't consistently beat the pro players with the budget that Google was willing to spend, and I believe the same was true of Dota II (and there they were even playing a limited form of the game, with fewer heroes and without the hero choice part, I believe).
As I recall, the Starcraft ones heavily involved being able to exploit the computer's advantage in "twitch" speed over any human, it's just a slightly more complicated way of how any aim-bot enabled AI will always beat a human in an FPS, the game is designed to reward a certain amount of physical speed and accuracy.
In other words, the Starcraft AIs that win do so by microing every single unit in the entire game at the same time, which is pretty clever, but if you reduce them to interfacing with the game in the same way a human does, they start losing.
One of my pet peeves when we talk about the various chess engines is yes, given a board state they can output the next set of moves to beat any human, but can they teach someone else to play chess? I'm not trying to activate some kinda "gotcha" here, just getting at what does it actually mean to "know how to play chess". We'd expect any human that claimed to know how to play to be able to teach any other human pretty trivially.
When I finished my degree, the idea that a software system could develop that level of expertise was relegated to science fiction. It is an unbelievable human accomplishment to get to that point and honestly, a bit of awe makes life more pleasant.
Less quality of life focused, I don’t believe that the models he uses for this research are capable of more. Is it really that revealing?
I wonder how much performance decreases if they just use slightly modified versions of the same game. Like a different color scheme, or a couple different sprites.
No, they (and many others before them) are genuinely trying to improve on the original research.
The original paper "Playing Atari with Deep Reinforcement Learning" (2013) from Deepmind describes how agents can play Atari games, but these agents would have to be specifically trained on every individual game using millions of frames. To accomplish this, simulators were run in parallel, and much faster than in real-time.
Also, additional trickery was added to extract a reward signal from the games, and there is some minor cheating on supplying inputs.
What Carmack (and others before him) is interested in, is trying to learn in a real-life setting, similar to how humans learn.
I'd say with confidence: we're living in the early days. AI has made jaw-dropping progress in two major domains: language and vision. With large language models (LLMs) like GPT-4 and Claude, and vision models like CLIP and DALL·E, we've seen machines that can generate poetry, write code, describe photos, and even hold eerily humanlike conversations.
But as impressive as this is, it’s easy to lose sight of the bigger picture: we’ve only scratched the surface of what artificial intelligence could be — because we’ve only scaled two modalities: text and images.
That’s like saying we’ve modeled human intelligence by mastering reading and eyesight, while ignoring touch, taste, smell, motion, memory, emotion, and everything else that makes our cognition rich, embodied, and contextual.
Human intelligence is multimodal. We make sense of the world through:
Touch (the texture of a surface, the feedback of pressure, the warmth of skin0; Smell and taste (deeply tied to memory, danger, pleasure, and even creativity); Proprioception (the sense of where your body is in space — how you move and balance); Emotional and internal states (hunger, pain, comfort, fear, motivation).
None of these are captured by current LLMs or vision transformers. Not even close. And yet, our cognitive lives depend on them.
Language and vision are just the beginning — the parts we were able to digitize first - not necessarily the most central to intelligence.
The real frontier of AI lies in the messy, rich, sensory world where people live. We’ll need new hardware (sensors), new data representations (beyond tokens), and new ways to train models that grow understanding from experience, not just patterns.
> Language and vision are just the beginning — the parts we were able to digitize first - not necessarily the most central to intelligence.
I respectfully disagree. Touch gives pretty cool skills, but language, video and audio are all that are needed for all online interactions. We use touch for typing and pointing, but that is only because we don't have a more efficient and effective interface.
Now I'm not saying that all other senses are uninteresting. Integrating touch, extensive proprioception, and olfaction is going to unlock a lot of 'real world' behavior, but your comment was specifically about intelligence.
Compare humans to apes and other animals and the thing that sets us apart is definitely not in the 'remaining' senses, but firmly in the realm of audio, video and language.
> Language and vision are just the beginning — the parts we were able to digitize first - not necessarily the most central to intelligence.
I probably made a mistake when i asserted that -- should have thought it over. Vision is evolutionarily older and more “primitive”, while language is uniquely human [or maybe, more broadly, primate, cetacean, cephalopod, avian...] symbolic, and abstract — arguably a different order of cognition altogether. But i maintain that each and every sense is important as far as human cognition -- and its replication -- is concerned.
Vision is interesting in that it leverages the maximum speed with which it is easily possible to gather information about our surroundings in this universe. I believe that is what makes it special and very valuable. I also believe this aspect makes it a strong attractor for convergent evolution.
Language allows encoding and compression of information about the world, which is of course incredibly powerful and increases communication bandwidth enormously (as well as tons of other stuff).
I'd say that for high level cognitive processes, hearing and speaking were an important stepping stone because for some reason evolving organs that can generate relatively high bandwidth signals in audio seems to be easier than evolving something that does that for visuals (very few Teletubby screens on tummies in the natural world).
Interesting games to think about in this sense: Pictionary/drawing games and charades.
Organic adaption and persistence of memory I would say are the two major advancements that need to happen.
Human neural networks are dynamic, they change and rearrange, grow and sever. An LLM is fixed and relies on context, if you give it the right answer it won't "learn" that is the correct answer unless it is fed back into the system and trained over months. What if it's only the right answer for a limited period of time?
To build an intelligent machine, it must be able train itself in real time and remember.
Based on the architectures we have they may also be the ending. There’s been a lot of news in the past couple years about LLMs but has there been any breakthroughs making headlines anywhere else in AI?
> There’s been a lot of news in the past couple years about LLMs but has there been any breakthroughs making headlines anywhere else in AI?
Yeah, lots of stuff tied to robotics, for instance; this overlaps with vision, but the advances go beyond vision.
Audio has seen quite a bit. And I imagine there is stuff happening in niche areas that just aren't as publicly interesting as language, vision/imagery, audio, and robotics.
> The real frontier of AI lies in the messy, rich, sensory world where people live. We’ll need new hardware (sensors), new data representations (beyond tokens), and new ways to train models that grow understanding from experience, not just patterns.
Like Dr. Who said: DALEKs aren't brains in a machine, they are the machine!
Same is true for humans. We really are the whole body, we're not just driving it around.
There are many people who mentally developed while paralyzed that literally drive around their bodies via motorized wheelchair. I don't think there's any evidence that a brain couldn't exist or develop in a jar, given only the inputs modern AI now has (text, video, audio).
> any evidence that a brain couldn't exist or develop in a jar
The brain could. Of course it could. It's just a signals processing machine.
But would it be missing anything we consider core to the way humans think? Would it struggle with parts of cognition?
For example: experiments were done with cats growing up in environments with vertical lines only. They were then put in a normal room and had a hard time understanding flat surfaces.
This isn't remotely a hypothetical, so I imagine there are some examples out there, especially from back when polio was a problem. Although, for practical reasons, they might have had limited exposure to novelty, which could have negative consequences.
They took 1970s dead tech and deployed it on machines 1 million times more powerful. I'm not sure I'd qualify this as progress. I'd also need an explanation as to what systemic improvements in models and computations that give an exponential growth in performance are planned.
> They took 1970s dead tech and deployed it on machines 1 million times more powerful. I’m not sure I’d qualify this as progress
If this isn’t meant to be sarcasm or irony, you’ve got some really exciting research and learning ahead of you! At the moment it reads very “computers are just addition and multiplication and we’ve had that for thousands of years!”
Sometimes we get confused by the difference between technological and scientific progress. When science makes progress it unlocks new S-curves that progress at an incredible pace until you get into the diminishing returns region. People complain of slowing progress but it was always slow, you just didn’t notice that nothing new was happening during the exponential take off of the S-curve, just furious optimization.
And at the same time I have noticed that people don’t understand the difference between an S-curve and an exponential function. They can look almost identical at certain intervals.
As far back as 2017 I copped a lot of flak for suggesting that the coming automation revolution will be great at copying office workers and artists but wont be in order of replacing the whole human race. A lot of the time moores law got thrown back in my face. But thats how this works, we unlock something new, we exploit it as far as possible, the shine wears off and we deal with the aftermath.
If you work with model architecture and read papers, how could not know there are a flood of new ideas? Only few yield interesting results though.
I kind of wonder if libraries like pytorch have hurt experimental development. So many basic concepts no one thinks about anymore because they just use the out of the box solutions. And maybe those solutions are great and those parts are "solved", but I am not sure. How many models are using someone else's tokenizer, or someone else's strapped on vision model just to check a box in the model card?
That's been the very normal way of the human world.
When the foundation layer at a given moment doesn't yield an ROI on intellectual exploration - say because you can overcompensate with VC funded raw compute and make more progess elsewhere -, few(er) will go there.
But inevitably, as other domains reach diminishing returns, bright minds will take a look around where significant gains for their effort can be found.
And so will the next generation of PyTorch or foundational technologies evolve.
Yeah and even then, it's been like ~ 2-3 years since the last rather major Architectural improvement, major enough for a lot of people to actually hear about it and use it daily. I think some people lose perspective on how short of a time frame 3 years is.
But yes, there's a ton of interesting and useful stuff (beyond datasets and data related improvements) going on right now, and I'm not even talking about LLMs. I don't do anything related to LLM and even then I still see tons of new stuff popping up regularly.
To be fair, if you imagine a system that successfully reproduced human intelligence, then 'changing datasets' would probably be a fair summary of what it would take to have different models. After all, our own memories, training, education, background, etc are a very large component of our own problem solving abilities.
I will respectfully disagree. All "new" ideas come from old ideas. AI is a tool to access old ideas with speed and with new perspectives that hasn't been available up until now.
Innovation is in the cracks: recognition of holes, intersections, tangents, etc. on old ideas. It has bent said that innovation is done on the shoulders of giants.
So AI can be an express elevator up to an army of giant's shoulders? It all depends on how you use the tools.
Access old ideas? Yes. With new perspectives? Not necessarily. An LLM may be able to assist in interpreting data with new perspectives but in practice they're still fairly bad at greenfield work.
As with most things, the truth lies somewhere in the middle. LLMs can be helpful as a way of accelerating certain kinds and certain aspects of research but not others.
I wonder if we can mine patent databases for old ideas that never worked out in the past, but now are more useful. Perhaps due to modern machining or newer materials or just new applications of the idea.
Imagine a human had read every book/publication in every field of knowledge that mankind has ever produced AND couldn’t come up with anything entirely new. Hard to imagine.
My hypothesis of the mismatch is centered around "read" - I think that when you wrote it, and when others similarly think about that scenario, the surprise is because our version of "read" is the implied "read and internalized" or at bare minimum "read for comprehension" but as very best I can tell the LLM's version is "encoded tokens into vector space" and not "encoded into semantic graph"
I welcome the hair-splittery that is sure to follow about what it means to "understand" anything
The article is discussing working in AI innovation vs focusing on getting more and better data. And while there have been key breakthroughs in new ideas, one of the best ways to increase the performance of these systems is getting more and better data. And how many people think data is the primary avenue to improvement.
It reminds me of an AI talk a few decades ago, about how the cycle goes: more data -> more layers -> repeat...
Anyways, I'm not sure how your comment relates to these two avenues of improvement.
> I will respectfully disagree. All "new" ideas come from old ideas.
The insight into the structure of the benzene ring famously came in a dream, hadn't been seen before, but was imagined as a snake bitings its own tail.
The empirical formula for benzene had been long known, but its highly unsaturated structure was a challenge to determine. Archibald Scott Couper in 1858 and Joseph Loschmidt in 1861 suggested possible structures that contained multiple double bonds or multiple rings, but the study of aromatic compounds was in its earliest years, and too little evidence was then available to help chemists decide on any particular structure.
More evidence was available by 1865, especially regarding the relationships of aromatic isomers.
[ Kekule claimed to have had the dream in 1865 ]
--- end quote ---
The dream claim came from Kekule himself 25 years after his proposal that he had to modify 10 years after he proposed it.
What about actively obtained data - models seeking data, rather than being fed. Human babies put things in their mouths, they try to stand and fall over. They “do stuff” to learn what works. Right now we’re just telling models what works.
What about simulation: models can make 3D objects so why not give them a physics simulator? We have amazing high fidelity (and low cost!) game engines that would be a great building block.
What about rumination: behind every Cursor rule for example, is a whole story of why a user added it. Why not take the rule, ask a reasoning model to hypothesize about why that rule was created, and add that rumination (along with the rule) to the training data. Providing opportunities to reflect on the choices made by their users might deepen any insights, squeezing more juice out of the data.
Simulation and embodied AI (putting the AI in a robotic arm or a car so it can try stuff and gather information about the results) are very actively being explored.
Yeah, at a very high level it's similar to an actor-critic reinforcement learning algorithm. The rule text is a value function and one could build a critic model that takes as input the rule text and the main model's (the actor's) output to produce a reward.
This is easier said than done though because this value function is so noisy it's often hard to learn from it. And also whether or not a response (the model output) matches the value function (the Cursor rules) is not even that easy to grade. It's been easier to train the chain-of-thought style reasoning since one can directly score it via the length of thinking.
You are just a lot of interactions of waves. All meaning is assigned. I prefer to think of this like the Goedel generator that found new formal expressions for the Principia - because we have a way of indexing concept-space, there's no telling what we might find in the gaps.
And architecture stuff like actually useful long context. Whatever they did with gemini 2.5 is miles ahead in long context useful results compared to the previous models. I'd be very surprised if gemini 2.5 is "just" gemini 1 w/ better data.
i dont know what all the hype is with gemini 2.5, at least the currently running instance. from my experience at least in conversation mode, it cannot remember my instructions to avoid apologies and similar platitudes from either the “persona”, personal instructions, or from ine message to the next.
You raise a really interesting point. I'm sure it's just missed my notice, but I'm not familiar with any projects from antediluvian AI that have been resurrected to run on modern hardware and see where they'd really asymptote if they'd had the compute they deserved.
This paper “were RNNs all we needed?” explores this hypothesis a bit, finding that some pre-transformer sequence models can match transformers when trained at appropriate scale. Though they did have to make some modifications to unlock more parallelism
can you name a couple so i can see how much work is involved? markov chains compile fast and respond fast, sure, and neural nets train pretty quick too, so i'm wondering where the cutoff is; expert systems?
Big difference between a perfect information, completely specified zero sum game and the real world.
As a simple analogy, read out the following sentence multiple times, stressing a different word each time.
"I never said she stole my money"
Note how the meaning changes and is often unique?
That is a lens I to the frame problem and it's inverse, the specification problem.
The above problem quickly becomes tower-complete, and recent studies suggest that RL is reinforcing or increasing the weight of existing patterns.
As the open domain frame problem and similar challenges are equivalent to HALT, finding new ways to extract useful information will be important for generalization IMHO.
Synthetic data is useful, but not a complete solution, especially for tower problems.
and as far as synthetic vs real data, there's a lot of gaps in LLM knowledge; and vision models suffer from "limited tags", which used to have workarounds with textual embeddings and the like, but those went by the wayside as LoRA, controlnet, etc. appeared.
There's people who are fairly well known that LLMs have no idea about. There's things in books i own that the AI confidently tells me are either wrong or don't exist.
That one page about compressing 1 gig wikipedia as small as possible implicitly and explicitly states that AI is "basically compression" - and if the data isn't there, it's not in the compressed set (weights) either.
And i'll reply to another comment here, about "24/7 rolling/ for looped" AI - i thought of doing this when i first found out about LLMs, but context windows are the enemy, here. I have a couple of ideas about how to have a continuous AI, but i don't have the capital to test it out.
There are new ideas, people are finding new ways to build vision models, which then are applied to language models and vice versa (like diffusion).
The original idea of connectionism is that neural networks can represent any function, which is the fundamental mathematical fact. So we should be optimistic, neural nets will be able to do anything. Which neural nets? So far people have stumbled on a few productive architectures, but it appears to be more alchemy than science. There is no reason why we should think there won't be both new ideas and new data. Biology did it, humans will do it too.
> we’re engaged in a decentralized globalized exercise of Science, where findings are shared openly
Maybe the findings are shared, if they make the Company look good. But the methods are not anymore
AI training is currently a process of making the AI remember the dataset. It doesn't involve the AI thinking about the dataset and drawing (and remembering) conclusions.
It can probably remember more facts about a topic than a PhD in that topic, but the PhD will be better at thinking about that topic.
Its a bit more complex than that. Its more about baking out the dataset into heuristics that a machine can use to match a satisfying result to an input. Sometimes these heuristics are surprising to a human and can solve a problem in a novel way.
"Thinking" is too broad a term to apply usefully but I would say its pretty clear we are not close to AGI.
If datasets are what we are talking about, I'd like to bring attention to the biological datasets out there that have yet to be fully harnessed.
The ability to collect gene expression data at a tissue specific level has only been invented and automated in the last 4-5 years (see 10X Genomics Xenium, MERFISH). We've only recently figured out how to collect this data at the scale of millions of cells. A breakthrough on this front may be the next big area of advancement.
I don't know if it matters. Even if the best we can do is get really good at interpolating between solutions to cognitive tasks on the data manifold, the only economically useful human labor left asymptotes toward frontier work; work that only a single-digit percentage of people can actually perform.
Hardware isn't even close to being out of steam. There are some breathtakingly obvious premature optimizations that we can undo to get at least 99% power reduction for the same amount of compute.
For example, FPGAs use a lot of area and power routing signals across the chip. Those long lines have a large capacitance, and thus cause a large amount of dynamic power loss. So does moving parameters around to/from RAM instead of just loading up a vast array of LUTs with the values once.
An interesting step forward, although an old idea we seem close to is recursive self improvement. Get the AI to make a modified version of itself to try to think better.
Here's an idea: make the AIs consistent at doing things computers are good at. Here's an anecdote from a friend who's living in Japan:
> i used chatgpt for the first time today and have some lite rage if you wanna hear it. tldr it wasnt correct. i thought of one simple task that it should be good at and it couldnt do that.
> (The kangxi radicals are neatly in order in unicode so you can just ++ thru em. The cjks are not. I couldnt see any clear mapping so i asked gpt to do it. Big mess i had to untangle manually anyway it woulda been faster to look them up by hand (theres 214))
> The big kicker was like, it gave me 213. And i was like, "why is one missing?" Then i put it back in and said count how many numbers are here and it said 214, and there just werent. Like come on you SHOULD be able to count.
If you can make the language models actually interface with what we've been able to do with computers for decades, i imagine many paths open up.
This needs to be generalized however. For example, if you present an AI with a drawing of some directed graph (a state diagram, for example), it should be able to answer questions based on the precise set of all possible paths in that graph, without someone having to write tooling for diagram or graph processing and traversal. Or, given a photo of a dropped box of matches, an AI should be able to precisely count the matches, as far as they are individually visible (which a human could do by keeping a tally while coloring the matches). There are probably better examples, these are off the cuff.
There’s an infinite repertoire of such tasks that combine AI capabilities with traditional computer algorithms, and I don’t think we have a generic way of having AI autonomously outsource whatever parts require precision in a reliable way.
What you're describing sounds like agentic tool usage. Have you kept up with the latest developments on that? it's already solved depending on how strict you define your criteria above
My understanding is that you need to provide and configure task-specific tools. You can’t combine the AI with just a general-purpose computer and have the AI figure out on its own how to make use of it to achieve with reliability and precision whatever task it is given. In other words, the current tool usage isn’t general-purpose in the way the LLM itself is, and also the LLM doesn’t reason about its own capabilities in order to decide how to incorporate computer use to compensate for its own weaknesses. Instead you have to tell the LLM what it should apply the tooling for.
"There weren't really any advancements from around 2018. The majority of the 'advancements' were in the amount of parameters, training data, and its applications. What was the GPT-3 to ChatGPT transition? It involved fine-tuning, using specifically crafted training data. What changed from GPT-3 to GPT-4? It was the increase in the number of parameters, improved training data, and the addition of another modality. From GPT-4 to GPT-40? There was more optimization and the introduction of a new modality. The only thing left that could further improve models is to add one more modality, which could be video or other sensory inputs, along with some optimization and more parameters. We are approaching diminishing returns." [1]
10 months ago around o1 release:
"It's because there is nothing novel here from an architectural point of view. Again, the secret sauce is only in the training data.
O1 seems like a variant of RLRF https://arxiv.org/abs/2403.14238
Soon you will see similar models from competitors." [2]
This seems simplistic, tech and infrastructure play a huge part here. A short and incomplete list of things that contributed:
- Moore's law petering out, steering hardware advancements towards parallelism
- Fast-enough internet creating shift to processing and storage in large server farms, enabling both high-cost training and remote storage of large models
- Social media + search both enlisting consumers as data producers, and necessitating the creation of armies of Mturkers for content moderation + evaluation, later becoming available for tagging and rlhf
- A long-term shift to a text-oriented society, beginning with print capitalism and continuing through the rise of "knowledge work" through to the migration of daily tasks (work, bill paying, shopping) online, that allows a program that only produces text to appear capable of doing many of the things a person does
We may have previously had the technical ideas in the 1990s but we certainly didn't have the ripened infrastructure to put them into practice. If we had the dataset to create an LLM in the 90s, it still would have been astronomically cost-prohibitive to train, both in CPU and human labor, and it wouldn't have as much of an effect on society because you wouldn't be able to hook it up to commerce or day-to-day activities (far fewer texts, emails, ecommerce).
if datasets are the new codebases ,then the real IP can be dataset version control. how you fork ,diff ,merge and audit datasets like code. every team says 'we trained on 10B tokens' but what if we can answer 'which 5M tokens made reasoning better', 'which 100k made it worse'. then we can start being targeted leverage
disagree, there are a few organisations exploring novel paths. It's just that throwing new data at an "old" algorithm is much easier and has been a winning strategy. And, also, there's no incentive for a private org to advertise a new idea that seems to be working (mine's a notable exception :D).
Until these "AI" systems become always-on, always-thinking, always-processing, progress is stuck. The current push button AI - meaning it only processes when we prompt it - is not how the kind of AI that everyone is dreaming of needs to function.
Things haven't changed much in terms of truly new ideas since electricity was invented. Everything else is just applications on top of that. Make the electrons flow in a different way and you get a different outcome.
This brings us to the problem AI companies are facing, the lack of data, they have already hoovered as much as they can from the internet and desperately need more data.
Which make sit blatantly obvious why we're beginning to see products being marketed under the guise of assistants/tools to aid you whose actual purpose is to gather real world picture and audio data, think meta glasses and what Ives and Altman are cooking up with their partnership.
Paradigm shifts are often just a conglomeration of previous ideas with one little tweak that suddenly propels a technology ahead 10x which opens up a whole new era.
The iPhone is a perfect example. There were smartphones with cameras and web browsers before. But when the iPhone launched, it added a capacitive touch screen that was so responsive there was no need for a keyboard. The importance of that one technical innovation can't be overstated.
Then the "new new thing" is followed by a period of years where the innovation is refined, distributed, applied to different contexts, and incrementally improved.
The iPhone launched in 2007 is not really that much different than the one you have in your pocket today. The last 20 years has been about improvements. The web browser before that is also pretty much the same as the one you use today.
We've seen the same pattern happen with LLMs. The author of the article points out that many of AI's breakthroughs have been around since the 1990s. Sure! And the Internet was created in the 1970s and mobile phones were invented in the 1980s. That doesn't mean the web and smartphones weren't monumental technological events. And it doesn't mean LLMs and AI innovation is somehow not proceeding apace.
Each crawl on the internet is actually a discrete chunk of a more abstractly defined, constant influx of information streams. Let's call them rivers (it's a big stream).
These rivers can dry up, present seasonal shifts, be poisoned, be barraged.
It will never "get there" and gather enough data to "be done".
--
Regarding "new ideas in AI", I think there could be. But this whole thing is not about AI anymore.
What John Carmack is exploring is pretty revealing. Train models to play 2D video games to a superhuman level, then ask them to play a level they have not seen before or another 2D video game they have not seen before. The transfer function is negative. So, in my definition, no intelligence has been developed, only expertise in a narrow set of tasks.
It’s apparently much easier to scare the masses with visions of ASI, than to build a general intelligence that can pick up a new 2D video game faster than a human being.
Seeing comments here saying “this problem is already solved”, “he is just bad at this” etc. feels bad. He has given a long time to this problem by now. He is trying to solve this to advance the field. And needless to say, he is a legend in computer engineering or w/e you call it.
It should be required to point to the “solution” and maybe how it works to say “he just sucks” or “this was solved before”.
IMO the problem with current models is that they don’t learn categorically like: lions are animals, animals are alive. goats are animals, goats are alive too. So if lions have some property like breathing and goats also have it, it is likely that other similar things have the same property.
Or when playing a game, a human can come up with a strategy like: I’ll level this ability and lean on it for starting, then I’ll level this other ability that takes more time to ramp up while using the first one, then change to this play style after I have the new ability ready. This might be formulated completely based on theoretical ideas about the game, and modified as the player gets more experience.
With current AI models as far as I can understand, it will see the whole game as an optimization problem and try to find something at random that makes it win more. This is not as scalable as combining theory and experience in the way that humans do. For example a human is innately capable of understanding there is a concept of early game, and the gains made in early game can compound and generate a large lead. This is pattern matching as well but it is on a higher level .
Theory makes learning more scalable compared to just trying everything and seeing what works
He is not using appropriate models for this conclusion and neither is he using state of the art models in this research and moreover he doesn't have an expensive foundational model to build upon for 2d games. It's just a fun project.
A serious attempt at video/vision would involve some probabilistic latent space that can be noised in ways that make sense for games in general. I think veo3 proves that ai can generalize 2d and even 3d games, generating a video under prompt constraints is basically playing a game. I think you could prompt veo3 to play any game for a few seconds and it will generally make sense even though it is not fine tuned.
Veo3's world model is still pretty limited. That becomes obvious very fast once you prompt out of distribution video content (i.e. stuff that you are unlikely to find on youtube). It's extremely good at creating photorealistic surfaces and lighting. It even has some reasonably solid understanding of fluid dynamics for simulating water. But for complex human behaviour (in particular certain motions) it simply lacks the training data. Although that's not really a fault of the model and I'm pretty sure there will be a way to overcome this as well. Maybe some kind of physics based simulation as supplement training data.
Is any model currently known to succeed in the scenario that Carmack’s inappropriate model failed?
No monolithic models but us ng hybrid approaches we've been able to beet humans for some time now.
To confirm: hybrid approaches can demonstrate competence at newly-created video games within a short period of exposure, so long as similar game mechanics from other games were incorporated into their training set?
What you're thinking of is much more like the Genie model from DeepMind [0]. That one is like Veo, but interactive (but not publically available)
[0] https://deepmind.google/discover/blog/genie-2-a-large-scale-...
> I think veo3 proves that ai can generalize 2d and even 3d games, generating a video under prompt constraints is basically playing a game.
In the same way that keeping a dream journal is basically doing investigative journalism, or talking to yourself is equivalent to making new friends, maybe.
The difference is that while they may both produce similar, "plausible" output, one does so as a result of processes that exist in relation to an external reality.
> generating a video under prompt constraints is basically playing a game
Besides static puzzles (like a maze or jigsaw) I don't believe this analogy holds? A model working with prompt constraints that aren't evolving or being added over the course of "navigating" the generation of the model's output means it needs to process 0 new information that it didn't come up with itself — playing a game is different from other generation because it's primarily about reacting to input you didn't know the precise timing/spatial details of, but can learn that they come within a known set of higher order rules. Obviously the more finite/deterministic/predictably probabilistic the video game's solution space, the more it can be inferred from the initial state, aka reduce to the same type of problem as generating a video from a prompt), which is why models are still able to play video games. But as GP pointed out, transfer function negative in such cases — the overarching rules are not predictable enough across disparate genres.
> I think you could prompt veo3 to play any game for a few seconds
I'm curious what your threshold for what constitutes "play any game" is in this claim? If I wrote a script that maps button combinations to average pixel color of a portion of the screen buffer, by what metric(s) would veo3 be "playing" the game more or better than that script "for a few seconds"?
edit: removing knee-jerk reaction language
It's not ideal, but you can prompt it with an image of a game frame, explain the objects and physics in text and let it generate a few frames of gameplay as a substitute for controller input as well as what it expects as an outcome. I am not talking about real interactive gameplay.
I am just saying we have proof that it can understand complex worlds and sets of rules, and then abide by them. It doesn't know how to use a controller and it doesn't know how to explore the game physics on its own, but those steps are much easier to implement based on how coding agents are able to iterate and explore solutions.
[flagged]
fair, and I edited my choice of words, but if you're reading that much aggression from my initial comment (which contains topical discussion) to say what you did, you must find the internet a far more savage place than it really is :/
[flagged]
> I think veo3 proves that ai can generalize 2d and even 3d games
It doesn't. And you said it yourself:
> generating a video under prompt constraints is basically playing a game.
No. It's neither generating a game (that people can play) nor is it playing a game (it's generating a video).
Since it's not a model of the world in any sense of the word, there are issues with even the most basic object permanenece. E.g. here's veo3 generating a GTA-style video. Oh look, the car spins 360 and ends up on a completely different street than the one it was driving down previously: https://www.youtube.com/watch?v=ja2PVllZcsI
It is still doing a great job for a few frames, you could keep it more anchored to the state of the game if you prompt it. Much like you can prompt coding agents to keep a log of all decisions previously made. Permanenece is excellent, it slips often but it mostly because it is not grounded to specific game state by the prompt or by the decision log.
I think we need a spatial/physics model handling movement and tactics watched over by a high level strategy model (maybe an LLM).
I wonder if this is a case of overfitting from allowing the model to grow too large, and if you might cajole it into learning more generic heuristics by putting some constraints on it.
It sounds like the "best" AI without constraint would just be something like a replay of a record speedrun rather than a smaller set of heuristics of getting through a game, though the latter is clearly much more important with unseen content.
The subject you are referring to is most likely Meta-Reinforcement Learning [1]. It is great that John Carmack is looking into this, but it is not a new field of research.
[1] https://instadeep.com/2021/10/a-simple-introduction-to-meta-...
I don't get why people are so invested in framing it this way. I'm sure there are ways to do the stated objective. John Carmack isn't even an AI guy why is he suddenly the standard.
Who is an "AI guy"? The field as we know it is fairly new. Sure, neural nets are old hat, but a lot has happened in the last few years.
John Carmack founded Keen technology in 2022 and has been working seriously on AI since 2019. From his experience in the video game industry, he knows a thing or two about linear algebra and GPUs, that is the underlying maths and the underlying hardware.
So, for all intent and purposes, he is an "AI guy" now.
But the logic seems flawed.
He has built an AI system that fails to do X.
That does not mean there isn't an AI system that can do X. Especially considering that a lot is happening in AI, as you say.
Anyway, Carmack knows a lot about optimizing computations on modern hardware. In practice, that happens to be also necessary for AI. However, it is not __sufficient__ for AI.
> That does not mean there isn't an AI system that can do X.
You are holding the burden of proof here...
This is exactly how Science works. He’s right until proven wrong. And so are you.
"He has built an AI system that fails to do X."
Perhaps you have put your finger on the fatal flaw ...
Keen includes researchers like Richard Sutton, Joseph Modayil etc. Also John has being doing it full time for almost 5 years now so given his background and aptitude for learning I would imaging by this time he is more of an AI guy then a fairly large percentage of AI PhDs.
Ah some No True Scotsman
Not sure why justanotherjoe is a credible resource on who is and isn’t expert in some new dialectic and euphemism for machine state management. You’re that nobody to me :shrug:
Yann LeCun is an AI guy and has simplified it as “not much more than physical statistics.”
WWhole lot of AI is decades old info theory books applied to modern computer.
Either a mem value is or isn’t what’s expected. Either an entire matrix of values is or isn’t what’s expected. Store the results of some such rules. There’s your model.
The words are made up and arbitrary because human existence is arbitrary. You’re being sold on a bridge to nowhere.
What in your opinion constitutes an AI guy?
Credentialism is bad, especially when used as a stick
Because it "confirms" what they already believe in.
Names >> all, and increasingly so.
One phenomena that bared this to me, in a substantive way, was noticing an increasing # of reverent comments re: Geohot in odd places here, that are just as quickly replied to by people with a sense of how he works, as opposed to the keywords he associates himself with. But that only happens here AFAIK.
Yapping, or, inducing people to yap about me, unfortunately, is much more salient to my expected mindshare than the work I do.
It's getting claustrophobic intellectually, as a result.
Example from the last week is the phrase "context engineering" - Shopify CEO says he likes it better than prompt engineering, Karpathy QTs to affirm, SimonW writes it up as fait accompli. Now I have to rework my site to not use "prompt engineering" and have a Take™ on "context engineering". Because of a couple tweets + a blog reverberating over 2-3 days.
Nothing against Carmack, or anyone else named, at all. i.e. in the context engineering case, they're just sharing their thoughts in realtime. (i.e. I don't wanna get rolled up into a downvote brigade because it seems like I'm affirming the loose assertion Carmack is "not an AI guy", or, that it seems I'm criticizing anyone's conduct at all)
EDIT: The context engineering example was not in reference to another post at the time of writing, now one is the top of front page.
> Now I have to rework my site to not use "prompt engineering" and have a Take™ on "context engineering". Because of a couple tweets + a blog reverberating over 2-3 days.
The difference here is that your example shows a trivial statement and a change period of 3 days, whereas what Carmack is doing is taking years.
Right. Nothing against Carmack. Grew up on the guy. I haven't looked into, at all, into any of the disputed stuff, and should actively proclaim I'm a yuge Carmack fanboy.
[dead]
These questions of whether the model is “really intelligent” or whatever might be of interest to academics theorizing about AGI, but to the vast swaths of people getting useful stuff out of LLMs, it doesn’t really matter. We don’t care if the current path leads to AGI. If the line stopped at Claude 4 I’d still keep using it.
And like I get it, it’s fun to complain about the obnoxious and irrational AGI people. But the discussion about how people are using these things in their everyday lives is way more interesting.
This generalization issue in RL in specific was detailed by OpenAI in 2018
https://arxiv.org/pdf/1804.03720
Can you please explain "the transfer function is negative"?
I'm wondering whether one has tested with the same model but on two situations:
1) Bring it to superhuman level in game A and then present game B, which is similar to A, to it.
2) Present B to it without presenting A.
If 1) is not significantly better than 2) then maybe it is not carrying much "knowledge", or maybe we simply did not program it correctly.
According to Carmack's recent talk [0], SOTA models that have been trained on game A don't perform better or train faster on game B. Even worse, training on game B negatively affects performance in game A when returning to it.
[0] https://www.youtube.com/watch?v=3pdlTMdo7pY
I think the problem is we train models to pattern match, not to learn or reason about world models
In the Physics of Language Models[1] they argue that you must augment your training data by changing sentences and such, in order for the model to be able to learn the knowledge. As I understand their argument, language models don't have a built-in way to detect what is important information and what is not, unlike us. Thus the training data must aid it by presenting important information in many different ways.
Doesn't seem unreasonable that the same holds in a gaming setting, that one should train on many variations of each level. Change the lengths of halls connecting rooms, change the appearance of each room, change power-up locations etc, and maybe even remove passages connecting rooms.
[1]: https://physics.allen-zhu.com/part-3-knowledge/part-3-1
I think this is clearly a case of over fitting and failure to generalize, which are really well understood concepts. We don't have to philosophize about what pattern matching really means.
In other words, they learn the game, not how to play games.
They memorize the answers not the process to arrive at answers
They learn the value of specific actions in specific contexts based on the rewards they received during their play time. Specific actions and specific contexts are not transferable for various reasons. John quoted that varying frame rates and variable latency between action and effect really confuse the models.
Okay, so fuzz the frame rate and latency? That feels very easy to fix.
Good point, you should write to John Carmack and let him know you've figured out the problem.
This has been disproven so many times... They clearly do both. You can trivially prove this yourself.
> You can trivially prove this yourself.
Given the long list of dead philosophers of mind, if you have a trivial proof, would you mind providing a link?
It’s really easy: go to Claude and ask it a novel question. It will generally reason its way to a perfectly good answer even if there is no direct example of it in the training data.
When LLM's come up with answers to questions that aren't directly exampled in the training data, that's not proof at all that it reasoned its way there — it can very much still be pattern matching without insight from the actual code execution of the answer generation.
If we were taking a walk and you asked me for an explanation for a mathematical concept I have not actually studied, I am fully capable of hazarding a casual guess based on the other topics I have studied within seconds. This is the default approach of an LLM, except with much greater breadth and recall of studied topics than I, as a human, have.
This would be very different than if we sat down at a library and I applied the various concepts and theorems I already knew to make inferences, built upon them, and then derived an understanding based on reasoning of the steps I took (often after backtracking from several reasoning dead ends) before providing the explanation.
If you ask an LLM to explain their reasoning, it's unclear whether it just guessed the explanation and reasoning too, or if that was actually the set of steps it took to get to the first answer they gave you. This is why LLMs are able to correct themselves after claiming strawberry has 2 rs, but when providing (guessing again) their explanations they make more "relevant" guesses.
LLMs clearly don't reason in the same way that humans or SMT solvers do. That doesn't mean they aren't reasoning.
How do you know it’s a novel question?
You have probably seen examples of LLMs doing the "mirror test", i.e. identifying themselves in screenshots and referring to the screenshot from the first person. That is a genuinely novel question as an "LLM mirror test" wasn't a concept that existed before about a year ago.
Elephant mirror tests existed, so it doesn’t seem all that novel when the word “elephant” could just be substituted for the word “LLM”?
The question isn't about universal novelty, but whether the prompt/context is novel enough such that the LLM answering competently demonstrates understanding. The claim of parroting is that the dataset contains a near exact duplicate of any prompt and so the LLM demonstrating what appears to be competence is really just memorization. But if an LLM can generalize from an elephant mirror test to an LLM mirror test in an entirely new context (showing pictures and being asked to describe it), that demonstrates sufficient generalization to "understand" the concept of a mirror test.
It's not exactly difficult to come up with a question that's so unusual the chance of it being in the training set is effectively zero.
Can you provide some examples of these genuinely unique questions?
And as any programmer will tell you: they immediately devolve into "hallucinating" answers, not trying to actually reason about the world. Because that's what they do: they create statistically plausible answers even if those answers are complete nonsense.
Just go and ask ChatGPT or Claude something that can't possibly be in its training set. Make something up. If it is only memorising answers then it will be impossible for it to get the correct result.
A simple nonsense programming task would suffice. For example "write a Python function to erase every character from a string unless either of its adjacent characters are also adjacent to it in the alphabet. The string only contains lowercase a-z"
That task isn't anywhere in its training set so they can't memorise the answer. But I bet ChatGPT and Claude can still do it.
Honestly this is sooooo obvious to anyone that has used these tools, it's really insane that people are still parroting (heh) the "it just memorises" line.
LLMs don't "memorize" concepts like humans do. They generate output based on token patterns in their training data. So instead of having to be trained on every possible problem, they can still generate output that solves it by referencing the most probable combination of tokens for the specified input tokens. To humans this seems like they're truly solving novel problems, but it's merely a trick of statistics. These tools can reference and generate patterns that no human ever could. This is what makes them useful and powerful, but I would argue not intelligent.
People who say that LLMs memorize stuff are just as clueless who assume that there's any reasoning happening.
They generate statistically plausible answers (to simplify the answer) based on the training set and weights they have.
What if that’s all we’re doing, though?
Well yeah... If you only ever played one game in your life you would probably be pretty shit at other games too. This does not seem very revealing to me.
I don't think thats true. If you'd only ever played Doom, I think you could play, say, counterstrike or half-life and be pretty good at it, and i think Carmack is right that its pretty interesting that this doesn't seem to be the case for ai models
I am decent at chess but barely know how the pieces in Go move.
Of course, this because I have spent a lot of time TRAINING to play chess and basically none training to play go.
I am good on guitar because I started training young but can't play the flute or piano to save my life.
Most complicated skills have basically no transfer or carry over other than knowing how to train on a new skill.
yeahhhh why isnt there a training structure where you play 5000 games, and the reward function is based on doing well in all of them?
I guess its a totaly different level of control: instead of immediately choosing a certain button to press, you need to set longer term goals. "press whatever sequence over this time i need to do to end up closer to this result"
There is some kind of nested multidimensional thing to train on here instead of immediate limited choices
I kinda think I'm more or less the same...OK maybe we have different definitions of "pattern matching".
It's Plato's cave:
We train the models on what are basically shadows, and they learn how to pattern match the shadows.
But the shadows are only depictions of the real world, and the LLMs never learn about that.
100%
Where do you draw the line between pattern matching and reasoning about world models?
A lot of intelligence is just pattern matching and being quick about it.
The line is: building an internal world model requires interfacing with the world, not a model of it, and subsequent failing (including death and survivorship over generations) and adaptation. Plus pattern matching.
Current AI only does one of those (pattern matching, not evolution), and the prospects of simulating evolution is kind of bleak, given I don’t think we can simulate a full living cell yet from scratch? Building a world model requires life (or something that has undergone a similar evolutionary survivorship path), not something that mimics life.
Just sounds like an example of overfitting. This is all machine learning at its root.
Indeed, it's nothing but function fitting.
I've wondered about the claim that the models played those Atari/2D video games at superhuman levels, because I clearly recall some humans achieving superhuman levels before models were capable of it. Must have been superhuman compared to average human player, not someone who spent an inordinate amount of time mastering the game.
I'm not sure why you think so. AI outperforms humans in many games already. Basically all the games we care to put money to train a model.
AI has beat the best human players in Chess, Go, Mahjong, Texas hold'em, Dota, Starcraft, etc. It would be really, really surprising that some Atari game is the holy grail of human performance that AI cannot beat.
I recall this not being true at all for Dota and Starcraft. I recall AlphaStar performed much better than the top non-pro players, but it couldn't consistently beat the pro players with the budget that Google was willing to spend, and I believe the same was true of Dota II (and there they were even playing a limited form of the game, with fewer heroes and without the hero choice part, I believe).
As I recall, the Starcraft ones heavily involved being able to exploit the computer's advantage in "twitch" speed over any human, it's just a slightly more complicated way of how any aim-bot enabled AI will always beat a human in an FPS, the game is designed to reward a certain amount of physical speed and accuracy.
In other words, the Starcraft AIs that win do so by microing every single unit in the entire game at the same time, which is pretty clever, but if you reduce them to interfacing with the game in the same way a human does, they start losing.
One of my pet peeves when we talk about the various chess engines is yes, given a board state they can output the next set of moves to beat any human, but can they teach someone else to play chess? I'm not trying to activate some kinda "gotcha" here, just getting at what does it actually mean to "know how to play chess". We'd expect any human that claimed to know how to play to be able to teach any other human pretty trivially.
When I finished my degree, the idea that a software system could develop that level of expertise was relegated to science fiction. It is an unbelievable human accomplishment to get to that point and honestly, a bit of awe makes life more pleasant.
Less quality of life focused, I don’t believe that the models he uses for this research are capable of more. Is it really that revealing?
I wonder how much performance decreases if they just use slightly modified versions of the same game. Like a different color scheme, or a couple different sprites.
this is what deepmind did 10 years ago lol
No, they (and many others before them) are genuinely trying to improve on the original research.
The original paper "Playing Atari with Deep Reinforcement Learning" (2013) from Deepmind describes how agents can play Atari games, but these agents would have to be specifically trained on every individual game using millions of frames. To accomplish this, simulators were run in parallel, and much faster than in real-time.
Also, additional trickery was added to extract a reward signal from the games, and there is some minor cheating on supplying inputs.
What Carmack (and others before him) is interested in, is trying to learn in a real-life setting, similar to how humans learn.
I'd say with confidence: we're living in the early days. AI has made jaw-dropping progress in two major domains: language and vision. With large language models (LLMs) like GPT-4 and Claude, and vision models like CLIP and DALL·E, we've seen machines that can generate poetry, write code, describe photos, and even hold eerily humanlike conversations.
But as impressive as this is, it’s easy to lose sight of the bigger picture: we’ve only scratched the surface of what artificial intelligence could be — because we’ve only scaled two modalities: text and images.
That’s like saying we’ve modeled human intelligence by mastering reading and eyesight, while ignoring touch, taste, smell, motion, memory, emotion, and everything else that makes our cognition rich, embodied, and contextual.
Human intelligence is multimodal. We make sense of the world through:
Touch (the texture of a surface, the feedback of pressure, the warmth of skin0; Smell and taste (deeply tied to memory, danger, pleasure, and even creativity); Proprioception (the sense of where your body is in space — how you move and balance); Emotional and internal states (hunger, pain, comfort, fear, motivation).
None of these are captured by current LLMs or vision transformers. Not even close. And yet, our cognitive lives depend on them.
Language and vision are just the beginning — the parts we were able to digitize first - not necessarily the most central to intelligence.
The real frontier of AI lies in the messy, rich, sensory world where people live. We’ll need new hardware (sensors), new data representations (beyond tokens), and new ways to train models that grow understanding from experience, not just patterns.
> Language and vision are just the beginning — the parts we were able to digitize first - not necessarily the most central to intelligence.
I respectfully disagree. Touch gives pretty cool skills, but language, video and audio are all that are needed for all online interactions. We use touch for typing and pointing, but that is only because we don't have a more efficient and effective interface.
Now I'm not saying that all other senses are uninteresting. Integrating touch, extensive proprioception, and olfaction is going to unlock a lot of 'real world' behavior, but your comment was specifically about intelligence.
Compare humans to apes and other animals and the thing that sets us apart is definitely not in the 'remaining' senses, but firmly in the realm of audio, video and language.
> Language and vision are just the beginning — the parts we were able to digitize first - not necessarily the most central to intelligence.
I probably made a mistake when i asserted that -- should have thought it over. Vision is evolutionarily older and more “primitive”, while language is uniquely human [or maybe, more broadly, primate, cetacean, cephalopod, avian...] symbolic, and abstract — arguably a different order of cognition altogether. But i maintain that each and every sense is important as far as human cognition -- and its replication -- is concerned.
Vision is interesting in that it leverages the maximum speed with which it is easily possible to gather information about our surroundings in this universe. I believe that is what makes it special and very valuable. I also believe this aspect makes it a strong attractor for convergent evolution.
Language allows encoding and compression of information about the world, which is of course incredibly powerful and increases communication bandwidth enormously (as well as tons of other stuff).
I'd say that for high level cognitive processes, hearing and speaking were an important stepping stone because for some reason evolving organs that can generate relatively high bandwidth signals in audio seems to be easier than evolving something that does that for visuals (very few Teletubby screens on tummies in the natural world).
Interesting games to think about in this sense: Pictionary/drawing games and charades.
People who lack one of those senses, or even two of them, tend to do just fine.
Mostly thanks to other humans helping them.
If all humans lacked vision, the human race would definitely not do just fine.
Organic adaption and persistence of memory I would say are the two major advancements that need to happen.
Human neural networks are dynamic, they change and rearrange, grow and sever. An LLM is fixed and relies on context, if you give it the right answer it won't "learn" that is the correct answer unless it is fed back into the system and trained over months. What if it's only the right answer for a limited period of time?
To build an intelligent machine, it must be able train itself in real time and remember.
Yes and: and forget.
> Language and vision are just the beginning..
Based on the architectures we have they may also be the ending. There’s been a lot of news in the past couple years about LLMs but has there been any breakthroughs making headlines anywhere else in AI?
> There’s been a lot of news in the past couple years about LLMs but has there been any breakthroughs making headlines anywhere else in AI?
Yeah, lots of stuff tied to robotics, for instance; this overlaps with vision, but the advances go beyond vision.
Audio has seen quite a bit. And I imagine there is stuff happening in niche areas that just aren't as publicly interesting as language, vision/imagery, audio, and robotics.
Two Nobel prizes in chemistry: https://www.nature.com/articles/s41746-024-01345-9
Sure. In physics, math, chemistry, biology. To name a few.
> The real frontier of AI lies in the messy, rich, sensory world where people live. We’ll need new hardware (sensors), new data representations (beyond tokens), and new ways to train models that grow understanding from experience, not just patterns.
Like Dr. Who said: DALEKs aren't brains in a machine, they are the machine!
Same is true for humans. We really are the whole body, we're not just driving it around.
There are many people who mentally developed while paralyzed that literally drive around their bodies via motorized wheelchair. I don't think there's any evidence that a brain couldn't exist or develop in a jar, given only the inputs modern AI now has (text, video, audio).
> any evidence that a brain couldn't exist or develop in a jar
The brain could. Of course it could. It's just a signals processing machine.
But would it be missing anything we consider core to the way humans think? Would it struggle with parts of cognition?
For example: experiments were done with cats growing up in environments with vertical lines only. They were then put in a normal room and had a hard time understanding flat surfaces.
https://computervisionblog.wordpress.com/2013/06/01/cats-and...
This isn't remotely a hypothetical, so I imagine there are some examples out there, especially from back when polio was a problem. Although, for practical reasons, they might have had limited exposure to novelty, which could have negative consequences.
I agree it’s not hypothetical and also as a layperson I don’t know how much impact on cognition has been studied. Would be cool if it has!
I do know of studies that showed blind people start using their visual cortex to process sounds. That is pretty cool imo
Yeah, but are there new ideas or only wishes?
It’s pure magical thinking that would be correctly dismissed if it didn’t have AI attached to it. Imagine talking this way about anything else.
“We’ve barely scratched the surface with Rust, so far we’re only focused on code and haven’t even explored building mansions or ending world hunger”
AI has some real possibilities of building mansions and ending hunger in a way that Rust doesn't.
> has made jaw-dropping progress
They took 1970s dead tech and deployed it on machines 1 million times more powerful. I'm not sure I'd qualify this as progress. I'd also need an explanation as to what systemic improvements in models and computations that give an exponential growth in performance are planned.
I don't see anything.
Winning two Nobel prizes wasn't enough progress?
> They took 1970s dead tech and deployed it on machines 1 million times more powerful. I’m not sure I’d qualify this as progress
If this isn’t meant to be sarcasm or irony, you’ve got some really exciting research and learning ahead of you! At the moment it reads very “computers are just addition and multiplication and we’ve had that for thousands of years!”
Sometimes we get confused by the difference between technological and scientific progress. When science makes progress it unlocks new S-curves that progress at an incredible pace until you get into the diminishing returns region. People complain of slowing progress but it was always slow, you just didn’t notice that nothing new was happening during the exponential take off of the S-curve, just furious optimization.
Fully agree.
And at the same time I have noticed that people don’t understand the difference between an S-curve and an exponential function. They can look almost identical at certain intervals.
As far back as 2017 I copped a lot of flak for suggesting that the coming automation revolution will be great at copying office workers and artists but wont be in order of replacing the whole human race. A lot of the time moores law got thrown back in my face. But thats how this works, we unlock something new, we exploit it as far as possible, the shine wears off and we deal with the aftermath.
The crypto mind cannot comprehend
You're being awfully generous to describe basic hype as "technological progress."
[dead]
If you work with model architecture and read papers, how could not know there are a flood of new ideas? Only few yield interesting results though.
I kind of wonder if libraries like pytorch have hurt experimental development. So many basic concepts no one thinks about anymore because they just use the out of the box solutions. And maybe those solutions are great and those parts are "solved", but I am not sure. How many models are using someone else's tokenizer, or someone else's strapped on vision model just to check a box in the model card?
The hardware(GPU)'s architectural limitations may slow research more than PyTorch. The hardware lottery https://hardwarelottery.github.io/
That's been the very normal way of the human world.
When the foundation layer at a given moment doesn't yield an ROI on intellectual exploration - say because you can overcompensate with VC funded raw compute and make more progess elsewhere -, few(er) will go there.
But inevitably, as other domains reach diminishing returns, bright minds will take a look around where significant gains for their effort can be found.
And so will the next generation of PyTorch or foundational technologies evolve.
The people who don't think about such things probably wouldn't develop experimentally sans pytorch either.
Yeah and even then, it's been like ~ 2-3 years since the last rather major Architectural improvement, major enough for a lot of people to actually hear about it and use it daily. I think some people lose perspective on how short of a time frame 3 years is.
But yes, there's a ton of interesting and useful stuff (beyond datasets and data related improvements) going on right now, and I'm not even talking about LLMs. I don't do anything related to LLM and even then I still see tons of new stuff popping up regularly.
It's the opposite.
Frameworks like pytorch are really flexible. You can implement any architecture, and if it's not enough, you can learn CUDA.
Keras it's the opposite, it's probably like you describe things.
To be fair, if you imagine a system that successfully reproduced human intelligence, then 'changing datasets' would probably be a fair summary of what it would take to have different models. After all, our own memories, training, education, background, etc are a very large component of our own problem solving abilities.
I will respectfully disagree. All "new" ideas come from old ideas. AI is a tool to access old ideas with speed and with new perspectives that hasn't been available up until now.
Innovation is in the cracks: recognition of holes, intersections, tangents, etc. on old ideas. It has bent said that innovation is done on the shoulders of giants.
So AI can be an express elevator up to an army of giant's shoulders? It all depends on how you use the tools.
Access old ideas? Yes. With new perspectives? Not necessarily. An LLM may be able to assist in interpreting data with new perspectives but in practice they're still fairly bad at greenfield work.
As with most things, the truth lies somewhere in the middle. LLMs can be helpful as a way of accelerating certain kinds and certain aspects of research but not others.
> Access old ideas? Yes. With new perspectives?
I wonder if we can mine patent databases for old ideas that never worked out in the past, but now are more useful. Perhaps due to modern machining or newer materials or just new applications of the idea.
Imagine a human had read every book/publication in every field of knowledge that mankind has ever produced AND couldn’t come up with anything entirely new. Hard to imagine.
The difficult part is proposing an experiment to check a new idea.
My hypothesis of the mismatch is centered around "read" - I think that when you wrote it, and when others similarly think about that scenario, the surprise is because our version of "read" is the implied "read and internalized" or at bare minimum "read for comprehension" but as very best I can tell the LLM's version is "encoded tokens into vector space" and not "encoded into semantic graph"
I welcome the hair-splittery that is sure to follow about what it means to "understand" anything
It is possible that such a human wouldn't come up with anything new, even if they could.
The article is discussing working in AI innovation vs focusing on getting more and better data. And while there have been key breakthroughs in new ideas, one of the best ways to increase the performance of these systems is getting more and better data. And how many people think data is the primary avenue to improvement.
It reminds me of an AI talk a few decades ago, about how the cycle goes: more data -> more layers -> repeat...
Anyways, I'm not sure how your comment relates to these two avenues of improvement.
> I will respectfully disagree. All "new" ideas come from old ideas.
The insight into the structure of the benzene ring famously came in a dream, hadn't been seen before, but was imagined as a snake bitings its own tail.
And as we all know, it came in a dream to a complete novice in chemistry with zero knowledge of any old ideas in chemistry: https://en.wikipedia.org/wiki/August_Kekul%C3%A9
--- start quote ---
The empirical formula for benzene had been long known, but its highly unsaturated structure was a challenge to determine. Archibald Scott Couper in 1858 and Joseph Loschmidt in 1861 suggested possible structures that contained multiple double bonds or multiple rings, but the study of aromatic compounds was in its earliest years, and too little evidence was then available to help chemists decide on any particular structure.
More evidence was available by 1865, especially regarding the relationships of aromatic isomers.
[ Kekule claimed to have had the dream in 1865 ]
--- end quote ---
The dream claim came from Kekule himself 25 years after his proposal that he had to modify 10 years after he proposed it.
Exactly!
Can you imagine if we applied the same gatekeeping logic to science?
Imagine you weren't allowed to use someone else's scientific work or any derivative of it.
We would make no progress.
The only legitimate defense I have ever seen here revolves around IP and copyright infringement, which I couldn't care less about.
Man I can't wait for this '''''AI''''' stuff to blow over. The back and forth gets a bit exhausting.
What about actively obtained data - models seeking data, rather than being fed. Human babies put things in their mouths, they try to stand and fall over. They “do stuff” to learn what works. Right now we’re just telling models what works.
What about simulation: models can make 3D objects so why not give them a physics simulator? We have amazing high fidelity (and low cost!) game engines that would be a great building block.
What about rumination: behind every Cursor rule for example, is a whole story of why a user added it. Why not take the rule, ask a reasoning model to hypothesize about why that rule was created, and add that rumination (along with the rule) to the training data. Providing opportunities to reflect on the choices made by their users might deepen any insights, squeezing more juice out of the data.
Simulation and embodied AI (putting the AI in a robotic arm or a car so it can try stuff and gather information about the results) are very actively being explored.
What about at inference time? ie. in response to a query.
We let models write code and run it. Which gives them a high chance of getting arithmetic right.
Solving the “crossing the river” problem by letting the model create and run a simulation would give a pretty high chance of getting it right.
That would be reinforcement learning. The juice is quite hard to squeeze.
Agreed for most cases.
Each Cursor rule is a byproduct of tons of work and probably contains lots that can be unpacked. Any research on that?
Yeah, at a very high level it's similar to an actor-critic reinforcement learning algorithm. The rule text is a value function and one could build a critic model that takes as input the rule text and the main model's (the actor's) output to produce a reward.
This is easier said than done though because this value function is so noisy it's often hard to learn from it. And also whether or not a response (the model output) matches the value function (the Cursor rules) is not even that easy to grade. It's been easier to train the chain-of-thought style reasoning since one can directly score it via the length of thinking.
This new paper covers some of the difficulties of language-based critic models: https://openreview.net/pdf?id=0tXmtd0vZG
Generally speaking, the algorithm and approach is not new. Being able to do it in a reasonable amount of compute is the new part.
The latest LLMs are simply multiplying and adding various numbers together... Babylonians were doing that 4000 years ago.
You are just a lot of interactions of waves. All meaning is assigned. I prefer to think of this like the Goedel generator that found new formal expressions for the Principia - because we have a way of indexing concept-space, there's no telling what we might find in the gaps.
But on clay tables, not in semi-conductive electron prisons separated by one-atom-thick walls.
Slight difference to those methods, wouldn't you agree?
No it's exactly the same. Everything old is new again...
Reinforcement learning from self-play/AlphaWhatever? Nah must just be datasets. :)
And architecture stuff like actually useful long context. Whatever they did with gemini 2.5 is miles ahead in long context useful results compared to the previous models. I'd be very surprised if gemini 2.5 is "just" gemini 1 w/ better data.
i dont know what all the hype is with gemini 2.5, at least the currently running instance. from my experience at least in conversation mode, it cannot remember my instructions to avoid apologies and similar platitudes from either the “persona”, personal instructions, or from ine message to the next.
https://en.wikipedia.org/wiki/TD-Gammon
For that matter, https://gist.github.com/deebs67/8fbcf8b127a63e70d4a3f8590c97... .
You raise a really interesting point. I'm sure it's just missed my notice, but I'm not familiar with any projects from antediluvian AI that have been resurrected to run on modern hardware and see where they'd really asymptote if they'd had the compute they deserved.
This paper “were RNNs all we needed?” explores this hypothesis a bit, finding that some pre-transformer sequence models can match transformers when trained at appropriate scale. Though they did have to make some modifications to unlock more parallelism
https://arxiv.org/abs/2410.01201
To be fair, usually those projects would need considerable work to be ported to modern multicore machines, let alone GPUs.
can you name a couple so i can see how much work is involved? markov chains compile fast and respond fast, sure, and neural nets train pretty quick too, so i'm wondering where the cutoff is; expert systems?
Self-play gives you a large explosion of data.
Big difference between a perfect information, completely specified zero sum game and the real world.
As a simple analogy, read out the following sentence multiple times, stressing a different word each time.
"I never said she stole my money"
Note how the meaning changes and is often unique?
That is a lens I to the frame problem and it's inverse, the specification problem.
The above problem quickly becomes tower-complete, and recent studies suggest that RL is reinforcing or increasing the weight of existing patterns.
As the open domain frame problem and similar challenges are equivalent to HALT, finding new ways to extract useful information will be important for generalization IMHO.
Synthetic data is useful, but not a complete solution, especially for tower problems.
The one we use is "I always pay my taxes"
and as far as synthetic vs real data, there's a lot of gaps in LLM knowledge; and vision models suffer from "limited tags", which used to have workarounds with textual embeddings and the like, but those went by the wayside as LoRA, controlnet, etc. appeared.
There's people who are fairly well known that LLMs have no idea about. There's things in books i own that the AI confidently tells me are either wrong or don't exist.
That one page about compressing 1 gig wikipedia as small as possible implicitly and explicitly states that AI is "basically compression" - and if the data isn't there, it's not in the compressed set (weights) either.
And i'll reply to another comment here, about "24/7 rolling/ for looped" AI - i thought of doing this when i first found out about LLMs, but context windows are the enemy, here. I have a couple of ideas about how to have a continuous AI, but i don't have the capital to test it out.
There are new ideas, people are finding new ways to build vision models, which then are applied to language models and vice versa (like diffusion).
The original idea of connectionism is that neural networks can represent any function, which is the fundamental mathematical fact. So we should be optimistic, neural nets will be able to do anything. Which neural nets? So far people have stumbled on a few productive architectures, but it appears to be more alchemy than science. There is no reason why we should think there won't be both new ideas and new data. Biology did it, humans will do it too.
> we’re engaged in a decentralized globalized exercise of Science, where findings are shared openly
Maybe the findings are shared, if they make the Company look good. But the methods are not anymore
AI training is currently a process of making the AI remember the dataset. It doesn't involve the AI thinking about the dataset and drawing (and remembering) conclusions.
It can probably remember more facts about a topic than a PhD in that topic, but the PhD will be better at thinking about that topic.
Maybe that's why PhDs keep the textbooks they use at hand, so they don't have to remember everything.
Why should the model need to memorize facts we already have written down somewhere?
Its a bit more complex than that. Its more about baking out the dataset into heuristics that a machine can use to match a satisfying result to an input. Sometimes these heuristics are surprising to a human and can solve a problem in a novel way.
"Thinking" is too broad a term to apply usefully but I would say its pretty clear we are not close to AGI.
> It can probably remember more facts about a topic than a PhD in that topic
So can a notebook.
If datasets are what we are talking about, I'd like to bring attention to the biological datasets out there that have yet to be fully harnessed.
The ability to collect gene expression data at a tissue specific level has only been invented and automated in the last 4-5 years (see 10X Genomics Xenium, MERFISH). We've only recently figured out how to collect this data at the scale of millions of cells. A breakthrough on this front may be the next big area of advancement.
I don't know if it matters. Even if the best we can do is get really good at interpolating between solutions to cognitive tasks on the data manifold, the only economically useful human labor left asymptotes toward frontier work; work that only a single-digit percentage of people can actually perform.
Hardware isn't even close to being out of steam. There are some breathtakingly obvious premature optimizations that we can undo to get at least 99% power reduction for the same amount of compute.
For example, FPGAs use a lot of area and power routing signals across the chip. Those long lines have a large capacitance, and thus cause a large amount of dynamic power loss. So does moving parameters around to/from RAM instead of just loading up a vast array of LUTs with the values once.
An interesting step forward, although an old idea we seem close to is recursive self improvement. Get the AI to make a modified version of itself to try to think better.
Inbreeding is illegal for a reason
There are new things being tested and yielding results monthly in modelling. We've deviated quite a bit from the original multi head attention.
Here's an idea: make the AIs consistent at doing things computers are good at. Here's an anecdote from a friend who's living in Japan:
> i used chatgpt for the first time today and have some lite rage if you wanna hear it. tldr it wasnt correct. i thought of one simple task that it should be good at and it couldnt do that.
> (The kangxi radicals are neatly in order in unicode so you can just ++ thru em. The cjks are not. I couldnt see any clear mapping so i asked gpt to do it. Big mess i had to untangle manually anyway it woulda been faster to look them up by hand (theres 214))
> The big kicker was like, it gave me 213. And i was like, "why is one missing?" Then i put it back in and said count how many numbers are here and it said 214, and there just werent. Like come on you SHOULD be able to count.
If you can make the language models actually interface with what we've been able to do with computers for decades, i imagine many paths open up.
Many of us have solved this with internal tooling that has not yet been shared or released to the public.
This needs to be generalized however. For example, if you present an AI with a drawing of some directed graph (a state diagram, for example), it should be able to answer questions based on the precise set of all possible paths in that graph, without someone having to write tooling for diagram or graph processing and traversal. Or, given a photo of a dropped box of matches, an AI should be able to precisely count the matches, as far as they are individually visible (which a human could do by keeping a tally while coloring the matches). There are probably better examples, these are off the cuff.
There’s an infinite repertoire of such tasks that combine AI capabilities with traditional computer algorithms, and I don’t think we have a generic way of having AI autonomously outsource whatever parts require precision in a reliable way.
What you're describing sounds like agentic tool usage. Have you kept up with the latest developments on that? it's already solved depending on how strict you define your criteria above
My understanding is that you need to provide and configure task-specific tools. You can’t combine the AI with just a general-purpose computer and have the AI figure out on its own how to make use of it to achieve with reliability and precision whatever task it is given. In other words, the current tool usage isn’t general-purpose in the way the LLM itself is, and also the LLM doesn’t reason about its own capabilities in order to decide how to incorporate computer use to compensate for its own weaknesses. Instead you have to tell the LLM what it should apply the tooling for.
> If data is the only thing that matters, why are 95% of people working on new methods?
Because new methods unlock access to new datasets.
Edit: Oh I see this was a rhetorical question answered in the next paragraph. D'oh
I wrote about it around a year ago here:
"There weren't really any advancements from around 2018. The majority of the 'advancements' were in the amount of parameters, training data, and its applications. What was the GPT-3 to ChatGPT transition? It involved fine-tuning, using specifically crafted training data. What changed from GPT-3 to GPT-4? It was the increase in the number of parameters, improved training data, and the addition of another modality. From GPT-4 to GPT-40? There was more optimization and the introduction of a new modality. The only thing left that could further improve models is to add one more modality, which could be video or other sensory inputs, along with some optimization and more parameters. We are approaching diminishing returns." [1]
10 months ago around o1 release:
"It's because there is nothing novel here from an architectural point of view. Again, the secret sauce is only in the training data. O1 seems like a variant of RLRF https://arxiv.org/abs/2403.14238
Soon you will see similar models from competitors." [2]
Winter is coming.
1. https://news.ycombinator.com/item?id=40624112
2. https://news.ycombinator.com/item?id=41526039
And when winter does arrive, then what? The technology is slowing down while its popularity picks up. Can sparks fly out of snow?
the trillion dollar funding tap is turned off, the prices charged then will have to reflect the costs
shortly thereafter the entire ecosystem will collapse
This seems simplistic, tech and infrastructure play a huge part here. A short and incomplete list of things that contributed:
- Moore's law petering out, steering hardware advancements towards parallelism
- Fast-enough internet creating shift to processing and storage in large server farms, enabling both high-cost training and remote storage of large models
- Social media + search both enlisting consumers as data producers, and necessitating the creation of armies of Mturkers for content moderation + evaluation, later becoming available for tagging and rlhf
- A long-term shift to a text-oriented society, beginning with print capitalism and continuing through the rise of "knowledge work" through to the migration of daily tasks (work, bill paying, shopping) online, that allows a program that only produces text to appear capable of doing many of the things a person does
We may have previously had the technical ideas in the 1990s but we certainly didn't have the ripened infrastructure to put them into practice. If we had the dataset to create an LLM in the 90s, it still would have been astronomically cost-prohibitive to train, both in CPU and human labor, and it wouldn't have as much of an effect on society because you wouldn't be able to hook it up to commerce or day-to-day activities (far fewer texts, emails, ecommerce).
if datasets are the new codebases ,then the real IP can be dataset version control. how you fork ,diff ,merge and audit datasets like code. every team says 'we trained on 10B tokens' but what if we can answer 'which 5M tokens made reasoning better', 'which 100k made it worse'. then we can start being targeted leverage
Why is DeepSeek specifically called out?
disagree, there are a few organisations exploring novel paths. It's just that throwing new data at an "old" algorithm is much easier and has been a winning strategy. And, also, there's no incentive for a private org to advertise a new idea that seems to be working (mine's a notable exception :D).
Until these "AI" systems become always-on, always-thinking, always-processing, progress is stuck. The current push button AI - meaning it only processes when we prompt it - is not how the kind of AI that everyone is dreaming of needs to function.
From a technical perspective, we can do that with a for loop.
The reason we don't do it isn't because it's hard, it's because it yields worse results for increased cost.
Things haven't changed much in terms of truly new ideas since electricity was invented. Everything else is just applications on top of that. Make the electrons flow in a different way and you get a different outcome.
> Make the electrons flow in a different way and you get a different outcome.
This happens to be the basis of every aspect of our biology.
This brings us to the problem AI companies are facing, the lack of data, they have already hoovered as much as they can from the internet and desperately need more data.
Which make sit blatantly obvious why we're beginning to see products being marketed under the guise of assistants/tools to aid you whose actual purpose is to gather real world picture and audio data, think meta glasses and what Ives and Altman are cooking up with their partnership.
Paradigm shifts are often just a conglomeration of previous ideas with one little tweak that suddenly propels a technology ahead 10x which opens up a whole new era.
The iPhone is a perfect example. There were smartphones with cameras and web browsers before. But when the iPhone launched, it added a capacitive touch screen that was so responsive there was no need for a keyboard. The importance of that one technical innovation can't be overstated.
Then the "new new thing" is followed by a period of years where the innovation is refined, distributed, applied to different contexts, and incrementally improved.
The iPhone launched in 2007 is not really that much different than the one you have in your pocket today. The last 20 years has been about improvements. The web browser before that is also pretty much the same as the one you use today.
We've seen the same pattern happen with LLMs. The author of the article points out that many of AI's breakthroughs have been around since the 1990s. Sure! And the Internet was created in the 1970s and mobile phones were invented in the 1980s. That doesn't mean the web and smartphones weren't monumental technological events. And it doesn't mean LLMs and AI innovation is somehow not proceeding apace.
It's just how this stuff works.
I mean there's no new ideas for saas but just new applications and that worked out pretty well
[dead]
[flagged]
Everyone knows we'll always need horseshoes /s
Hmmm
Dataset? That's so 2000s.
Each crawl on the internet is actually a discrete chunk of a more abstractly defined, constant influx of information streams. Let's call them rivers (it's a big stream).
These rivers can dry up, present seasonal shifts, be poisoned, be barraged.
It will never "get there" and gather enough data to "be done".
--
Regarding "new ideas in AI", I think there could be. But this whole thing is not about AI anymore.