Seeing the dramatic differences in scores just going from high to xhigh is just another demonstration of the bitter lesson: Just keep scaling search and learning.
We are probably going to need a lot more GPUs.
GPT-6 is supposed to be using a much larger base model that just finished pretraining so the "dump another chunk of all written language" approach is still going strong.
Modern pretraining also consists of expensive human-led specialized task creation and grading loops. Synthetic generation and distillation from previous models is another input for training. I wonder how much new text contributes beyond keeping knowledge up-to-date.
A lot of data these days is synthetically generated. As an example, to make a model good at understanding assembly you simply need to round-trip code through a compiler and disassembler, then train against the source of truth and the assembly. You can generate arbitrary algebra expressions and have it solve it.
A lot of pretraining is also choosing the right type of data, you don't want to just have it ingest garbage (although I read that some amount of garbage actually helps the model be more robust). Pretraining crystallizes a lot of the inductive biases that post-training builds on, so by crafting the right data mixture you can make it easier for it to start off with a good foundation. There is also a lot of focus on mid-training these days, which I understand is basically either the name for the synthetic data stage, or the SFT phase before all the RL
Not sure where I implied they are "raw base models" and not sure what "various adjustments" means here or how "ton of RLHF" contradicts anything. If we look at research for open source models, "adjustments" usually come in the form of efficiency gains which directly contributes to the ability to scale or synthetic data pipelines to increase the dataset and increasing the context window.
Not just RLHF but also RLVR, and isn't that the litter lesson though?
My sense of the Sutton Dwarkesh interview was that he was calling out that he didn't mean just longer datasets, but rather learning through exploration and that's exactly RL.
The scaling with reasoning models is more and more with things like verifiable rewards (coding and math), in line with bitter lesson and also Sutton invented lots of modern RL.
Nah, the last few generations have more RLVR in the data mix. Which is more CPU intensive and very much amenable to the bitter lesson as you can reduce the loss by doing more rollouts in your tool environment.
While I think this is true, remember as we get more efficient we just decide to scale even bigger. So more GPUs, and more efficient.
I agree with the sibling comment, effiency is probably the more important component at this point. We are hitting not just a practical engineering roadblock for scaling with current technology, I think we have definitely hit a financial and logistical roadblock for up scaling with the number of GPUs (on an immediate basis)
Sam Altman was saying we'll eventually need to build a dyson sphere. We'll see how far human society can withstand this kind of escalation. I'm guessing one more year before there are major fractures in the basic economic/social/political relationship that get so bad it actually prevents more buildout.
I actually think we're in a strange situation with AI compute.
Right now, we have models that are statistical models of language, with a world model and reasoning "falling out" of a lot of effort.
It's like we've made something that's a little bit intelligent, and now we're trying to amplify that trick to create something that's quite intelligent. And - don't get me wrong - it works.
But it's also super, super inefficient. We're having machines "think out loud" to compensate for the quality of their thought processes. We elongate the path to make up for the progress made on a given step.
I tink there's probably a much smarter way of doing things that will require qualitative architectural (and quite possibly hardware) innovations. Right now we're on the path to a Dyson sphere: that's probably not going to be necessary once we figure out a smarter way to think.
I agree. But I think you're missing that LLMs can internalise a lot of the thinking process in their layers without explicit CoT. That System 1-style reasoning is bounded depth computation but very, very broad. Yudkowsky called it "cached thoughts" and I think it's an incredibly important idea [1]. It's really stiking how the best LLMs don't even need to think where smaller LLMs do.
So as more thinking is cached in their weights through increased RL training, those weights are doing more useful work and the efficiency is increasing.
Kind of refreshing though that the "throw more processing at it" scaling we saw in the 90s has returned in a different way. For a while we were really bottlenecked in our advances by relatively low levels of parallelism (most software used by your average user doesn't scale cleanly with more than a few threads).
Not always, in some cases, changing to a higher reasoning makes the AI doubt itself too much, and skip over the correct answer by overcomplicating the problem and polluting the context.
It would be nice to see on which categories of problems the extra thinking makes it better and on which it makes it worse.
This shows up in OpenAI's graphs on their announcement page. There is a peak performance datapoint in the graphs past which (to the right on the graph indicating more resources spent) peformance declines. And it's on every graph on that page!
Moreover we've known for quite a while now that glial cells also participate in cognition and moderate learning (e.g.: [1]). When you take those connections into account the numbers get really staggering. 85 billion glial cells with trillions of protein channels facilitating communication between the glial syncytium [2].
The real question is not how many "weights" the human brain has (neurons+synapses may or may not translate into "weights", and brain might be also inefficient for what it is), but rather how much evolutionary and social "compute" was necessary to pack everything into that capacity.
It’s still 20W. We have living proof what is possible within 20W. The message has always been clear - try to get silicon computer to be as power efficient as the brain is as it is obviously possible.
I keep thinking about the cyborg chicken mechs in Metal Gear Solid 4. If stuff like Neuralink gets advanced enough, we might find that cutting up animals and sticking electrodes in their brains is the best way to make robots.
Yeah, people might object, but it can be argued that we are already subjecting scores of animals to horrors beyond comprehension just to get a bucket of chicken wings. And even if we manage to get silicon to do what brains do, it will likely cost 1000x as much and consume 1000x the power like you said.
I think you're helping GPs point: there is a lot of efficiency gains to be made to match the processing power of the brain, given it's size and power draw.
For intelligence, I expect the next breakthrough to be colocation of memory and compute in the same chip. And we'll need much more of this memory, probably a few petabytes.
I mean, theoretically you can solve every finitary problem with a brute force solution...
Richard Sutton specifically states that the search has to be smart. We know that the brain uses recurrent connections and is shallow. I think a lot more money has to go into architecture. Feed Forward transformers can only scale so far
This isn’t really how it works anymore. Agents rely heavily on tool use and the agentic harness to perform tasks. Pre-training is no longer very effective.
I said a few months ago, "man, Opus is great, but sometimes when talking with it I have the feeling like, this thing should be about 10 times bigger."
When Mythos was announced after that, I was pleasantly surprised to hear about it. But when it turned out to be only two times bigger, I was a little disappointed!
(I am even more disappointed with the safety filters, but that's kind of a separate discussion... "Fortunately" I find that I can usually edit my prompt by single character and get through...)
We have it slightly ahead of Fable in our multi-agent coding evaluations.
Fable's main advantage is that its average solution size is smaller. However, GPT 5.6 Sol is a substantial improvement from GPT 5.4/5.5 which would write verbose, defensive code. 31KB for GPT 5.4/5.5 down to 26KB for GPT 5.6 Sol, with better performance for Sol.
Fable scores slightly lower, but with an average solution size of 12.2 KB.
This looks like a good benchmark. Time and time again I keep giving OpenAI models the chance to win me back, but Opus (and Fable especially) just writes more elegant code and is a significantly more productive rubber duck for interactive discussions. I feel vindicated seeing your description of verbose and defensive code, and I’m a bit disappointed that 5.6 Sol’s solution is still >5x longer than the human solution and 2x as verbose as Fable’s. Do you have any insight whether any of that is comments?
I wonder why nobody has tried to optimize for actual code size or complexity metric, or at least why I haven’t seen more benchmarks that display this. GPT5.5 just keeps pushing more and more pointless indirection into every function it writes in my main project, it’s borderline negative productivity.
The human solutions are all written in Python, which creates a significant length bias, whereas the AI models are assigned to create solutions randomly distributed across 11 relevant programming languages, most of which are inherently more verbose than Python.
I have not broken down the comment/code ratio, but that's actually a really interesting idea for a metric.
I would also like to test Cursor, but our policy is to only test models available on public routers for now.
Not at all. The model could (and sometimes should) burn all the money it wants, and then produce a single line of actual production code. Only some things, e.g. full rewrites, have clear cost - LoC scaling.
For my usage, I would very much prefer if those $/task were being spent in thinking and experimenting, and the actual output would be as short and maintainable as possible. “maintainability” is a vague target of course, but it’s at least somewhat correlated with code size.
Gemini models struggle with agentic coding/tool use/exploration, but they are actually quite smart in one-shot reasoning. They're not as far behind as people think. It's mostly post-training and productization issues, which are easier to fix than pre-training/mid-training issues.
It seems a weird and arbitrary challenge for a language model to be expected to perform. It also seems like there are some harness/visual issues even in the first few steps, where it states that it hasn't moved when it clearly has.
> Can someone explain how it in any way indicates how close we are to "AGI"?
I think it is historical name.
At some point when benchmarking was very undeveloped, this was targeting abstract reasoning and generalization, hence AGI.
the problems are general and abstract, domain specific knowledge and memorization don't help. Figuring out the rules, the goal, the controls, and how to solve in a reasonable budget all indicate some level of general ability.
So basically they're well suited for like, an octopus or a crow?
I was thinking about those species earlier in the context of, what does intelligence mean outside of language.
The benchmark appears to be testing the same thing. Although I don't know how much transfer there would be between this data set and the kind of situations a crow or an octopus would encounter.
Edit: Huh, it's just a Game boy game? I just did a couple of the tasks. It looks like C64 era game to me. Navigating levels. A lot of overlap with animal intelligence then.
I'm surprised it is that low. Are not all top AI labs "cheating" and workaround LLMs's low sample efficiency by hiring people to generate more data points - similar problems with answers, so they can train models on those and improve scores? A good benchmark for general intelligence probably should be a complete black box, no sample data given/leaked at all.
Oh it's because the bench is lying.
You need to pass each level without failing, if you fail a level, it count as you "lost" the minigame.
The fact it start to get a score means it managed to get a 100% score on one of the minigame.
Very interesting. My prediction is that Mythos would outperform Sol.
Also what does this tell about Yann LeCuns whole world model theory? Bro has been going on and on about it. He has made multiple wrong predictions on the trajectory of LLMs.
At some point his claim should be fully falsified no?
Are you joking? They spend billions of dollars training LLMs to get a 7.8% on arc agi 3 whereas DINO models are near sota in image classification, provide meaningful embeddings to the point where image segmentation is just PCA. The spend on DINO cannot be more than five million (correct me if I'm wrong)
His main anti-LLM predictions have been consistently either wrong or misleading.
There's many ways to skin a cat so you can probably do something with a JEPA approach as well, but I doubt he actually catches up to having agents on the level of where Anthropic/OpenAI will be at any point.
His main LLM predictions have almost nothing to do with Arc AGI...
What exactly was he dead wrong about that is proven by any of this?
GPT getting better has absolutely nothing to do with completely disproving anything LeCun has been saying.
He never said LLMs couldn't get better. He never said they couldn't score 7.6% on Arc AGI 3.
He's merely said they don't think, and you probably want something that actually thinks if you want a model that can be trained cheaply on a small amount of data and provide a ton of value.
Spending $5B to train a model that scores better than an older model does not disprove any of that in any way.
He said years ago even 'GPT 5000' couldnt do things that they ended up doing fine a month later, let alone by 5000. His later predictions are just moving that goal post including towards them not being able to do more general, harder problems of which Arc AGI is a counter-example.
That does not at all look cherry picked or taken out of context...
LeCun's ideas cannot be reduced to a 6 second clip...
You're missing the forrest for the trees, taking a singular example of a problem and thinking that if an LLM can solve the singular example it completely disproves LeCun is comical...
You can watch the whole Lex Friedman interview, it's on youtube. It's not out of context at all. He goes on about how LLMs will never be able to do things that they do trivially. And he has just doubled down for years.
Ive read and watched more of his interviews and lectures it seems, it feels like you just have a rosier idea of his views than the views he repeatedly presents.
I think you misunderstand what he thinks LLMs can and can't do. He says repeatedly that LLMs may be great for generating text and code, but that a fundamental model of artificial intelligence should be able to perceive and use information outside of text. Also, that a supervised learning approach is not exactly representative for how we learn. That it's a small piece but we largely learn with unsupervised learning. His main criticisms of LLMs are that they are supervised, probabilistic, and learn largely from text instead of observations. His claims about performance come downstream and I'd still argue that he's been (somewhat) right about those as well.
That LLMs don't have common sense and don't have good physical reasoning abilities; that you can't scale LLMs all the way to AGI; or that they can't predict the consequences of their actions which is the foundation of agentic behavior all seem like still (mostly) accurate predictions to me.
While LeCun has his share of problems, I think largely his criticisms on LLMs are more right than wrong. What remains to see is how good JEPA can be at filling in the gaps left behind from the brittleness of LLMS.
I think he argues more that LLMs are a form of supervision because you need a whole batch of text about specifically what you want the LLM to learn for it to be useful at that one topic. That they predict on human supplied data instead of learning purely from existing in the world and from observations.
Either way, there's something fundamentally inefficient about the whole business of training LLMs.
The human brain manages to self-organize with only a fraction of the information that LLMs get trained on. To train an LLM you need a lot of high-quality data.
There are two threads in history: firstly the compute thread leading to GPUs and AlexNet in 2012. Secondly the model architecture thread that started long before we had the compute and lead to transformers in 2017.
If the compute thread had been 30 years behind then we might be spending this century coming up with better architectures to make do with the more limited compute. However since the compute came first, we settled on the first thing that worked (transformers) and all effort went into polishing that.
There's something wrong with transformers though. No matter how many trillions of dollars get thrown at the problem, they still don't learn like humans do.
>The human brain manages to self-organize with only a fraction of the information that LLMs get trained on.
So? The question isnt can we get to ASI as efficiently as a brain, the question is can we get there, which we likely can. The inefficiencies can also be fixed after that.
>No matter how many trillions of dollars get thrown at the problem, they still don't learn like humans do.
Again, so? Humans are efficient but also bad at many things that transformers are already better at because of it. You are looking at the wrong thing if you think it needs to be like humans.
To me LLMs have gotten better since 2024, but their fundamental flaws still seem there.
They hallucinate when it comes to really challenging tasks such as math proofs. They still do not reuse code well and will rewrite functions instead of perusing the standard library.
But this is good news. LLMs are awesome and they are only the first step towards AI being applied everywhere. They are a Model T
> What exactly was he dead wrong about that is proven by any of this?
He said as you need more and more tokens models will fall apart because each additional token is a chance for a mistake and they will just exponentially fall apart. But in practice models have learned to identify and self-correct mistakes and if you look at the graphs more inference reasoning tokens almost always give far better accuracy.
That's true, but he's still correct, it's just that the context is now so large that only people using agent loops see "context rot"
His other criticism of LLMs that I like better is that they try to predict tokens instead of learned embeddings. Tokens are arbitrary and in order to decode LLMs you need technical analysis (see mechanistic interpretability).
With JEPA models so far, it seems that PCA on latent vectors suffices.
tldr: embeddings have a lot more room for improvement
One of the big things that enabled extended context to work was training techniques for extending LLMs with reasoning, part of the thing he was saying wouldn't work.
From CoreWeave, at current prices (~$2.46/hr spot to ~$6.16/hr on demand) would correspond to $22M–$55M.
The dataset is really where the cost is though - they used LVD-1689M - 1.6B images of curated web data from roughly 17B instagram images. This probably cost a huge amount of hours in human annotation, compute for algorithmic filtering, etc and not to mention probably a 20-50 person team working on this model.
You might want to change assumptions about how expensive these models are.
Thanks for the correction on the order of magnitude for the whole training process.
The 9M GPU hours includes the DINO v2 inference used in order to curate the data set.
The final training run used like 300000 dollars of compute.
Unfortunately we don't know how much RLVR + Agent training costs these companies. I'm just gonna say it's in the hundreds of millions, because they are supposedly making billions of profit on inference yet making billion dollar losses
DINO was created independent of JEPA but uses a similar principle of self supervised learning through minimizing the prediction error of a latent.
The difficulty in predicting a latent is so called "collapse"; the embedding neutral network can always output the zero vector and this would predict the output correctly.
There are different ways to solve this, DINO uses two different models - a teacher and a student and LeCunn uses an explicit term against collapsing to a single output.
My main takeaway from LeCun's thesis isn't that you can't build LLMs to do useful things better than the best human, it's that these systems don't learn arbitrary skills efficiently, like humans do. And the question is, why not? 8% on ARC-AGI-3 is amazing for a machine considering how far we've come since digital computers were first built. But it is pretty poor if you're claiming something is well on its way to exhibiting human-like intelligence.
Mythos can do some amazing things (I'm assuming, I've never seen it). A young child can learn to control its body without reading any books on dynamical systems and kinematics. Mythos cannot learn to control a humanoid robot after sucking in every piece of data Anthropic can get their hands on.
Sol is the first verified frontier model to ever beat an ARC-AGI-3 game
https://arcprize.org/results/openai-gpt-5-6