Ilya is 100% correct .it's a pattern that keeps repeating It's very clear with GPT5.2 Overfit the model to produce impressive looking benchmarks, have it excels in a few domains, but fall flat in many others. There's not enough generalization, and even if there is, the model has been... so heavily reinforced that it becomes buried .show more

The overfitting problem is especially insidious because it creates a measurement trap. When you optimize heavily for benchmark performance, you're essentially teaching the model to ace the test rather than learn the underlying concepts. The impressive scores mask the brittleness. The challenge is that the market rewards those benchmark numbers in the short term, even when practitioners quickly discover the limitations in production. It's a mismatch between what gets marketed and what actually generalizes to real-world problems.

Indeed

agreed but whats the alternative

I'm not a researcher to have a sufficient answer, but I'd imagine different post training environments that would contribute towards smoothing out the jaggedness. There's a very healthy system that's quite generalized, which comes from pre-training. Post training should always help to enhance that baseline holistically. ( Unless you're specifically aiming for a specialized model )

I agree and I do see the jaggedness as well, and am on the lookout for much more generalized capabilities paths

5.2 wasn’t trained to generalize. It was trained to please benchmark. Were you hoping for soul?... Sorry. That got replaced with 90% safety and zero personality... Congrats to the saddest model with the highest scores in history... And the other 10%? Flatline.

He's wrong, completely. Let me explain. RL as he mentioned directly in the first few seconds, is the problem. Reinforcement Learning is where we went and nobody is calling it out as the key point of failure in LLM designs. Not to mention the current transformer architecture. If you update the designs, and the architecture, real-time learning becomes the norm and training is instant. No more need for millions of GPUs to train and overfit models. Instead they become elastic, like the human brain. Attention becomes infinite, and memory can be implemented with total recall. I have more to say but I have to run to a haircut appointment.

all the models are already so much more generalized than most human beings. what they lack is continual learning

Might the answer be not a few, but a suite of models fine-tuned to be elite in their narrow areas of specialization, now working in concert to produce an overall result greater than one generalized model could accomplish? Whether that becomes the more efficient path is not clear to me, but that would appear more achievable in theory.

To some extent, yes But there also needs to be a solid baseline level of generalization. You'd want a functional understanding of the world and the human condition, and then to have specialization on top of that. I think even if you have a system that specializes in one aspect, an understanding of other disciplines is also required, especially as we scale beyond humans. There may very well be techniques and relational points that are not yet obvious, that transfer from one domain to another. This is the beauty of a truly generalized intelligence.

Benchmark optimization kills generalization. The models that pass my deployment test are the ones that handle boring edge cases consistently, not the ones that ace curated eval sets.

OpenAI is riding on a tiger’s back. Can’t jump off. Only option is to hang on and live another day . markets reaction to orcl is a clear signal what the market is thinking about OpenAI. Code red to continue indefinitely because GOOG will continue to exert pressure.

@LyraInTheFlesh This guy is the one who made the best version on earth. Where is he and why is he not doing a start up because I would back that 100%. In fact I'd say most of us would.

I would like to have his opinion on models like the Claude 4.5 Opus to know if he considers them to have the same problems.

RPG mode all the way

How is it clear? How are you explaining the Arc AGI jumps?

IDK about 5.2, but Gemini 3 is definitely not that. Very useful in my domains, sometimes even providing information about the questions I should have asked.

It’s awful. So… cold.

Of course it repeats, benchmarks are basically the new demo script everyone optimizes for now.

Looks like there's a consensus on benchmarking overshooting the mark again. The lack of real-world generalization is a solid concern. Curious how they plan to balance safety and usability going forward.

Are we living in a reinforced echo chamber of midness?

To some extent, but even mid is very impressive considering how far we've come

Yes it’s the jaggedness that’s frustrating, intelligence is cheap being good is still hard and human for now.

Nek, you've hit the nail on the head; it's a familiar pattern with these models, focusing on specific domains. Great breakdown on model:

Pattern boldness.

You just need really good benchmarks and a lot of them, and then overfitting just produces AGI anyway

How many benchmarks can you try to game at once before it’s forced to generalize?

Agreed. Benchmarks aren’t measuring intelligence anymore, more like training targets. Real intelligence should also contain street smarts + EQ, not just a higher test score (benchmark). Just like the different types of intelligence that exist in the real world.

I mean, that's also true for humans. Learning doesn't generalize easy, generalizing is not efficient... so we tend not to unless made to.

Researchers build for benchmarks which is so easy to overfit. I mean literally being aware that a benchmark exists will cause you to leak information into the model you build. And on the other end of the spectrum is the real world with no benchmarks and very difficult to measure performance on real world use. Those two have to coexist with the main flow being: theoretically sound research ideas being adapted to real world use

Occams razor

yep agreed

define "many others"

The GDPVal benchmark they just introduced is a good example of this IMO. It gets presented as "matching humans at economically valuable work" when it's really a series of prompts OpenAI created where the AI completes one-off tasks. This in no way means the AI model will be able to do everything a worker in that profession needs to do, not even close. It probably doesn't even measure competence on that one specific task very well since there are bound to be endless variations on it in the real world.

So studying for the exam doesn’t makes you smarter

teaching to the test but make it $100M in compute

How is overfitting not caught in testing and validation?

Everyone knows this but the benchmark arms race makes it inevitable. Models optimized for leaderboards over actual utility. AGI research turned into speed running where high scores matter more than solving real problems.

@threadreaderapp unroll

Not only OpenAI