[jijijijij]

It doesn't matter where we draw the line, or what someone else said at some point. I merely pointed out your argument wasn't a sound rebuttal. Criticizing subjective experiences to assess the situation is, but that cuts both ways.

> HN commenters have been saying that LLMs plateaued ever since the first ChatGPT release.

Your earliest example was 2 years after release, when LLMs were already widely used and there is literally a source to support the claim. Now, you need to show research efforts, time and resource investments, ... are producing proportional results, disproving diminishing returns. If there are diminishing returns, LLMs are plateauing.

Also, manual capability extensions, or use/edge case adaptations, which may improve subjective usability are not exactly advances in AI as technology. LLMs still hallucinate. LLms still fundamentally struggle with certain classes of problems (e.g. counting), but you increasingly need to come up with different problem dress-ups because of targeted interventions to manufacture hype and a limited supply of test cases. Can you make the case AI got actually more intelligent, fundamentally? That is, not an increase in case specific usability, but a decrease in fundamental limitations. And is this proportional to improvement efforts?

[usef-]

> but you increasingly need to come up with different problem dress-ups because of targeted interventions to manufacture hype and a limited supply of test cases. Can you make the case AI got actually more intelligent, fundamentally? That is, not an increase in case specific usability, but a decrease in fundamental limitations. And is this proportional to improvement efforts?

I think you're setting up expectations here that nothing could falsify.

Any advance could be dismissed as "case-specific", couldn't it?

What's something concrete that you would judge as a fundamental step forward?

My own claim would be that the increase in long-horizon tasks without heavy harnesses is a sign of a decrease in "fundamental limitations".

METR's time-horizon tests are going up, and that seems to match actual problems LLMs are solving in daily usage. Eg. Agent workloads. Or games: see the newest models being able to beat longer-term tasks like Pokemon from just a minimal vision-only harness. Models last year had cheating harnesses and yet still failed.

> I merely pointed out your argument wasn't a sound rebuttal. Criticizing subjective experiences to assess the situation is, but that cuts both ways.

My original "doesn't mean progress stopped" comment was replying to the "no improvement" guy, not the plateau guy.

Yes, I agree there's vibe-based judgement happening on both sides. I was quoting old HN not to make a "boy who cried wolf" argument, but to say that feeling a vibe of slow progress is not enough, since people repeatedly have expressed that vibe.

> Your earliest example was 2 years after release, when LLMs were already widely used

Fair. I was going backwards and thought a few would suffice. There was a lot of doubt earlier too if you search. From March 2023, four months after chatgpt "We are heading towards a limit and the AIs aren't getting much better." https://news.ycombinator.com/item?id=35332537

> Now, _you_ need to show research efforts, time and resource investments, ... are producing _proportional_ results, disproving diminishing returns. If there are diminishing returns, LLMs are plateauing.

Yes, it's a good point, though we don't have good numbers on investment so it's hard to answer in either direction. But the METR speedup in 2024-25 appears to be mainly from RL/algorithm gains not training compute, judging as an outsider.

"Is progress efficiency per dollar decreasing" is a different claim to the earlier "is progress slowing?", though.

> LLMs still hallucinate. LLms still fundamentally struggle with certain classes of problems (e.g. counting),

Measured hallucination rates have gone down, and don't seem to be an issue for achieving things, as per METR and other results. I don't see hallucination issues in practice as all real work I know about includes the ability to validate.

If by "counting" you mean the "counting r's" issue, that's a tokenisation artifact (the models are blind to individual characters): ask it to count in Python and it's fine (ie. reasoning is not the problem).

(the viral online counting videos of voice-mode are using an ancient gpt4o-based model that is trained to keep answers short. I do think this is giving much of the public the vibe that current text models are worse than they actually are)

[jijijijij]

> Any advance could be dismissed as "case-specific", couldn't it?

and

> If by "counting" you mean the "counting r's" issue, that's a tokenisation artifact (the models are blind to individual characters): ask it to count in Python and it's fine (ie. reasoning is not the problem).

The problem is a lack of transparency. That's on the AI companies, not me. Everybody can attach a mathematics engine to an LLM and say "Look, it does math now!". The intelligence part hasn't changed though, the model didn't figure out mathematics, didn't get any closer to AGI.

I am also not saying, testing is impossible (making claims unfalsifiable), but you have to acknowledge it's inherently becoming harder to argue either way. Now, if I asked to count to 100 the LLMs deliver. I can even ask to count to 100 in base16. But it falls apart in less common bases like base33. The problem clearly isn't solved fundamentally. For counting in arbitrary bases you need a knowledge transfer and creative adjustment since you may exceed the symbol space. I think, this is also a good indicator because for the same reason there is no training material. Again, this can be addressed algorithmically, so this test is not futureproof.

Regarding character counting, the excuse about tokenisation is a bit questionable in this context, I think, because inherent limitations to the current approach is exactly the matter debated. Do we know, the extent of this particular problem space in the real world?

>What's something concrete that you would judge as a fundamental step forward?

I want reproducibility and transparency how a feature happened. If you threw a bazillion dollars onto Wolfram Alpha 15 years ago, you would also have a very useful tool for most questions. Spending a lot of money to implement edge case solution is not technological progress. That's always been possible. I mean, in the beginning there were thousands of humans curating chatgpt answers. That's not intelligence, just reckless spending, exploitation and borderline fraud. And it shows exactly how much "customer experience" can be delivered just by throwing money around.

> "Is progress efficiency per dollar decreasing" is a different claim to the earlier "is progress slowing?", though.

Well, I guess it is a matter of opinion. In my opinion, a "thermodynamic" perspective is useful. If you hire 100000 more workers to harvest a small field, you get it done quicker, but is it a breakthrough in farming? That is, trivially, if we spent more energy, resources, money we can expect to get more anything. However, those investments are not spent elsewhere, the progress of technological/scientific advancement is slowing down elsewhere. So, going by relative contributions to overall progress things are slowing down when progress is disproportionally low in an resource intense field. After all resources on Earth are limited and there is only so much sunlight refueling our energy supply. If the AI hype isn't panning out, the damage by hardware prices alone will be enormous. You need compute for more than AI and a lot of that is unaffordable right now. AI has to be 10X for every field, or it's gonna be a net global set back. Look at the volatility and social disruption in the US. It's a gamble, not covered by growth, but debt. Putting progression into relation to effort is required IMO.

> Measured hallucination rates have gone down, and don't seem to be an issue for achieving things, as per METR and other results. I don't see hallucination issues in practice as all real work I know about includes the ability to validate.

Are you in tech? Because I think there is a bias related to the above, since agentic AI can compile/run code to validate answers and most problems are documented ad nauseam. In my experience, if you search for very specific information in other sciences, you still frequently get confident non-sense. Outside of IT, you can't really have a "fact checking" co-routine, you need actual cognition. Of course, if you "chain AIs" things get a little better, but I claim again that's multiplying effort, not fundamentally making anything smarter. If we compare to biological intelligence, we absolutely do not see the same relationship. You may think this argument is moving the goal post, but if I ask you this: If we double the AI workload (eg. agentic AI) to get 10% better results, doesn't this sound like approaching a plateau to you? We are literally putting the fabric of civilization and humanity's future at risk for this (by social disruption, increasing energy/resource expenditure and not spending the money elsewhere in most critical times). Expectations must be high.

Anyway, thank you for expanding the argument, nice discussion.