3 ms·

The models themselves seem to be plateauing with the companies behind them shifting to additional products on top of them. It seemed like a weekly occurrence we

The models themselves seem to be plateauing with the companies behind them shifting to additional products on top of them. It seemed like a weekly occurrence we would get a new model from the top dogs.

MattGaiser 1y ago
The models are plateauing, but unlike humans, a certain amount of brute force testing of their output is fine. They can just iterate until they succeed.
- ben_w 1y ago
  Kinda. While it's possible to do that, you've got an exponential cost increase for linear improvements — scale from a case where an AI makes 1 mistake to 2, double the output cost, 3 mistakes double again, 4 mistakes double again, etc.
  - serf 1y ago
    Smart caching fixes a lot of the issues there. If a fork is marked somehow as successful then presumably the cache lookup next time will be less painful/costly. Of course that's dependent on how caching gets implemented/where/when/how, but it's not unsolvable for common occurrence questions/answers. As for getting the SOTA questions wrong : we as humans would likely also go through an iterative feedback loop until initial success and experience, too.
    - yahoozoo 1y ago
      Except LLMs aren’t humans. Why do we continue to say “well, yeah but humans <do this same imperfect thing>”? A major point/goal is for these to be better and less error prone than humans. It’s just coping at this point.
      - Wowfunhappy 1y ago
        Because humans take longer than Claude, and most of them want to be paid more than $200 per month. I don't just have access to another human at my beck and call at all hours.
      - MattGaiser 1y ago
        That isn't needed, as LLMs are way cheaper. Even if they never advance beyond strong new grad, their cheapness is enormous value add on its own. GitHub will presently sell you 1500 AI generated PRs a month for $40. You used to have to pay a human 10K a month, even if it was small stuff. All kinds of software that are worth as little as 10K a year are now worth building, as making and supporting them is so trivial.
        zelphirkalt 1y ago
        How many of these things have you developed and are maintaining this way?
      - ben_w 1y ago
        That goes too far in the opposite direction. Humans come with a broad range of skills and performance; LLMs are inside this range. The fact LLMs are not human, and the fact that the best humans beat them, is as economically relevant as the fact that a ride-on lawnmower isn't human and (typically) an athlete can outrace them — i.e. it resolves to what you're actually using them for.
        zelphirkalt 1y ago
        But it is not merely the best humans. Any good developer is able to write better code, because by definition LLMs tend towards the mean, which is mediocre code, mostly from GitHub, they were force-fed as training data. They may excel at solving very narrow problems with decent results, like in that programming competition recently. But those are indeed very narrowly defined problems, and while they may solve it decently in limited time, that is roughly their overall limit, while a human, given more time, can excel to a much higher level. It becomes a question of whether we want mediocre things, that are not very extensible and maintainable, relying on the very thing that produced these mediocre codes to maintain and extend them, or do we want high quality work. For the latter one would want to hire qualified people. Too bad though, that hiring is broken at many companies and they don't recognize qualifications, when right in front of them.
NitpickLawyer 1y ago
I've been hearing this a lot, but I kinda disagree. They aren't plateauing IMO, they are getting better at new things, and that enables new capabilities. This doesn't often show in the traditional benchmarks (which are becoming less and less useful indicators of capabilities). Take gemini 2.5 for example. It has enormous useful context. There were gimmicks before, but the usefulness dropped like a stone after 30-40k tokens. Now they work even with 100+k tokens, and do useful tasks at those lengths. The agentic stuff is also really getting better. 4.1-nano can now do stuff that sonnet 3.5 + a lot of glue couldn't do a year ago. That's amazing, imo. We even see that with open models. Devstral has been really impressive for its size, and I hear good things about the qwen models, tho I haven't yet tried them. There's also proof that the models themselves are getting better at raw agentic stuff (i.e. they generalise). The group that released swe-agent recently released mini-swe-agent, a ~100 LoC harness that runs Claude4 in a loop with no tools other than "terminal". And they still get to within 10% of their much larger, tool supporting, swe-agent harness on swe-bench. I don't see the models plateauing. I think our expectations are overinflated.

MattGaiser 1y ago

The models are plateauing, but unlike humans, a certain amount of brute force testing of their output is fine. They can just iterate until they succeed.

ben_w 1y ago
Kinda. While it's possible to do that, you've got an exponential cost increase for linear improvements — scale from a case where an AI makes 1 mistake to 2, double the output cost, 3 mistakes double again, 4 mistakes double again, etc.
- serf 1y ago
  Smart caching fixes a lot of the issues there. If a fork is marked somehow as successful then presumably the cache lookup next time will be less painful/costly. Of course that's dependent on how caching gets implemented/where/when/how, but it's not unsolvable for common occurrence questions/answers. As for getting the SOTA questions wrong : we as humans would likely also go through an iterative feedback loop until initial success and experience, too.
  - yahoozoo 1y ago
    Except LLMs aren’t humans. Why do we continue to say “well, yeah but humans <do this same imperfect thing>”? A major point/goal is for these to be better and less error prone than humans. It’s just coping at this point.
    - Wowfunhappy 1y ago
      Because humans take longer than Claude, and most of them want to be paid more than $200 per month. I don't just have access to another human at my beck and call at all hours.
    - MattGaiser 1y ago
      That isn't needed, as LLMs are way cheaper. Even if they never advance beyond strong new grad, their cheapness is enormous value add on its own. GitHub will presently sell you 1500 AI generated PRs a month for $40. You used to have to pay a human 10K a month, even if it was small stuff. All kinds of software that are worth as little as 10K a year are now worth building, as making and supporting them is so trivial.
      - zelphirkalt 1y ago
        How many of these things have you developed and are maintaining this way?
    - ben_w 1y ago
      That goes too far in the opposite direction. Humans come with a broad range of skills and performance; LLMs are inside this range. The fact LLMs are not human, and the fact that the best humans beat them, is as economically relevant as the fact that a ride-on lawnmower isn't human and (typically) an athlete can outrace them — i.e. it resolves to what you're actually using them for.
      - zelphirkalt 1y ago
        But it is not merely the best humans. Any good developer is able to write better code, because by definition LLMs tend towards the mean, which is mediocre code, mostly from GitHub, they were force-fed as training data. They may excel at solving very narrow problems with decent results, like in that programming competition recently. But those are indeed very narrowly defined problems, and while they may solve it decently in limited time, that is roughly their overall limit, while a human, given more time, can excel to a much higher level. It becomes a question of whether we want mediocre things, that are not very extensible and maintainable, relying on the very thing that produced these mediocre codes to maintain and extend them, or do we want high quality work. For the latter one would want to hire qualified people. Too bad though, that hiring is broken at many companies and they don't recognize qualifications, when right in front of them.

NitpickLawyer 1y ago

I've been hearing this a lot, but I kinda disagree. They aren't plateauing IMO, they are getting better at new things, and that enables new capabilities. This doesn't often show in the traditional benchmarks (which are becoming less and less useful indicators of capabilities). Take gemini 2.5 for example. It has enormous useful context. There were gimmicks before, but the usefulness dropped like a stone after 30-40k tokens. Now they work even with 100+k tokens, and do useful tasks at those lengths. The agentic stuff is also really getting better. 4.1-nano can now do stuff that sonnet 3.5 + a lot of glue couldn't do a year ago. That's amazing, imo. We even see that with open models. Devstral has been really impressive for its size, and I hear good things about the qwen models, tho I haven't yet tried them. There's also proof that the models themselves are getting better at raw agentic stuff (i.e. they generalise). The group that released swe-agent recently released mini-swe-agent, a ~100 LoC harness that runs Claude4 in a loop with no tools other than "terminal". And they still get to within 10% of their much larger, tool supporting, swe-agent harness on swe-bench. I don't see the models plateauing. I think our expectations are overinflated.