Today's models depend on inference time compute to get these results. The inference time compute available on any claude subscription is not comparable to the ones used to get some of these results (yes, in this case, it is 2000 USD total as noam confirmed, but some previous results took more).
In general, you can think of the process as generating massive rollouts in generation N, and then compiling in the verifier/human feedback("gradient") signal into generation N+1. The time taken to make the rollout in generation N, and separately the time taken to get the same rollout in generation N+1, each grows constant in some tasks, linear in more, and exponential in some.
In the end, this becomes bottlenecked by time. Today, we can make statements like "I generated all these successful trajectories with 2 weeks of compute, in the next model it will be able to do it in 7 hours of compute", but very soon you'll find yourself making statements like "I generated.... with 8 months of compute, in the next model it can do it in 6 months", which isn't really enticing the same way you can _technically_ brute force passwords but it just needs prohibitive amounts of time and money. That is the "plateau". Note that, this point is quite far away. For example, at any point if we agree it plateaus, today's known hardware techniques such as fixed function accelerators give you a 10-100x timeline reduction immediately allowing for a few more cycles of improvement. This is not to mention future innovations, but of course none of that is helping with the benchmarks where the time needed is growing superlinearly.
In many math and coding benchmarks, we are still in the constant phase. These are the massive improvements we see every few months. I'm not making any prediction of what will plateau and what will not as it's not possible to make an informed prediction about these things IMO. But the observed fact is that some have already plateaud as in, they don't improve with reasonable inference time (likely superlinear growth).
> will we need mathematicians to translate
Let's take a sudoku analogy. The model is initially just doing the random value algorithm, but lets say you the human are watching it. You make one of the usual reductions and interject "hey you can stop trying 8 here because of ....". Over enough examples, you get to a point where the model is _forced_ to learn the logical pattern. Next generation, it will skip that number. After this, you can peak the distribution using simple 1/0 RL. Doing _pure_ 1/0 RL works decent, but its not frontier as its a very sparse signal.
For that lift, human (or even a better LLM, but if you're trying to improve a frontier LLM, there is by definition no better LLM) feedback becomes necessary. This is _why_ it is crucial that these models interface in natural language and is also why the labs are hiring AI tutors by the hundreds. The "better LLM" case is what Kimi etc are doing by "distilling"(bad term for this) claude.
> But the long term is completely bewildering if you believe any of these trends can continue at a similar pace for the next few years.
For math and coding, for now we are in the phase where the times are just ... constant, so there's little reason to think it will stop soon. We still need humans to expand the frontier. It just becomes a matter of if its worth the cost of compute for running this generalized The Algorithm or not.
Given how well chess players internalized _many_ (not all) of alphazero's emergent chess knowledge, I am confident we wont have too much trouble figuring out any new math LLMs come up with, which will let us keep expanding the frontier by giving the LLM the next "lift". Only when we reach the stage where the time growth become exponential will this stop, IMO.