[solomatov]

>For coding you always want to go with the best model in the category, not something that would be the best model if we went 1 year back which GLM 5.1 is, and I'm saying that as a big fan of GLM cause I run a translation site where GLM is good enough for the price.

Currently, the difference is substantial, but what happens if capabilities saturate?

[aspenmartin]

Then the house of cards comes crumbling down, but there is so much evidence to point to this not happening that it requires a bit of a theory for how that may happen

[solomatov]

> but there is so much evidence to point to this not happening

Could you explain this?

[aspenmartin]

Well I think there are several fairly stable trends that paint a pretty compelling picture:

- performance scales with compute very very reliably. We have “scaling laws” (and have for years) and they are almost miraculously stable and show no sign of being invalidated at all even at the very largest scales. There are some theoretical bases for this though I’m not as familiar with the details

- these scaling laws are on an unintuitive quantity (validation loss on pretraining datasets), so we can look at downstream performance. Benchmarks are a minefield of junk but there are many decent ones and enough variety of techniques and data sources and scoring methods etc that in aggregate they are useful. The single number that I think is the best summary statistic across the crazy (O(100k)) number of benchmarks is the “epoch capability index” (just some branding over a reasonably standard statistical model that was really well thought out and a great idea). The trends in this are extremely stable. Eyeballing the trend over time on their graph we’re getting basically a GPT-4 to GPT-5 level capability improvement every ~18 months

- coding agents are not limited by the quality of the human training data they’re trained on, this is such a massive misconception: human data is only a bootstrap to a reinforcement learning phase. This combined with the fact that we have verifiable rewards means it’s just a matter of when not if for any given level of reliability.

- the massive compute investment implies that the compute that we’re building over the next 2-3 years will 10x the effective compute for training models. That combined with various R&D contributions (historically which have been very significant and there is no shortage of wins here), better data curation and flywheels, richer data (wait until conversation capability gets good) means we have several orders of magnitude of runway that we know of, today.

In short I don’t see any compelling evidence to suggest all of the trends we observe in many different ways will end any time soon.