> We haven't seen a significant increase in the quality of LMM output since 2023 that hasn't been the result of throwing even more energy and compute at it.
This is completely false. Most of the dramatic improvements in LLM quality in the last two years were due to the application of new post-training methods, especially RLVR. It’s really interesting to read about (you should!) and it is the whole secret to why LLMs did not plateau in 2024 or 2025 like many people confidently predicted. Sure, RLVR requires compute to do, but this is not just throwing more compute at 2023 LLMs.
That's interesting. If you have a source that shows that RLVR was primarily responsible for model improvement, I'd be interested to see it. In any case, it sounds like it has its own set of limitations and there are applications where it does not help at all.
This is completely false. Most of the dramatic improvements in LLM quality in the last two years were due to the application of new post-training methods, especially RLVR. It’s really interesting to read about (you should!) and it is the whole secret to why LLMs did not plateau in 2024 or 2025 like many people confidently predicted. Sure, RLVR requires compute to do, but this is not just throwing more compute at 2023 LLMs.