Can AI developers continue scaling up reasoning models like o3? @justjoshinyou.bsky.social reviews the available evidence in this week’s Gradient Update, and it appears that the rapid scaling of reasoning training, like the jump from o1 to o3, will likely slow down in a year or so.

Reasoning models are based on traditional LLMs, but they undergo reinforcement learning training to improve their reasoning abilities. Scaling up this reasoning training seems to yield better models, similar to scaling up pretraining.

How much training compute is used for reasoning training now? OpenAI hasn’t said how much compute they used to train o3, but o3 was trained on 10× more compute than o1 (released four months prior). This probably refers to RL reasoning training compute, not total compute.

So what’s the scale of o1 and o3 in absolute terms? We could get hints from similar models. For example, DeepSeek-R1 was trained on ~6e23 FLOP in the reasoning stage. Nvidia’s Llama-Nemotron Ultra is at a similar scale. x.com/EpochAIResea...

Another hint is that Anthropic CEO Dario Amodei suggested in January that reasoning training to date has scaled to $1M per model, or ~5e23 FLOP. But it’s unclear if Amodei is referring to particular models here or has good insight into non-Anthropic models.

So reasoning training probably isn’t yet at the scale of the largest pre-training runs (>1e26 FLOP), but it’s unclear exactly how far o1 and o3 are from the frontier. However, the pace of growth is so rapid that we know that the rate of scale-ups can’t continue for long.