AI Frontier

Written 8 days ago

From observing the AI frontier closely in the past few years, a few patterns are emerging on how models are improving:

  1. Your AI company must first produce a big, powerful, high compute-cost model like claude-mythos-5 (DeepSWE[1], CursorBench[2]). In some cases, like o1, a pioneering method like Chain of Thought is used too.
  2. You must then hype it up by finding X vulnerabilities in Y program[3] that millions use or a security incident[4]. Gatekeeping your latest model to researchers because it is "too dangerous" adds to this hype: release claude-fable-5 instead.
  3. Then, once the model releases, competitors will attempt to compete with your frontier: kimi-k3, qwen-3.8, grok-4.5. Some will beat it (gpt-5.6-sol). They could possibly distill your model against your ToS OR they're just slightly behind your frontier: it's suspicious so many models emerge above your previous flagship model's performance so fast right after your better one drops.
  4. Then, your own company will be able to distill it and improve smaller models to match the pioneer model's performance at a lower compute-cost (gpt-5.6-luna from gpt-5.6-sol, claude-opus-5 from claude-fable-5, o3-mini from o1).

Learning from this, I think you should throw a massive amount of compute at something at the beginning. Then you make it more efficient. This has been a long-standing law in engineering.

[1] https://deepswe.datacurve.ai
[2] https://cursor.com/cursorbench
[3] https://www.anthropic.com/research/glasswing-initial-update
[4] https://openai.com/index/hugging-face-model-evaluation-security-incident/