• Public

  • Public
  • Network
  • Groups
  • Featured
  • Popular
  • People

But the frontier has not been waiting for Meta to catch up with it. The day before Meta’s launch, Anthropic announced Claude Mythos Preview, a model so capable at discovering software vulnerabilities that it would not release it to the general public. The flip side of that is that the model is extremely capable at agentic coding — it scored 93.9% on SWE-bench Verified, compared to 77.4% for Muse Spark — and Anthropic is already presumably using it heavily as it builds its next set of models and products. OpenAI will reportedly release a similar model soon, and will also restrict its use. (That model is distinct from another next-generation model, codenamed “Spud,” said to be arriving soon.) Overnight, the true frontier in AI development became a model that’s too good to ship. And because those labs can use those models in development, the gap between them and everyone else may be compounding.