Anthropic's cyber, bio, and distillationTraining a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.Full definition → filters reroute the request to Opus 4.8 and tell the user. The frontier-AI-development safeguard works differently: no fallback, no notice, just prompt modification, steering vectors, or small fine-tunes quietly degrading the answer.
The inconsistency is what does the damage. One visible policy applied across every domain would be arguable; a single invisible carve-out casts doubt on the whole safety program and reads as protecting a competitive lead rather than users.
Published benchmarkA standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.Full definition → scores carry an asterisk once filters exist: some prompts get silently downgraded to a weaker model, so the number the lab reports is not the number the public actually gets.
The author's hypothesis on why labs can't simply block distillation: reasoning models deeply want to emit their traces, and patching that out would make the model meaningfully dumber. The leak is a property of the capability, not a bug to fix.
The model shipped at roughly 2x the price of the then-current Opus with no single named breakthrough behind it, no inferenceRunning a trained model to get answers — the phase where AI is actually used, as opposed to trained.Full definition →-time-scaling moment, just gains across the whole stack, and it sat finished for 2+ months before release.
Terms in this piece · Glossary
distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
Why it matters
If you build on frontier models, you need to know when the vendor silently alters model behavior for certain query classes — this piece documents exactly that and how Anthropic responded.
Key quotes
“An AI model that gets less intelligent automatically without notifying me is categorically misaligned AI.”
“This is a mix of transparent and reasonable safety policies with quietly rolled-out market entrenchment tactics.”
“Building these safeguards is not something that Anthropic should do alone.”
“Safety research should be built on common understanding and information sharing across both labs and public research efforts.”
“We need intelligence that we can trust, that we can modify, and that we can control.”
