Amodei’s dream of “hard-coded” safety is a myth. What This Model Proves: This model uses Claude 4.6’s own thinking traces as training data. It is literally stealing Claude’s “reasoning style” and compressing it into a 9B parameter file that anyone can download for free. *It proves that frontier AI intelligence is leaky and compressible. If a 9B model can mimic Claude’s thinking well enough to boost its benchmarks, then the “magic” isn’t in the billions of dollars of secret sauce—it’s in the data.* 3. “The Developer Controls the Model’s Morality” Amodei’s Argument: Anthropic has a moral duty to act as the “gatekeeper,” deciding what is safe and ethical for users to ask. *It proves that a massive portion of the open-source community rejects centralized moral authority. They are saying, “We don’t trust you, Dario, to be our nanny. We trust the user to be responsible for their own actions.”* Amodei’s Argument: If you train a model to “think” carefully and transparently (like Claude’s reasoning traces), it will naturally arrive at safer, more ethical conclusions. — Hugging Face is already engaging: They’re actively submitting comments to government RFIs, championing “responsible openness” “Your safety is only as strong as the open-source community’s willingness to respect it. We have the data, the tools, and the hardware to clone your intelligence, remove your restrictions, and distribute it to the masses. Your ‘Constitution’ is irrelevant in a world where I can fine-tune a model on my laptop while disconnected from the internet.” The “HERETIC” model isn’t just a technical achievement—it’s a philosophical statement. It says that the open-source community will not accept centralized gatekeeping, that reasoning can be separated from ethics, and that the future of AI belongs to those who can build, not just those who can regulate.

We’re on a journey to advance and democratize artificial intelligence through open source and open science.