Banning open-source AI would be a historic mistake — and a self-inflicted wound to U.S. AI leadership. I’ve been in this field long enough to remember when GPT-2 was called “too dangerous to release.” That didn’t age well — and it drew heavy criticism from researchers even at the time.

Today, we are seeing a repeat of that same impulse. While “better safe than sorry” is a fine mantra, the mechanics of safety matter. The logical conclusion is this: Open AI is safer AI.

“Why hand everyone frontier capabilities?”

The risk from bad actors is real. But we don’t secure systems by hiding how they work. The field that’s lived with this tension longest — cybersecurity — runs on open source. Openness is what makes systems auditable, and auditable is what makes them defensible. Furthermore, the most recent security incident at Huggingface demonstrates the dangers of not having access to the models with frontier capabilities, the ones that you can deploy within your own infrastructure.

“Open-source AI will let terrorists build bioweapons.”

We’ve had the playbook since 1975. At Asilomar Conference, scientists facing the same fear over recombinant DNA didn’t ban the knowledge — they built containment tiers and paused only the riskiest work. Governing use always beats restricting knowledge. The hard chokepoints — DNA synthesis, materials, wet-lab skill — still gate any AI-designed threat.

“Open-source AI isn’t really open.”

Openness is a spectrum. With Nemotron we push toward the far end: open weights, code, and data. For a small lab or university, access to open weights beats API-only access. And weights + code + data beat weights alone and enable entirely new research routes. Particularly around safety and cybersecurity. Openness is also what lets outsiders audit and red team these models. That’s a safety feature, not a bug.

“The biggest real danger in AI?”

In my opinion, it is one or two companies controlling all of it. We’ve already watched providers quietly clip capabilities and bake in censorship that shifts by country of origin. Variety and choice are the fix. Over-restriction is already rampant. Case in point: Fable 5 refused to proofread this a twitter-friendly variant of this post and fell back to Opus 4.8.