A reading list, not a take

Nathan Lambert pulled together his research materials while prepping public and policy writing on open models. It is the best writing of the last few years, in one place.

“This is my list of the best writing on open models in the last few years.”

Start with the business case

Bill Gurley's piece traces how open-source software became business strategy, and what the early signals say about AI following the same path.

Why Meta gave Llama away

Zuckerberg's post around the Llama 3 release is still the clearest statement of why a company would hand out its models on purpose.

“Open Source AI is the Path Forward”

Open is a dial, not a switch

Irene Solaiman argues release should be read on a gradient: license terms, cost to run, data access. Binary open versus closed hides most of what matters.

Open models as complement

The bet is that open weights power custom agentic workflows inside enterprises while strong closed models sit alongside them. Catalini makes the economic version of the same case.

Perpetual catch-up

Open models will keep trailing closed ones on raw performance. But adoption follows a different curve, so the two are on separate exponentials.

“Open and closed models are on different exponentials”

Safety without the panic

Thinking Machines Lab lays out how to ship powerful open weights responsibly. Kapoor and Bommasani found text LLMs added only marginal documented risk.

The real risk is already here

Florian Brand's point: guardrails on closed models get bypassed constantly, causing actual harm, while the hypothetical open-weight catastrophes have yet to show up.

“The Myth of unsafe Open Source AI”

The data commons is closing

Longpre's Consent in Crisis documents the rapid collapse of openly available training data, which has quietly hobbled genuinely open AI research.

America needs to build

The ATOM Project argues the US must invest in open models for fundamental R&D. Lambert and Kevin Xu add that banning them would cut against education, innovation and competition.

“Banning Open Source AI Would Be A Mistake”

Six months to live

Vague federal oversight, what Lambert calls vibe regulation, sets up a collision with frontier open models and possibly a ban in the near term.

“6 months to live for open models”

China's structural advantage

Kevin Xu traces Chinese open source back before AI and explains the structural edge. Lambert's notes from inside Chinese labs show how differently the industry operates.

“Get it out fast. We open source it within a few hours.”

Washington noticed

Lawmakers have probed DoorDash, Airbnb, Cursor and Apple over using Chinese models. Perplexity adopted DeepSeek R1 within days; Thomson Reuters moved off Claude to Qwen.

The gap is four to six months

SemiAnalysis, Epoch AI and Artificial Analysis all point the same way. Every leading open model since roughly 2024 has come from a Chinese lab.

Cheap enough to matter

Open models sit on the Pareto cost frontier even when they miss the absolute performance frontier. DeepSeek V4 Flash is the current example.

You cannot ban access

Joshua Saxe argues bad actors will always reach cyber-capable models, so policy should observe and respond downstream. Helen Toner: nonproliferation is the wrong frame entirely.

“Nonproliferation is the wrong approach to AI misuse”

The distillation panic

Training on another model's outputs is 2026's loudest fight. Anthropic documented at-scale misuse, and a paper showed reasoning traces could be extracted from proprietary APIs.

“Stealing Reasoning Traces from Proprietary LLM APIs”