Open Source

The dirty secret of the open-source debate at the frontier labs is that it was settled years ago, quietly, by the people who once championed it most loudly. OpenAI's founding bet, in Ilya Sutskever's own telling, was that "the best way to do it is by just open sourcing a lot of Technology" (No Priors, Nov 2023). Nobody open-sources their frontier weights now. What's left is a much more interesting argument about where the line sits, what you can give away without giving away the keys, and whether "open" was ever about weights at all... or about who gets to use the thing once it's built.

The original sin, recanted

Sutskever is the cleanest witness here because he watched the conversion happen from the inside. "The goal of open AI from the very beginning has been to make sure that artificial general intelligence... benefits all of humanity," he says. "The initial thinking has been that maybe the best way to do it is by just open sourcing a lot of Technology" (No Priors, Nov 2023). What changed wasn't the goal. It was the realization that "to make progress in AI for real you need a lot of compute," and a nonprofit giving everything away couldn't raise billions for clusters. The name "OpenAI" is now a fossil of a strategy its own co-founder abandoned.

Greg Brockman tells the same story as a tale of corporate plumbing rather than principle. They tried to raise $100 million as a nonprofit ("that's a staggering amount for a non-profit," one investor told him) and actually pulled it off, but "10x that... was not going to happen" (SXSW, Mar 2023). So they invented the capped-profit limited partnership, "super weird... all custom docs." Note what Brockman defends and what he doesn't: he defends distributing the benefits of AGI ("universal basic income," "lowcost doctors," "AGI services that Empower every individual"). He never once defends shipping the weights. "Open" quietly migrated from open weights to open access.

The line nobody can locate but everybody believes in

Here is the actual consensus, and it's a strange one. Almost every leader endorses openness up to a capability threshold, and refuses it past one... but none of them can tell you where the threshold is.

Sutskever draws it most explicitly: "up to a certain capability it's great." Then it gets "more complicated." His thought experiment: today's open models are fine, but "what about if you had a model which could autonomously start and build a large tech company," or "deliver on big science projects"? At that point "the argument there is a lot less clearcut a lot less straightforward" (AI TECH, Jul 2024). Asked for a signal that we've crossed the line, he has none. It's a threshold defined entirely by vibes and dread.

Demis Hassabis comes at it from the opposite emotional register but lands adjacent. "This is why science works, open science, is because... if there's a full sharing of everything, then progress overall is faster" (Semafor, Jan 2026). And then, in the same breath: "it may be a good thing that it's not as fast as that because there's safety to worry about." He wants the credit for openness (Google "putting transformers out there... AlphaGo we published that in nature") and harbors the now-standard worry that maybe faster isn't better. He also airs the free-rider grievance that has done more to kill open publication than any safety argument: "a lot of other people were using our stuff we put out in the open but we're not necessarily contributing... back."

So the unifying belief is: openness is good for science and bad for whatever comes after science. Nobody will say what comes after science.

What you can give away: tools, not weights

The frontier's compromise is to open-source everything except the thing that matters. Anthropic is the purest case. Dario Amodei's lab open-sourced its circuit-tracing interpretability library in May 2025, explicitly built to run on "popular open-weights models" like Gemma-2-2b and Llama-3.2-1b (Anthropic Research, May 2025). Read that carefully: Anthropic ships the microscope and lets you point it at other people's open models. Its own Claude weights stay locked. Open tools, closed weights... interpretability for the commons, capability for the vault.

Mira Murati's Thinking Machines is the loudest pushback, and tellingly it builds on top of the open-weight ecosystem rather than contributing frontier weights to it. Tinker fine-tunes "two open source models: Meta's Llama and Alibaba's Qwen" (WIRED, Jan 2025). Murati's pitch is genuinely democratic: "We're making what is otherwise a frontier capability accessible to all, and that is completely game-changing... we need as many smart people as possible to do frontier AI research." Her mission statement laments that "knowledge of how these systems are trained is concentrated within the top research Labs" (Feb 2025). But notice the move: she's democratizing the customization of open weights others released, not the frontier weights themselves. Even the leading open-source partisan operates downstream of Meta and Alibaba.

Jack Clark's ecology: rats, fruit flies, and superpredators

Nobody maps the actual landscape better than Anthropic's Jack Clark, who has stopped arguing about open source as a value and started observing it as an ecosystem. Proprietary frontier models are "superpredators... akin to lumbering elephants or whales... of supreme capability, but also in some sense slow-moving." Open-weight models are "rats and fruit flies and ants and cockroaches... fast-moving creatures that disseminate themselves through natural and artificial environments with amazing speed" (Import AI 435, Nov 2025).

And the rats are catching up. Citing the "Intelligence per Watt" paper, Clark notes local open models can now answer 88.7% of single-turn queries, up from 23% in 2023 and 48.7% in 2024, with accuracy-per-watt improving 5.3x in two years... GPT-OSS now runs on an Apple M4 Max. The capability gap is closing on retail use even as proprietary models "hold an edge" on hard reasoning (a B200 still gets 1.40x the intelligence-per-watt of an M4 Max). But Clark also documents openness's failure mode mercilessly: the Swiss Apertus models, trained openly on 4096 GH200s, "are not good," scoring 69.6 on MMLU against Llama-3.3-70B's 87.5. His verdict: "We're well past the stage where it is notable for non-corporates to train non-trivial LLMs... what matters is performance." Openness as virtue is dead; openness as competitive product is the only game.

Where "open" went: abundance, not weights

Watch Sam Altman in 2026 and you'll notice he barely discusses open-weighting OpenAI's models at all. The word "democratize" recurs, but it's been redefined as cheapness. "We need to make AI so inexpensive, so abundant, so good, such that people can use a ton of it" (News18, Feb 2026). India, he says approvingly, is building "the entire AI stack... from data centers to models to applications." Demand is "very correlated to price": cheap intelligence creates "huge amounts of demand." This is the abundance gambit dressed as access. You don't get the weights; you get oceans of cheap inference. Altman's genuine excitement is reserved for OpenCLAW and local agents ("the most exciting thing to happen in the AI space in quite some time"), where the "open" is in the agentic interoperability, not the model.

The economic stakes are why the threshold question won't stay theoretical. Clark, sign in hand in imaginary Washington, points at OpenAI's GDPval benchmark testing models across 44 occupations, where frontier models already hit a 47.6% win-or-tie rate against human experts "roughly 100x faster and 100x cheaper" (Import AI 429, Sep 2025). Altman concedes the displacement directly: there's "some AI washing" but also "some real displacement," and he expects "more of a ladder over time" (News18, Feb 2026). If models this economically potent become trivially copyable, "open source" stops being a licensing debate and becomes a question about who controls the means of production. That, not safety theater, is why the weights stay locked.

The tell is in the name. A company called OpenAI keeps its weights shut, while a company called Thinking Machines builds its whole identity on opening someone else's. Everyone agrees openness is good... right up to the capability where it isn't, a point each of them swears is coming and none of them can find.