The Wikipedia argument

In a February 2024 interview with Time, Yann LeCun made the open-weights argument about as compactly as it has ever been made.

“It’s as if you said,” he told the magazine, “can you have a commercial entity, somewhere on the West Coast of the U.S., produce Wikipedia? No. Wikipedia is crowdsourced because it works.” A few sentences later he drew the conclusion: “the future has to be open source.”

LeCun, the 2018 Turing Award co-laureate who founded Meta’s AI research lab in 2013, directed it until 2018, remained the company’s chief AI scientist until the end of 2025 and has since served as executive chairman of AMI Labs, has been making some version of this case in public since at least 2023. The line is a prediction about where the rents in a maturing technology stack end up.

The analogy survives or fails on a narrower question than the line itself admits: what about Wikipedia actually made it win.

The marginal cost of contribution was approximately zero, which meant the supply curve of editors was effectively limitless. The collective oversight loop made the content correct enough for the median use case. And the closed alternatives, Encarta and the printed Britannica chief among them, priced themselves into irrelevance because their unit economics demanded a price the user no longer had to pay. By 2009 Encarta was dead. By 2012 the Britannica had stopped its print run. The losing side did not lose because it produced an inferior product article-for-article. It lost because the cost structure underneath its product became indefensible against a thing that was free, good enough, and improving faster.

Those conditions do not hold for foundation models, at least not in the way they held for an encyclopedia. A frontier training run is not a marginal-cost-zero contribution. It is, by the public estimates of the last two years, a hundred-million-dollar capital expense at the bottom of the range, with a comparable order of magnitude in the talent and infrastructure required to absorb it. The contributor pool that built Wikipedia did not need a GPU cluster. The Wikipedia analogy gets the direction of travel right, toward commodification of the underlying artifact and value migration upward, but it picks the wrong precedent for the mechanism.

LeCun spent twelve years inside Meta, which has funded the open-weights bet since 2023.

Meta released Llama 1 in February 2023 under a research-only license; the weights leaked within weeks. Llama 2, in July 2023, came with open weights and a commercial use grant. Through 2024 the family iterated quickly. Llama 4 followed in 2025. Across that arc Meta did something few others in the top tier of compute have matched at this scale or cadence: it has placed the model weights of a near-frontier system into the hands of anyone with the hardware to run them, and it has done so on a release cadence that has materially compressed the gap between what a developer can self-host and what the closed labs are selling through APIs. OpenAI’s August 2025 release of the Apache-licensed gpt-oss-120b and gpt-oss-20b, with the larger model near o4-mini on core reasoning benchmarks by OpenAI’s account, put the maker of ChatGPT in an open-weight field that already included Google DeepMind’s Gemma models and Chinese labs such as DeepSeek.

The gap, however, is real. Llama at its release moments has typically trailed the closed frontier (GPT-4-class models from OpenAI, Claude from Anthropic, Gemini from Google DeepMind) on raw capability benchmarks, and by a wider margin on the agentic features the labs have shifted their commercial pitch toward over the last year. How long the lag runs depends on which models, benchmarks and release dates are paired. Capability lag is the easier of the two gaps to argue away. The agentic gap is harder. It is partly a function of training data the closed labs have gathered from their own deployed traffic and partly a function of post-training infrastructure which is not weight-shaped and therefore does not transfer when the weights do.

What the Llama strategy has demonstrated is that a single well-capitalized actor can keep a credible open-weight track running adjacent to the closed frontier. What it has not yet demonstrated is that the open track will close the gap and stay closed. That is the empirical question on which LeCun’s argument rises or falls.

DeepSeek’s V3 release in December 2024 and R1 in January 2025 produced models with capability roughly comparable to the closed frontier of the moment. DeepSeek put the official training run of V3 at 2.788 million H800 GPU-hours, about $5.6 million at an assumed rental price of $2 an hour, excluding earlier research and ablation experiments. That figure covers V3 alone, not R1 and not the cost of serving either model, and the American labs have published no training costs on the same basis, so the comparison with their spending is loose. The releases were open-weight. They were also Chinese, which complicates the geopolitics in ways the Wikipedia analogy does not anticipate. But on the narrow technical question of whether a frontier-comparable model can be trained on a compute budget of that reported size and released openly, DeepSeek is the existence proof. Its V4-Flash, released in April 2026, listed uncached input tokens at $0.14 per million, roughly a thirty-sixth of GPT-5.5’s standard rate; the comparison is of list prices, not capability. It does suggest the moat the closed labs have been describing is thinner than their spending implied.

LeCun is not claiming that the open releases will outperform the closed frontier head-to-head on any given benchmark in any given quarter. He is claiming that the value in the AI stack will end up in the application layer rather than in the weights, that the weights will commodify on a timeline measurable in years rather than decades, and that the labs charging premium API rates today are charging against an asset whose half-life is shorter than their valuations imply. The actual mechanism is closer to what happened to commercial Unix in the 1990s, or to the proprietary database layer over the 2000s: an open alternative reaches good-enough on a cost curve the proprietary version cannot match, and the value migrates to the layer above.

It also explains why the argument is cheap for LeCun to make and expensive for the people running the closed-weight labs to refute. LeCun sits outside the value chain in question. AMI Labs, his post-Meta venture, is pursuing world-model architectures and is not, on any current public read, planning to compete on closed-weight foundation models for chat and coding. He has nothing to defend on the proprietary side. Sam Altman and Dario Amodei run OpenAI and Anthropic, and Demis Hassabis chairs Google DeepMind; the current revenue lines and forward valuations behind all three labs depend on the answer to LeCun’s question being no.

The other place the asymmetry shows up is on safety. The standard framing in 2024 and 2025 treated open weights as a misuse vector. Closed weights, by the same argument, were a containment mechanism.

LeCun has been inverting this argument since 2023.

In an October 29, 2023 post on X, reported by Business Insider, he warned that if open-source AI is “regulated out of existence,” then “a small number of companies from the West Coast of the US and China will control AI platform and hence control people’s entire digital diet.”

If foundation models become the substrate through which most people access most information, then a market with three or four serving incumbents is a market in which three or four firms decide what the median person can be told. Concentration, in this reading, is the safety problem. Openness is the mitigation.

“I’m very much of the second category of thinking,” he told Time in the 2024 interview. “What needs to be done is for society in general, the good guys, to stay ahead by progressing. And then it’s my good AI against your bad AI.”

This is not a settled argument. A RAND red-team study published in January 2024 had teams plan a biological attack, some with a large language model of the day plus the internet and some with the internet alone, and found no statistically significant difference in the viability of their plans. It tested one kind of misuse with the models then available; it did not test the release of model weights as such. The finding has been cited in defense of open release, and contested by researchers who argue the relevant question is not the current generation but the next one.

In the October 29, 2023 post on X, LeCun also accused Altman, Hassabis and Amodei by name of seeking regulatory capture; the two charges rest on one view of concentration.

If the open-weight gap to the closed frontier widens, then LeCun’s analogy will look like it reached for the wrong precedent. If the gap narrows or holds, then the value migration LeCun is predicting will already be visible in the API pricing of the major labs.

LeCun’s commitment is on the record. He left Meta at the end of 2025 after twelve years; in remarks at Davos in January 2026 he called the company’s culture “completely LLM-pilled” — his coinage for total institutional buy-in to the autoregressive-transformer track — a critique of architectural commitment rather than a retraction of the open-weights case. The architectural disagreement and the economic prediction are separable. He can be wrong about JEPA and right about Wikipedia. He can be right about both. The pricing on premium API access over the next eighteen months will settle it.