Why I build open language models

Nathan Lambert wrote this on October 30, 2024, his 30th birthday and the end of his first year at the Allen Institute for AI (Ai2). It is a personal essay rather than an argument aimed at policymakers, and half of it is a recruiting pitch. Lambert's open-source-ai-reading-list files it under the ATOM Project as the lens for why open models drive research and good outcomes. Reading it next to his 2025-2026 pieces is useful mostly for what it does not contain: China is not mentioned once.

Open without Meta

Lambert's complaint about the existing case for open source AI is that it feels like damage control, because its loudest advocate is a large technology company acting in its own interest. Meta releases Llama to commoditize its complements and weaken competitors, which Lambert calls the right strategy for Meta. He finds Zuckerberg's essay compelling, but says only its section on why open source is good for the world is worth reading closely, since the benefits to Meta and to developers are obvious.

His concern is that one for-profit company holds a majority share of what open source AI means in practice, while its licenses and messaging do not match open source principles. Whatever definition the field settles on "needs to work without Meta, which may stop releasing models at any time and use licenses that are broadly prohibitive." That turned out to be the right worry. By November 2025 the ATOM Project listed Meta under "unclear", with "crickets since the Llama 4 fiasco".

Outside Meta, he locates the real motive for open source AI in fear of AI development going wrong. The stories people tell are about repeating the mistakes of recommender systems and social networks: negligent actors rather than malicious ones, shipping powerful technology without oversight.

What openness is supposed to buy

Lambert's list of expected benefits touches most of the usual issues. Software gets more secure, because open-source vulnerabilities attract monitoring in proportion to their reach, where the closed world produces its own CrowdStrike outages; the xz-utils-incident is his example of an open-source backdoor that got caught. More stakeholders can work on specific subproblems of training, as in his own work with social choice researchers. More people can use AI, organizations can keep their data, and profits flow to product builders instead of platform owners like Apple, Google and Meta. Regulators get direct access to systems rather than access mediated by companies.

The more original point is about culture. AI models are relationships for many users, not just a channel for interaction, so the values behind a model matter more than they did for search or routing. Anthropic's culture is why Claude feels distinctive, but it also puts the company in the path of cultural fights as it scales. Lambert's claim is that open source "de-escalates culture": control is not democratized, since few organizations can afford to train models, but the decisions are visible and the world can adjust to them.

He also rejects the binary. A healthy open ecosystem still has many closed models, openness is a spectrum, and oversight does not depend on how a model is released. On risk he is brief, citing Kapoor, Bommasani and others (societal-impact-of-open-foundation-models) for the claim that the disinformation, biosecurity and phishing risks experts predicted from open models had been disproven, while granting that some model types should not be released.

The allies he names are also worth a note. He points to Dean Ball's writing at Hyperdimensional as one of the few good cases for open source AI. In July 2026, arguments-against-open-source-ai opens by quoting Dean Ball warning that an open-weight-dominant world risks "full AI communism". That later quote is the newer statement of Ball's position.

White rice research

The second half describes the job. Lambert calls it "white rice research": Ai2 releases solid models that do not quite beat Meta's Llamas, makes them easy to build on, and documents every step so other groups can pick up unfinished ideas. Maintaining large data repositories on Hugging Face is boring work, and it is what makes the output dependable over years. The olmo models are best understood as research infrastructure rather than standalone artifacts. Occasionally Ai2 gets something right first, and he cites the dataset behind the Molmo multimodal model.

His more opinionated description of Ai2 is that it is what OpenAI originally seemed like it was going to be, with enough resources to train solid models ("maybe not 400B yet"), explain them and maintain the tooling.

The headwinds are specific. The web is closing its archives to AI crawlers (the trend measured in consent-in-crisis-ai-data-commons). Ai2 could use ten times its compute. In his own area, open RLHF is up to years behind the closed labs, and the best open preference-tuning dataset is more than a year old. Industrial post-training has moved on to things like OpenAI's o1, and open efforts are "barely staying above water."

There are also few real allies. When Lambert met Yann LeCun, what stuck was LeCun saying "I need the help fighting the good fight." Big tech is financially committed to building large models regardless, so the marginal impact is in open work, where a small handful of people drive most of the progress on OLMo. The pitch to experienced model trainers is direct: waiting for big-tech options to vest is too slow, and "there is far more marginal impact to be had in roles where outputs are open."

Read against the later pieces

The essay's case is domestic and civic: concentration of power, security, research access, culture. A year later the ATOM Project recast nearly the same goals as a national competition with China, and by 2026 banning-open-source-ai-would-be-a-mistake and six-months-to-live-for-open-models were defending open models against restriction in Washington. The Olmo line did grow up. The OLMo papers (olmo-accelerating-the-science-of-language-models, olmo-2-furious, olmo-3-technical-report) track that growth, and by late 2025 Lambert was describing Olmo 3 32B Think as the best fully open language model yet made.