There’s plenty of chatter around AI safety these days, but these exact concerns have been raised for years — with a lot lesser attention.
Yann LeCun has used seven seven-year-old ammunition to take yet another swipe at Dario Amodei, this time resurfacing a 2019 episode where the Anthropic CEO argued that OpenAI’s GPT-2 was too risky to release to the public.
The moment was triggered by a post from the account Pessimists Archive, which shared a 2019 Guardian article describing how OpenAI had decided to withhold the full release of GPT-2, a text-generating AI model that the company said was capable of producing convincing fake news, spam, and fabricated product reviews. Amodei, who was OpenAI’s research director at the time, was quoted in the piece explaining the scale of the breakthrough, noting that the underlying models were far larger and had been trained on a dataset many times bigger than anything that came before. OpenAI’s head of policy at the time, Jack Clark, was also quoted describing the company’s caution as an attempt to “perform experimentation to find out what they can and can’t do” before the technology became mainstream.
LeCun, Meta’s former Chief AI Scientist and one of the most prominent skeptics of AI doom narratives, quoted the Pessimists Archive post with a blunt response. “Right. Dario was already claiming that GPT2 was too dangerous to open source back in 2019,” he wrote, adding that he had mocked the claim back then and that “everyone should make fun of them now.”
The jab fits a pattern that has been building for years. LeCun has repeatedly gone after Amodei’s public statements on AI risk, at one point calling him “deluded” about the dangers of AI and accusing him of having a “huge superiority complex” for continuing to build frontier models while warning about their dangers. He has also dismissed Amodei’s vision of AI labs producing a “country of geniuses in a datacenter” as “complete BS”, and more recently pushed back hard on Amodei’s predictions about AI wiping out half of white-collar jobs, telling people not to listen to AI lab leaders, including himself, on questions of labour economics.
The GPT-2 episode itself was a genuinely unusual moment for the AI industry in 2019. OpenAI, still a nonprofit at the time and backed by Elon Musk and Sam Altman among others, said the model produced text so convincing that it could pass off fabricated quotes from politicians and generate infinite fake product reviews at will, and the organization framed its decision to withhold the full model as an attempt to give the world time to grapple with the implications before the technology went mainstream. In hindsight, OpenAI did release GPT-2 in full later that same year, and the model looks unremarkable next to the large language models that followed, including the ones now built by Amodei’s own company, Anthropic.
Whether the caution around GPT-2 was prescient or premature is still debated in AI circles, and that debate has only gotten louder in recent days. Amodei has continued to make the case for treading carefully, at one point comparing AI companies that stay silent on safety risks to cigarette companies that hid the dangers of their product, and just last week he published a lengthy essay arguing that the pace of AI capability development needs to be deliberately slowed. That essay actually found some unlikely support, with Sam Altman and Elon Musk both publicly backing Amodei’s call to pace AI progress, even as OpenAI itself has cited safety concerns as a reason it won’t be going public in 2026.
Not everyone in Amodei’s own backyard is convinced, either. The recent resignation of Anthropic researcher Jacob Coxon, who left the company warning about existential AI risk, has become its own flashpoint, with Nvidia CEO Jensen Huang calling Coxon’s comments “outlandish” and “deeply untrue”, and David Sacks suggesting the researcher’s viral resignation post was less organic than it appeared, while Friedberg accused AI doomers of using fear the way religious institutions do. Amodei himself has since struck a more measured tone, saying he agrees with Coxon more than he disagrees with him, even as he stops well short of Coxon’s bleaker predictions.
Against that backdrop, LeCun’s comment on a seven-year-old Guardian article feels less like a random bit of trolling and more like a deliberate contribution to that live argument: a reminder, from one of the loudest voices on the other side of the AI safety debate, that Amodei’s instinct to sound the alarm did not start with existential risk or job-loss warnings, but goes all the way back to a text generator that today would barely register as noteworthy.