Why Meta’s Yann LeCun isn’t buying the AI doomer narrative

Back in late January, when everyone thought Big Tech was caught sleeping by OpenAI, Meta’s chief AI scientist Yann LeCun tweeted a blunt reality check.

He argued that established tech giants had far more to lose than cash-hungry startups by shipping imperfect demos. Google and Meta didn't hold back because they couldn't build something like ChatGPT; they held back because they wouldn't.

At the time, he was right. Tech giants had all the research muscle, but none of OpenAI's risk tolerance. Yet less than a year later, the script completely flipped.

From pulling the plug to blowing open the gates

Meta's early caution was understandable given its recent scars.

Just under two weeks before ChatGPT debuted, Meta rolled out Galactica, a chatbot designed for scientific tasks. It lasted all of three days before being dragged offline after intense backlash over hallucinated nonsense. Meanwhile, ChatGPT arrived with similar flaws, but OpenAI leaned into the chaos.

Fast forward a few months, and Meta responded by taking an arguably bigger gamble than anyone else: releasing Llama 2.

While OpenAI keeps frontier models like GPT-4 behind a subscription fee and closed API, Meta made Llama 2 freely available for commercial and research use. Purists quickly pointed out that Meta kept its training data private, meaning it is not strictly open source in the traditional sense, but the move still blew the doors wide open for external builders.

Stripping power from a handful of engineers

For LeCun, who pioneered deep learning across the 1990s and early 2000s before founding FAIR in 2013, Meta’s shift is personal.

He envisions a world where search engines and social platforms fade into the background. Instead, everyone will rely on an AI virtual assistant as their primary interface to the digital world.

If that future pans out, LeCun argues, having critical AI infrastructure controlled by a tiny circle of Silicon Valley executives is a massive hazard. The only safe hedge is making these models widely accessible and transparent.

Why it matters

For developers and startups, Meta’s strategy is a massive win. Instead of being locked into OpenAI's billing meter and proprietary black box, teams can run, fine-tune, and inspect Llama 2 inside their own infrastructure without leaking data.

Strategically, Meta pulled off a classic commoditization move. Because Meta doesn't rely on selling model compute as its core business, giving away base models undercuts OpenAI's pricing power and forces the entire frontier ecosystem to compete on tooling rather than basic access.

We think the move is undeniably brilliant, but it is not without friction. Holding back the training data leaves open-source purists skeptical about true transparency, and distributing powerful weights freely means Meta essentially offloads safety enforcement to whoever downloads the model.