The open-source AI boom is precariously built on top of Big Tech's giant models like LLaMA and GPT-3, and could fold if Meta and OpenAI decide to close up shop
but the next generation of AI breakthroughs will be entirely back in the hands of the biggest, richest AI labs in the world” https://www.technologyreview.com/ ... Matt Rogerson / @mattrogerson : This is really good by @strwbilly. The fate of this boom rests in familiar hands. Everything has changed, nothing has changed! The open-source AI boom is built on Big Tech's handouts. How long will it last? https://www.technologyreview.com/ ... Max von Thun / @maxvonthun : Despite what Google's ‘leaked’ memo would have you think, this incisive piece by @strwbilly shows how most ‘open-source’ AI is built atop a handful of centralised models - and why access to those is likely to quickly narrow. https://www.technologyreview.com/ ... Alan Lau / @alanlau999 : So @OpenAI wont be open anymore? Smell like #BigTech monopolies all over again. AI will determine how info (and our world) is organised. Scary to think it's left in the hands of a few again https://www.technologyreview.com/ ... Michael Jackson / @workmj : “If the trend toward closing down access continues, then not only will the open-source crowd be cut adrift—but the next generation of AI breakthroughs will be entirely back in the hands of the biggest, richest AI labs in the world.” https://www.technologyreview.com/ ... @techreview : The future of how AI is made and used is at a crossroads. https://www.technologyreview.com/ ...
Context & Ripple Effects
A week after Google's leaked internal memo argued open-source models would outcompete the big labs, MIT Technology Review's piece makes the opposite point: most of what gets called open-source AI is a derivative layer on Meta's LLaMA and OpenAI's GPT-3, and the labs that handed those models out can take them back. The boom's independence, on this reading, is a license term, not a structural fact.
The argument has aged well. Two years later, [[a:887924|Meta's superintelligence lab under Alexandr Wang discussed abandoning its open-source flagship Behemoth for a closed model]] — exactly the withdrawal scenario the piece flagged — and researchers had already spent years criticizing unequal access to code, data, and hardware in the field.
First-order effects
- Startups and researchers building fine-tuned products on LLaMA and GPT-3 are exposed to a single decision by Meta or OpenAI: restrict the license or stop releasing weights, and their foundation disappears overnight.
Second-order effects
- If Meta closes its models, the 'open' ecosystem migrates to whichever lab still releases — turning weight releases themselves into a competitive weapon, the inverse of the Google memo's assumption that open-source wins by default.
Third-order effects
- The pattern points toward frontier breakthroughs reconcentrating in the richest labs, with model access becoming the chokepoint that determines who can build on AI at all — the dependence researchers flagged back in 2020, now enforced by licensing rather than scarcity.
The trend: Open-source AI is discovering that its supply chain runs through frontier labs' licensing desks, making model-access control — not community effort — the deciding force in whether the ecosystem stays independent.