Meta AI / Facebook AI Research (FAIR)

Contribution to AI

Two of the field's most-used pieces of infrastructure came out of this lab a decade apart, and neither was a product. PyTorch, first released in 2016 as a Python successor to Torch, built its computation graph as the code ran rather than compiling it in advance, so a model could be inspected with an ordinary Python debugger and its control flow written with ordinary loops and conditionals. Researchers migrated quickly, and by the end of the decade the majority of new papers shipping code shipped PyTorch code; governance passed to the PyTorch Foundation under the Linux Foundation in 2022 rather than staying in-house. Faiss was narrower and just as durable, making approximate nearest-neighbour search over billions of vectors fast enough on GPUs to sit underneath the retrieval systems and vector databases built years later. The LLaMA release of February 2023 was the larger intervention. Its argument was that a 13-billion-parameter model trained only on publicly available data could beat GPT-3's 175 billion on most benchmarks, shifting the question from how large a model should be to how many tokens it should see and what it costs to run. The weights went to researchers under a non-commercial licence and reached BitTorrent within a week; Alpaca, Vicuna and llama.cpp followed within a month, and a laptop became a plausible place to run a language model. Llama 2, licensed for commercial use that July, made that ecosystem legitimate rather than merely tolerated.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).