AI in 2022
10 events, oldest first. Each links to a page with details and sources.
- InstructGPT: RLHF aligns language models to follow instructions ★★★★★
OpenAI fine-tuned GPT-3 with reinforcement learning from human feedback (RLHF); labelers preferred outputs of the 1.3B InstructGPT over the 175B GPT-3, and the method became the recipe for ChatGPT.
- Chain-of-thought prompting elicits reasoning in LLMs ★★★★
Wei et al. showed that prompting large models to write out intermediate reasoning steps dramatically improves performance on math and logic tasks — an ability that emerges with scale.
- Deep reinforcement learning controls fusion plasma in the TCV tokamak ★★★★
DeepMind and EPFL (Nature, Feb 2022) trained a single deep-RL policy in simulation that commanded all of TCV's magnetic control coils on the real machine. It produced and held elongated, negative-triangularity and 'snowflake' plasmas, and even two separate 'droplet' plasmas at once.
- NVIDIA announces the H100 'Hopper' GPU ★★★★
NVIDIA unveiled the Hopper architecture and H100 GPU with a Transformer Engine and FP8 support; the H100 became the defining AI training chip of the generative AI boom.
- DeepMind's Chinchilla revises scaling laws toward more data ★★★★
Hoffmann et al. found that for compute-optimal training, parameters and training tokens should scale equally (~20 tokens per parameter); 70B Chinchilla outperformed the 280B Gopher.
- DALL·E 2 brings photorealistic text-to-image generation ★★★★
OpenAI's DALL·E 2 used a diffusion decoder conditioned on CLIP embeddings to generate high-resolution, photorealistic images from text, kicking off 2022's image-generation boom alongside Midjourney and Stable Diffusion.
- Stable Diffusion released as open weights ★★★★★
Stability AI and collaborators released Stable Diffusion, a latent diffusion text-to-image model small enough to run on consumer GPUs, with openly downloadable weights — democratizing image generation.
- AlphaTensor discovers faster matrix multiplication algorithms, beating Strassen's 1969 record for 4×4 mod 2 ★★★★
DeepMind's AlphaTensor (Nature, Oct 2022) framed matrix multiplication as a tensor-decomposition game. It found a 4×4 algorithm over GF(2) with 47 multiplications (Strassen-based: 49) and improved 5×5 to 96. Human researchers cut 5×5 further to 95 within days.
- OpenAI launches ChatGPT ★★★★★
OpenAI released ChatGPT, a conversational interface to a GPT-3.5 model fine-tuned with RLHF, as a free research preview; it became the fastest-growing consumer app to that point and triggered the generative AI boom.
- Anthropic introduces Constitutional AI (RLAIF) ★★★★
Anthropic's Constitutional AI trained a harmless-but-helpful assistant using AI feedback guided by a written set of principles (a 'constitution') instead of human harm labels.