{"event":{"id":"evt-chinchilla-20220329","dedupe_key":"scaling_law_research:2022-03-29:chinchilla-preprint-reframes-compute-optimal-language-model-scaling","event_type":"scaling_law_research","title":"Chinchilla preprint reframes compute-optimal language-model scaling","summary":"DeepMind's March 29 preprint argued that many large language models were undertrained for their compute budgets and reported better performance by balancing model size with substantially more training tokens.","occurred_at":"2022-03-29T13:38:03.000Z","published_at":"2022-03-29T13:38:03.000Z","observed_at":"2026-10-01T17:04:00.000Z","reconstructed_at":"2026-10-01T17:34:32.000Z","ingest_type":"live_world_scan","locations":null,"status":"verified","confidence":0.98,"metadata":{"actors":["DeepMind"],"what_happened_at_time":"A compute-optimal scaling analysis and Chinchilla results were publicly released as a preprint.","evidence_at_time":"The arXiv submission was dated March 29.","retrospective_inference":"Training data/model balance and inference efficiency became more explicit design variables inside the compute-model dependency.","uncertainty":"The empirical rule is not a universal physical law; later ecosystem adoption is outside the Q1 fact."},"created_at":"2026-10-01T18:05:36.371Z","updated_at":"2026-10-01T18:05:36.371Z"},"artifacts":[{"id":"art-deepmind-chinchilla-20220329","canonical_url":"https://arxiv.org/abs/2203.15556","artifact_type":"paper_preprint","title":"Training Compute-Optimal Large Language Models","summary":"DeepMind reported that many large language models were undertrained for their compute budgets and that model size and training tokens should scale together under its empirical compute-optimal analysis.","creator_entities":["DeepMind"],"released_at":"2022-03-29T00:00:00.000Z","content_hash":null,"metadata":{"historical_scope":"2022-Q1"},"created_at":"2026-10-01T18:05:36.362Z","updated_at":"2026-10-01T18:05:36.362Z","relation":"documented_by"}],"signals":[{"id":"sig-compute-optimal-scaling-2022q1","statement":"Model size alone was becoming an insufficient description of frontier progress; training-token allocation and inference cost emerged as important variables in the model-compute relationship.","signal_type":"efficiency_shift","direction":"increasing","confidence":0.95,"epistemic_status":"supported_inference","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T17:34:32.000Z","created_at":"2026-10-01T18:05:36.381Z","updated_at":"2026-10-01T18:05:36.381Z"},{"id":"sig-preagent-stack-coupling-2022q1","statement":"By the end of Q1 2022, the clearest stack coupling ran through research, models, software, compute, chips, networks, manufacturing and governance; the selected evidence does not yet support persistent agents or agent permissioning as independent operational layers.","signal_type":"civilization_stack_coupling","direction":"increasing","confidence":0.86,"epistemic_status":"supported_inference","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T17:34:32.000Z","created_at":"2026-10-01T18:05:36.382Z","updated_at":"2026-10-01T18:05:36.382Z"}]}