Organisation
Stanford HAI
Stanford's human-centred AI institute, best known for the annual AI Index, a widely cited survey of the field's progress, investment and policy.
Stanford's Institute for Human-Centered AI, founded in 2019, is an interdisciplinary research centre that studies AI's technical progress alongside its social, economic and policy effects. It is best known for the annual AI Index, a data-heavy survey of benchmarks, investment, hardware and regulation that has become one of the most cited reference points for how fast the field is moving. Its affiliated Center for Research on Foundation Models coined the term "foundation model" in a 2021 report that named the category now central to the technology, though some researchers argued the label legitimised an industry path rather than critiquing it. HAI continues to publish research and convene work at the intersection of AI capability and public interest.
- Category
- Academic labs & institutes
- Founded
- 2019
- HQ
- Stanford, US
- Key people
- Fei-Fei Li, James Landay
Appears alongside
Featured in threads
Tracks
- Ideas & essays 6
- Benchmarks & progress 4
- Open weights & ecosystem 1
- Compute & infrastructure 1
Stanford HAI releases 2026 AI Index Report
Stanford's AI Index reports coding-benchmark scores jumping from 60% to near 100% in a year, alongside a 'jagged frontier' where an IMO gold-medal model reads analogue clocks correctly only half the time.
Benchmarks & progress
Stanford HAI releases 2025 AI Index Report
The eighth annual report put US private AI investment at $109.1 billion in 2024, nearly twelve times China's $9.3 billion, and inference cost for GPT-3.5-level performance down over 280-fold since late 2022.
Benchmarks & progress · Ideas & essays
Stanford HAI releases 2024 AI Index Report
The report put GPT-4's training compute cost at roughly $78 million and Gemini Ultra's at $191 million, and found industry produced 51 notable models in 2023 to academia's 15.
Benchmarks & progress
Direct Preference Optimization paper reframes RLHF as a classification loss
The method skipped the separate reward model and reinforcement-learning loop, and was later adopted for post-training open models including Zephyr and Tulu.
Ideas & essays
'Are Emergent Abilities of Large Language Models a Mirage?' challenges emergence claims
Reanalysing the same benchmark results with linear metrics, the Stanford authors made the apparent phase transitions disappear; the paper won a NeurIPS 2023 outstanding paper award.
Ideas & essays · Benchmarks & progress
Stanford's Alpaca fine-tunes LLaMA for a few hundred dollars
Instruction-following behaviour was reproduced for under $600 total by fine-tuning Meta's 7B LLaMA on GPT-3.5-generated examples; the public demo was pulled within days.
Open weights & ecosystem · Ideas & essays
FlashAttention makes exact attention IO-aware
Reordering attention around GPU memory rather than approximating it cut training time and unlocked longer sequences — and became default infrastructure.
Ideas & essays · Compute & infrastructure
Stanford's foundation models report names the category
Over 100 researchers at Stanford's newly formed Center for Research on Foundation Models coined the term for models like GPT-3 and BERT, adapted rather than retrained for each task.
Ideas & essays