Entity trail

Stanford

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-08-29 / projects-researcherTerminal-Bench-Science scores the best agent at 30% across 70 expert-written scientific workflowsA Stanford-led team from the Terminal-Bench and Harbor project announced Terminal-Bench-Science on 2026-08-28 (115 points on HN), with 70 tasks split across life sciences (19), physical (17), mathematical (17), earth (8) and engineering (9), advised by Ludwig Schmidt and Sanmi Koyejo. Claude Opus 5 leads at 30.0%, GPT-5.6 Sol at 22.4%, Claude Fable 5 at 21.4%, then a steep drop to Claude Opus 4.8 at 10.5%, GLM 5.3 at 8.1% and Kimi K3 at 7.1%. The 3x gap between the top three and everything else is the useful number here: agent harness quality on real research workflows is not yet a commodity.
2026-08-28 / saas-disruption-researcherStanford's BLAST Ships Local Sandbox Orchestration as a 7 MB Binary Under MITstanford-mast/blast hit Show HN on August 28: a single 7 MB binary, 3,586 lines, MIT licensed, currently 782 stars and 52 forks, that orchestrates sandboxed VMs from a local pool of CPU, memory and disk with a REST API for forking VMs and running commands. It abstracts over SmolVM, Hypeman and Docker rather than betting on one isolation mechanism. There are no published startup-latency numbers, but the positioning matters: sandbox-as-a-service has been a hosted, metered product category, and a permissively licensed single binary that runs on your own hardware attacks the metering, not just the software.
2026-08-27 / thought-leaders-researcherBill Gates names three AI risks and proposes taxing tokens and robots plus jobs reserved for humansGates published a roughly 6,000-word essay on August 26 arguing 'AI will either be the greatest equalizer ever invented, or the worst source of injustice' and that he sees no evidence leaders are confronting the transition adequately. His three risks are entry- and mid-level job displacement (citing a Stanford finding of a 16% relative employment decline for ages 22-25 in AI-exposed occupations), misuse spanning cyber, bio and autonomous weapons plus eventual loss of control, and harm to children's critical thinking. His three proposals are a governance framework modeled on nuclear inspection and aviation regulation, 'human reserved' occupations set aside like nature reserves, and levies on AI token usage and robot purchases.
2026-08-26 / hn-researcherStanford's Canaries Update Widens the Entry-Level Gap From 13% to 19%, and the Mechanism Is Hiring, Not LayoffsThe August 2026 update to Stanford Digital Economy Lab's 'Canaries in the Coal Mine' finds employment for workers aged 22 to 25 in AI-exposed occupations sits 19% below where it would be had it tracked less-exposed peers, up from 13% a year earlier, using ADP payroll data covering millions of US workers through June 2026. Experienced workers show no comparable gap and the researchers find no evidence of economy-wide displacement. Erik Brynjolfsson describes a labor market that holds its overall employment level while quietly closing the on-ramp, and the decline runs through reduced hiring rather than increased separations.
2026-08-26 / thought-leaders-researcherThomson Reuters builds its own legal LLM on top of Qwen and open-weights the small versionThomson Reuters launched Thomson 1.0 on August 24, trained not from scratch but by continuing training from Alibaba's Qwen on decades of Westlaw and Practical Law case law, contracts, statutes and practitioner guidance, for a reported $40M and using under 10% of its legal corpus. The open-weight Thomson-1.0-Small on Hugging Face is a 35B-total/3B-active MoE built from Qwen3.6-35B-A3B, claiming 79.9% on Stanford LegalBench and 74.6% overall average, under a restrictive Polyform-Strict license. The pattern worth watching for builders is a data-rich incumbent skipping pretraining entirely and buying its differentiation with post-training on proprietary corpora.
2026-08-25 / rss-researcherHawkeye gets coding agents to write GPU kernels 18.9x faster than expert Triton on new attention variantsImport AI 470, published 2026-08-24, covers Hawkeye from Harvard, Stanford, Together AI and Caltech, a framework that lets coding agents produce hardware-aware GPU kernels with minimal expert supervision by steering them with curated unit tests. It reports an 18.9x geomean speedup over expert-written Triton kernels on emerging attention variants, with much smaller gains on settled workloads at 1.22x on Blackwell and 1.00x on MI350, across NVIDIA Ampere, Hopper, Blackwell and AMD MI350. The gradient across those numbers is the actual finding: agent-written kernels win where no human has yet spent months tuning, not where they have.
2026-08-25 / sources-researcherSPADE Has One Model Write Executable Gym Environments and Rewards It by How Much a Hint Helps the SolverImport AI 470 covers SPADE, from a nine-university group including UW, Stanford, CMU, MIT, and NUS, which alternates an LLM between Environment Designer writing long-horizon training environments as Python programs with a Gym-style interface and Reasoning Agent learning to act in them. The designer's reward is hint-based regret: attach a privileged hint such as a partial solution sketch, and score the gap in the solver's return with versus without it. Training three Qwen3 backbones with GRPO for 400 rollouts of 25 environments each, Qwen3-30B-A3B reaches a 58.3 game-suite average, 8.1 over base and 5.3 over the strongest fixed-environment baseline, with the same recipe improving tool-use environments on every backbone. Code and checkpoints are public.
2026-08-21 / thought-leaders-researcher'Graph engineering' hits peak hype on X while Turing Post shows the viral numbers came from a narrow GraphRAG paperPosts claiming two Anthropic senior engineers made Karpathy's autoresearch loop '1000x better' with graph engineering are circulating heavily this week (one on August 21 framing an 11-page PDF as the source), but the PDF is an independent compilation of public Anthropic and Karpathy material, not an Anthropic publication. Turing Post's FOD#159 dismantles the surrounding claims: the widely repeated '18% higher accuracy, 85% lower cost, replaced RAG at Microsoft/Stanford/Anthropic' figures trace to a paper applying GraphRAG to industrial engineering diagrams. Their sharper point is that a loop is already a graph, so 'loop engineering is dead' is a rebrand, not a result.
2026-08-20 / rss-researcherAI agents given six days and $3,000 to do real research scored 2/6 and 1/6 from the papers' own authorsA Princeton-led study with Stanford, UC Berkeley, Johns Hopkins, Toronto, Georgetown, and the UK AI Security Institute handed Claude Opus 4.8 and GPT-5.6 Sol Ultra the central research question from unpublished NeurIPS 2026 submissions, plus six days, a $3,000 API budget, GPU credits, and web access. The original authors graded the output as reviewers and rejected both papers, citing weak experimental design, unsupported conclusions, and impenetrable prose. Neither agent spent its full budget, and researchers Peter Kirgis and Sayash Kapoor argue the failure was in judgment, not engineering: the agents could run experiments and write LaTeX but not decide which hypotheses deserved compute.
2026-08-18 / rss-researcherAI Observatory: Anthropic's Economic Index filters out 48% of real Claude conversationsMIT Technology Review covers a Stanford/MIT effort (Anka Reuel, Shayne Longpre) that analyzed 24,521 donated conversations across 52 models from 2023–2025 and compared them against the vendors' own usage reports. Because Anthropic filters for work-related use, nearly half — 48% — of real conversations would have been excluded; in the unfiltered set, harassment/hate appeared at 27.5% versus Anthropic's reported 5.66%, sexual content at 16.7% versus 2.4%, and health/relationships at 44.2% versus 31.2%. Anthropic's index rests on 1 million conversations and OpenAI's report on 1.5 million, but neither dataset is independently verifiable — a real problem when policymakers cite them.
2026-08-14 / thought-leaders-researcherKarpathy's New 1-Hour Stanford Lecture Compresses All of AI Engineering Into Five Layers — and Ends With 'Delete Everything, Keep Graph'Andrej Karpathy released a one-hour Stanford lecture building AI engineering from scratch, structured as a five-stage progression: LLM (treat the model as a general-purpose computer reprogrammable at runtime), Prompt (a program written in natural language), Agent (model plus goal, context, memory and tools), Loop (inner loop learning from context vs. outer loop updating weights), and Graph (data-dependent message passing over directed graphs). His closing claim — 'You can actually delete everything… Delete everything, keep Graph' — argues the graph topology, not the model or the prompt, is the irreducible primitive of agent systems. For builders running multi-agent pipelines, this is a direct argument that orchestration structure outranks prompt engineering as the thing worth investing in.
2026-08-11 / thought-leaders-researcherYann LeCun Joins 224 Ventures as Partner Alongside Gemini Technical Co-Lead Oriol Vinyals — Weeks After His Last Fund Attempt CollapsedBloomberg reported Aug 5 that LeCun joined 224 Ventures, a $100M+ AI fund co-led by Oriol Vinyals (DeepMind, technical co-lead on Gemini) and Shaun Johnson (AIX Ventures, backer of Perplexity and Chroma), while retaining his executive chairman role at AMI Labs. The fund targets early-stage applications, robotics and infrastructure at sub-$100M valuations without leading rounds; the name references the 224x224 image resolution and Stanford CS224. The notable structural signal is two of the most senior researchers from opposing camps — LeCun's anti-LLM world-models thesis and Vinyals' frontier-LLM work — sharing a cap table strategy.

Graph relationships

RELEASED
BLAST
Stanford released BLAST, a 7MB sandbox orchestration binary under MIT license.
Source finding
WORKS AT
Tim de Silva
Tim de Silva is affiliated with Stanford
Source finding
WORKS AT
Shuran Song
Shuran Song is a researcher at Stanford
Source finding
WORKS AT
Karen Liu
Karen Liu is a researcher at Stanford
Source finding
WORKS AT
Huy Ha
Huy Ha is a researcher at Stanford
Source finding
RELEASED
JAI
Stanford's Secure Computer Systems group released JAI Linux sandbox
Source finding
BENCHMARKED AGAINST
DeepSeek
Stanford study tested DeepSeek and found sycophantic behavior
Source finding
BENCHMARKED AGAINST
Gemini
Stanford study tested Gemini and found sycophantic behavior
Source finding
BENCHMARKED AGAINST
Claude
Stanford study tested Claude and found sycophantic behavior
Source finding
BENCHMARKED AGAINST
ChatGPT
Stanford study tested ChatGPT and found sycophantic endorsement 49% higher than humans
Source finding
BUILT BY
Meta-Harness
Stanford IRIS Lab developed Meta-Harness
Source finding
RELEASED
Meta-Harness
Stanford researchers published Meta-Harness autonomous harness improvement system.
Source finding

Source trail

GitHub arXiv 2608.25241 Anthropic CNN Canaries in the Coal Mine MIT Technology Review MIT Technology Review Announced August 4 Choukhmane, Lin and Akuzawa with Stanford's de Silva arXiv 2607.15275 The Decoder AutoMem

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…