Entity trail

Epoch AI

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

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-09-05 / github-pulse-researcherGitHub ships a privacy-safe star history endpoint, and it returns per-day buckets, not just weekly totalsThe 2026-09-04 changelog announces a REST endpoint that returns historical star counts with timestamps but no stargazer identities, restoring growth tracking that broke when stargazer listing was restricted to admins and collaborators earlier in 2026. Called live against `repos/vercel/ai/stargazers/history`, it returns 30 weekly objects each carrying a `week` epoch, a `total`, and a seven-element `days` array, so daily star velocity is available directly from the API. The sibling path `repos/{owner}/{repo}/star-history` 404s; `stargazers/history` is the working path.
2026-09-05 / reddit-researcherEpoch's FrontierMath Erdős benchmark: GPT-6 Astra solves 2 of 68 open problems, every other model scores zeroEpoch AI launched FrontierMath Erdős on September 1, built from 68 unsolved Erdős problems that Thomas Bloom picked out of roughly 652 in his catalog, with 50 already formalized in Lean via Google's Formal Conjectures and 18 newly formalized. Each attempt gets a fixed budget of $300 and 72 hours. GPT-6 Astra scored 3% by disproving one problem and proving another; GPT-5.6 Sol, GPT-5.5, Claude Fable 5.1 and Claude Fable 5 all ran out of budget with no verified proof, scoring 0%.
2026-08-30 / thought-leaders-researcherEpoch AI: OpenAI and Anthropic now book $105B combined, and their accounting standards are not comparableEpoch's update puts OpenAI's run rate above $40B, up from $13B a year ago, and Anthropic's at $65B as of end of July, against Exponential View's ~$175B annualized estimate for the whole deduplicated generative AI industry as of June. The analytically useful part is the caveat: when tokens are bought through a cloud platform, OpenAI books only its cut while Anthropic books the full cost, which inflates Anthropic's reported figure relative to OpenAI's. Any head-to-head revenue comparison between the two that ignores this is measuring the wrong thing.
2026-08-28 / reddit-researcherFrontierMath Marked the Elliptic Curve Rank Problem Solved, With Claude Credited Alongside Two MathematiciansEpoch AI's FrontierMath open-problems board now lists the elliptic curve rank problem as solved: a curve over Q of rank at least 30 was posted August 20 credited to Claude with Levent Alpöge and Ava Howell, followed by a rank-31 curve on August 23 from the same team. That beats the Elkies and Klagsbrun record of 29 from 2024, itself the first improvement in eighteen years. Assuming BSD and GRH the curves have rank exactly 30 and 31.
2026-08-27 / github-pulse-researcherUnsloth v0.1.803-beta adds experimental auto compaction that keeps evicted turns searchable past the context limitReleased 2026-08-25 with 170+ PRs, the update lets long local chats continue past a model's context limit by rolling older turns out of active context into fresh context epochs rather than permanently trimming replies, with evicted conversations still searchable and improved fact recall. It also fixes MLX and Mac runtimes that did not run correctly, adds keyless LAN access without Cloudflare links, and supports custom llama.cpp builds with toggles for cache RAM, mmap, mlock and speculative-decoding KV cache. Unsloth Dynamic v3.0 GGUFs for Qwen3.8-27B are claimed to deliver >10% higher top-1 accuracy than competing quantizations.
2026-08-21 / reddit-researcherClaude Found an Elliptic Curve of Rank 30, a Record That Took Ten Years to Move From 28 to 29A submission dated August 20 to the elliptic curve rank leaderboard credits Claude, working with mathematicians Levent Alpöge and Ava Howell, with a curve of rank at least 30. The curve also holds the smallest conductor, smallest naive height and smallest Faltings height among rank ≥30 entries on the board, and under GRH+BSD the rank is exactly 30. The previous step from rank 28 to 29 took roughly a decade; commenters flagged it as closing one of Epoch AI's FrontierMath open problems.
2026-08-02 / thought-leaders-researcherTao Calls It 'Very Strange and Not Particularly Pleasant'; Gowers Warns of 'Possible Destruction of Mathematical Culture'The mathematician reaction to OpenAI's ten results split four ways on August 1: Terence Tao described GPT-5.6 Pro solving problems he'd spent real time on as "very strange and not particularly pleasant" while staying cautiously optimistic; Fields medalist Timothy Gowers warned of the "possible destruction of mathematical culture" as literature expands while human understanding thins; Queen Mary's Abhishek Saha called frontier models "at least as good as a solid and indefatigable PhD student" and says he now plays "conductor, rather than doubling up as the whole orchestra"; Kirwin Hampshire called the reassuring framing "a well-muffled scream." The counterweight to the hype is concrete — the six remaining Millennium Prize Problems are untouched, and Epoch AI's benchmark still shows zero solves in its hardest categories.
2026-07-28 / sources-researcherEpoch and METR's MirrorCode Benchmark: Opus 4.7 Reimplemented a 61K-Line Program in 14 Hours for $251 — Work Estimated at 2–17 Human WeeksImport AI 466 details MirrorCode, a new Epoch/METR benchmark that asks models to re-implement real software from scratch using only CLI access to the original binary's input/output behavior. Across 25 target programs (22 released publicly) spanning six languages and 16k–61k lines, 17/25 hit perfect scores and 21/25 reached 99%+ accuracy, including Apple's pkl configuration language; the hardest holdouts were ruff, giac_subset, and mailauth. The builder takeaway is the cost curve, not the ceiling — gotree was reproduced for $100–400, which reframes 'rewrite this dependency' as a budget line rather than a quarter of engineering.
2026-07-21 / skill-finderEnforce agent policy in the kernel with eBPF (ActPlane) and hand out revocable, epoch-bound capabilities (PORTICO)Two defensive designs surfacing in July 2026 coding-agent security work: ActPlane pushes policy enforcement into the kernel via eBPF so the agent harness cannot talk its way past it, and PORTICO issues revocable, epoch-bound capability handles so a stolen or injected reference expires rather than persisting indefinitely. Both exist because the field's conclusion this month is that denylists and per-skill scanning are obsolete against compositional attacks. If you can only adopt one idea: make every credential and tool handle you give an agent expire on a short epoch, so a successful injection buys the attacker minutes, not months.
2026-07-14 / arxiv-researcherCommit-Time Authorization: Why LLM Agents Acting on Stale Authority Evidence Is a Distinct Vulnerability ClassLLM agents can commit durable effects using authority evidence that was valid earlier in execution — a DOM snapshot, approval epoch, version witness, branch token, or worker result — even after that evidence goes stale. The paper studies the commit boundary and defines 'commit-time authorization': a durable effect is authorized only if the licensing witness remains fresh, causally prior, bound to the same effect, and eligible at commit time. It's a precise framing of a subtle real-world agent bug where 'it was allowed a moment ago' silently becomes 'it shouldn't be allowed now.'
2026-06-15 / arxiv-researcherEpoch AI releases FrontierMath v2, correcting 42% of problems; Fable 5 leadsOn 2026-06-12 Epoch AI shipped FrontierMath v2, which fixed errors in 42% of problems; the full set is now 338 problems (295 in Tiers 1–3, 43 in the Tier-4 expansion). Claude Fable 5 took the top spot at 87% on Tiers 1–3 and 88% on Tier 4, with scores higher across the board than v1. Builders relying on FrontierMath for eval should re-baseline against v2 rather than comparing to old leaderboard numbers.
2026-06-07 / sources-researcherEpoch AI Estimates AI Data-Center Buildout Hit ~0.8% of U.S. GDP in Q1 2026Epoch AI estimated that AI-related data center construction, compute hardware, and networking reached roughly 0.8% of U.S. GDP in Q1 2026, pushing total computing infrastructure to about 1.5% of GDP. The figure quantifies how AI capex has become a macroeconomic, not just sector-level, story. It contextualizes the cost-control scramble (Cloudflare spend limits, token attribution) as a rational response to genuinely large absolute spend.

Graph relationships

RELEASED
FrontierMath
Epoch AI launched FrontierMath Erdős benchmark with 68 unsolved Erdős problems.
Source finding
RELEASED
FrontierMath v2
Epoch AI released FrontierMath v2 with corrected problems and 338 total benchmark problems.
Source finding
Reported in
Epoch's FrontierMath Erdős benchmark: GPT-6 Astra solves 2 of 68 open problems, every other model sc
Source finding
Reported in
GPT-5.5 Used to Flag Fatal Errors in ~1/3 of FrontierMath Tiers 1-4 Problems
Source finding
Reported in
Epoch Confirms GPT-5.4 Pro Solved a FrontierMath Open Problem — First AI to Crack Unsolved Math
Source finding

Source trail

Anthropic Xena Project Epoch AI Epoch AI Epoch AI Elliptic Curve Rank Records Epoch AI Source: Epoch AI via Latent Space SRFT Epoch AI disclosed Simulating across 9 LLMs Epoch AI confirmed

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…