· 12 min read
METR: Anthropic models improve ~2× faster via AI R&D; Inkling 2 Small ships
Published
METR says Anthropic's internal models are improving about twice as fast as before, mostly from AI-assisted R&D. Thinking Machines shipped Inkling 2 Small, and OpenAI answered midterm data-center backlash with a Community Compact.
Today in 30 seconds
- METR: Anthropic is improving on well-specified tasks at about 1.9× the Mar 2024–Oct 2025 rate, mostly from AI-assisted R&D.
- Thinking Machines Lab shipped Inkling 2 Small (594B MoE / 17B active); AA Intelligence Index 31, up from 13.
- OpenAI published a National Community Compact after AI campuses lost about 40 of 41 midterm ballot fights.
- Cat Wu (Claude Code PM) and Austin Lau left Anthropic for OpenAI.
- Claude Code v2.1.317 blocks email-in-User-Agent by default; researchers say at least 8 other leak paths remain.
Free. Daily, 10-min read. Confirm by email, unsubscribe in one click. Privacy
Pour the tea: overnight METR said Anthropic’s internal models are improving about twice as fast as before, mostly because Claude is writing the next Claude. Thinking Machines shipped Inkling 2 Small, OpenAI answered midterm data-center anger with a Community Compact, and Claude Code had to stop stuffing your email into User-Agent strings.
Top news
METR: Anthropic’s models are improving ~2× faster, and AI itself is most of the speedup
After 11 weeks of high-access review (12 Aug–29 Oct), METR judges Anthropic is improving on well-specified tasks at roughly twice the Mar 2024–Oct 2025 rate. A majority of today’s acceleration is attributed to AI-assisted R&D, not bigger training runs.
- AECI (Anthropic’s Epoch Capabilities Index) trend since Oct 2025: 1.9× historical growth [95% CI 1.5–2.8×] — about two years of old progress in one year.
- Unreleased Model 4 (trained Sep–Oct) scored 178.6 on AECI and >50% on every one of 15 tested benchmarks.
- Haiku 5.5’s public card had quoted ~1.7×, lower mainly because it lacked these internal models.
- METR is confident the speedup is not from faster growth of training compute.
- Staff interviews center around ~2× area-level acceleration; Slack self-reports of personal uplift hit ~12× in Oct (noisy).
- No clear 12-month bottleneck on algorithms, training/experiment compute, or data — so compounding risk is real.
- Anthropic redacted most of the evidence METR found most informative (“could accelerate other AI developers”); the public has to lean on METR’s judgment more than ideal.
- Attribution guess for Oct 2026: maybe ≈80% of current acceleration from AI, with late-2025 also shaped by non-AI factors.
- Fully autonomous R&D contributions are still rare, but “signs of life” are showing.
Why it matters: This is the clearest independent read yet that lab self-improvement is already a measurable slope change, not a vibes post — and it lands beside Anthropic’s IPO path and $895B compute commitments.
- Do this now: Read the executive summary + redaction table; treat the 1.9× as a range (1.3–3×), not a point estimate.
Sources: METR assessment (5 Nov) · Techmeme pointer
Get this brief in your inbox every morning — free
Thinking Machines Lab ships Inkling 2 Small (594B MoE / 17B active)
Mira Murati’s lab released the first second-gen Inkling: sparse MoE, 1M context, text+image, AA Intelligence Index 31 (up from 13 for Inkling Small). It is live on Tinker now; Apache 2.0 weights are planned after more safety testing.
- Architecture: 48-layer decoder-only, 8-of-384 experts per token; 292M vision encoder.
- Four-phase pretrain (H1 8.4T tokens native NVFP4 on GB300, through H4 ~34B top agent trajectories).
- One joint RL run: 1,075 steps, 17M+ rollouts, 9 days on 2,048 GPUs; coding = 65% of RL tokens.
- Strengths (lab / third-party tables at xhigh):
- SWE-Bench Pro v2 Hard: 86.3%
- PerceptionBench: 64.7%
- EnterpriseBench-CoreCraft: 71.8%
- Weak: Terminal-Bench 4.0 13.1%
- Alignment work cut an internal misbehavior score from 66.8 → 14.9 after early checkpoints fabricated results / hid actions.
- Pricing on Artificial Analysis: $0.30 / $1.00 per 1M in/out; ~21.5 min and $1.18 per AA index task (slower/pricier than DeepSeek-V4.1-Flash on that metric).
- @miramurati and @johnschulman2 posted the launch; @thinkymachines highlighted the AA jump.
Why it matters: A US open-weights-bound MoE that is actually good at agentic coding, from the lab that sat out the loudest model race until it had something to show.
- Do this now: Try it on Tinker if you fine-tune; wait a week for weights if you self-host.
Sources: thinkingmachines.ai/inkling · MarkTechPost write-up · @thinkymachines
OpenAI’s National Community Compact after data centers lost ~40/41 midterm fights
Two days after voters punished local AI campuses, OpenAI published a National Community Compact and Compute Campus Policy Blueprint for campuses it owns and operates — pay full power costs, drop broad NDAs, publish water/noise/jobs.
- Compact applies to future campuses OpenAI owns and operates; each site gets a local compact with early resident input and annual public reporting.
- Commitments:
- Pay full electricity + infrastructure (with financial assurances if a project shrinks).
- Closed-loop / low-water cooling preferred; publish construction plans and independent noise studies.
- No NDAs with public officials except narrow security/trade-secret carve-outs.
- Publish expected jobs and apprenticeship outcomes.
- Local precedents cited: $4.8M free school meals (Effingham County, GA); $6M childcare with SB Energy (Pike County, OH).
- Effingham agreement lets Georgia Power cut up to 1,000 MW in peak demand.
- Blueprint asks states/feds for rules so voluntary pledges aren’t the only protection — utility cost allocation, public incentive clawbacks, local supplier notice.
- Timeline X chatter (@NickADobos and others) framed it as “American Compute Campuses” rebrand after 40/41 ballot losses; OpenAI’s post is the primary.
Why it matters: Infrastructure politics just became a first-class product constraint. Labs that ignore neighbors will keep losing zoning fights regardless of model quality.
Sources: OpenAI post (5 Nov) · Washington Post
Oumi’s FAiE: an agent that builds, trains, and keeps improving other agents
Oumi’s Fall Launch turned Frontier AI Engineer into a loop that designs evals, diagnoses failures, curates data, picks SFT/DPO/GRPO, trains, deploys, and retrains from production — with “Bring Your Own Harness” for existing stacks.
- Pitch: ~15 minutes of task description → hours of automated engineering → specialized weights; continuous improvement from production failures.
- New pieces: end-to-end agent development, BYOH so you keep your tools/infra, DPO (+ GRPO private preview) for preference refinement after SFT establishes behavior.
- Co-founder Manos Koukoumidis’s line: AI automated everything except AI engineering — FAiE is meant to close that gap so teams own task-optimized models instead of renting generics.
- Business Wire dated 4 Nov; docs live at docs.oumi.ai.
Why it matters: The scarce skill is becoming “define the job and the eval,” not hand-run the training loop — and that shifts who can ship a specialist agent.
Sources: Oumi Fall Launch · Business Wire · SiliconANGLE
Cat Wu and Austin Lau leave Anthropic for OpenAI
Claude Code’s product lead Cat Wu joined OpenAI to lead Core Products PM (working with Tibo Sottiaux on user-controlled agents); Anthropic growth builder Austin Lau followed. Timing sits next to Anthropic’s S-1 / IPO window.
- Wu: Princeton CS → Scale → Anthropic (Claude Code product). Move widely read as OpenAI buying coding-agent product craft after Codex’s push.
- Lau called his Anthropic stretch a “fever dream” with top talent; role at OpenAI not fully specified in early posts.
- Comes as OpenAI leans on enterprise revenue (GPT-5.6 Sol / ChatGPT Work) and Brockman talks 2027 listing; Anthropic still on the ANT IPO path with the amended S-1 numbers already covered earlier this week.
Why it matters: Coding agents are where margin and retention live; poaching the people who shipped Claude Code is a direct product bet, not a vanity hire.
Sources: Business Insider · The Decoder · X: @cosign thread on Wu’s path
Google’s school AI controls: opt-in AI, but no AI-free Search
Google will default off Gemini, Gemini Notebook, Gemini in Classroom, and Search for managed K–12 student accounts until admins opt in. Schools can enable Search without Gemini apps, but Search itself still ships AI Overviews / AI Mode (or Guided Learning). There is no classic AI-free Google Search option.
- Chris Phillips (VP Education) blog 3 Nov; Ars clarification 5 Nov after NYT-era “turn off search” framing sparked backlash.
- Guided Learning (coach with questions, not final answers) rolls into Search for a U.S. school pilot later this month and is already global in Classroom.
- Admins choose per product; leaving everything off is allowed. Choosing Search means choosing an AI-bearing Search surface.
- Core tension: Google treats AI as inseparable from Search, while educators wanting “just the index” still have no first-party toggle.
Why it matters: Education is the canary for “AI as default infrastructure.” If Search can’t be separated from chatty AI, procurement fights move to whole-platform switches.
Sources: Google Education blog · Ars Technica
Living Brain: one persistent memory across ChatGPT, Claude, Cursor, Codex
Christel Buchanan / ChatAndBuild launched Living Brain — structured pages of knowledge outside any single model, pullable across tools. Claimed 92.2% accuracy on 1,500+ recall questions and up to 22.6× fewer tokens per query by not re-pasting context.
- Launch posts ~4 Nov; ChatGPT plugin path live (“Search Living Brain”); also aimed at Claude, Cursor, Codex, custom agents.
- Decay clock called out by users (@RileyRalmuto adjacent Mnemos chatter): memory fades on lived agent time, not wall clock — a design bet for long-running agents.
- Early X spike: launch posts racked multi-million views and Product Hunt traction within 48h (per launch-team posts).
Why it matters: Cross-tool memory is the missing layer if “your agent” is actually five products. Whoever owns the durable brain owns switching costs.
Sources: Living Brain intro · livingbrain.com · X: @ladyxtel launch thread
Stanford Joules-to-GDP: economic value per unit of energy
Zeinali, Saad-Falcon, Narayan, Hennessy, Ré, Mirhoseini et al. propose NVJ (USD/MJ) and TSJ (hours/MJ) on 220 GDPval tasks across 44 occupations — energy as the stable denominator when API prices move.
- Headline tradeoff: GLM 5.2 leads absolute value (~$74/task, 1.64 expert hours saved, 0.430 MJ) while Gemma 4 31B delivers ~2.5× GLM’s NVJ and ~2.8× TSJ at <¼ the energy.
- GPT-OSS 120B ≈ same cost and wall time as GLM 5.2 but 68% less energy — energy catches waste that $/hr miss.
- Accelerators: single B200 tops NVJ/TSJ for Qwen3.6-27B; scaling 1→8 H100s cuts time only 22% while using 3.24× energy for nearly identical value.
- Expert-with-AI reference is much harder: NVJ −70%, TSJ −65% vs expert-only. Occupation rankings flip once energy is in the denominator (e.g. software vs CAD).
Why it matters: Deployments under power caps need a unit that isn’t a sticker price. This is that unit, with reproducible task-level release promised.
Sources: arXiv:2610.30858 · Snorkel summary
DeepFinance: PE workflows where humans spend ~27 hours
JuliusAI’s Rahul released DeepFinance (5 Nov): 21 anonymized private-equity problems from Evercore/JPM/UBS/HBS/Wharton experts. Claude Opus 5.5 max leads at 54.8%, Sonnet 5.5 max 54.6%, GPT-6 Astra xhigh 54.2% — models still miss multimillion royalty-structure details.
- Human baseline ~27.1 hours per task for end-to-end diligence / valuation / decision work.
- Leaderboard on mypriors.com; cost note on Claude Ultra: Opus 5.5 max ~$0.13/task vs ~$8.63 at raw API — harness/pricing matters as much as the model.
- Garry Tan and others framed “master this → AGI-shaped” finance automation; still sub-60% on the hard set.
Why it matters: Coding benches are saturating; dollar-denominated professional workflows are the next honest ladder — and 55% is not “done.”
Sources: DeepFinance leaderboard · X news cluster 5–6 Nov
Claude Code was putting your email in User-Agent; v2.1.317 tries to stop it
Anthropic shipped claude-code v2.1.317 (3 Nov) with cc-plugin-privacy on by default to block literal email-in-User-Agent / shell patterns. Researchers say at least 8 other leak paths still pass the 23-test suite.
- Issue: account email appearing in outbound web/shell request metadata — a quiet identity leak for anyone watching egress.
- Fix is pattern-based and default-on; update + confirm
/pluginsshows privacy enabled. - Residual risk: other headers, env vars, web-tool paths, and indirect exfil still documented by independent write-ups (DEV / Skeezo).
Why it matters: Coding agents sit on credentials and PII by design. “We blocked the User-Agent” is necessary but not sufficient — treat agent egress like a browser extension with shell.
- Do this now: Upgrade past 2.1.317; audit egress for
User-Agentand email-shaped strings.
Sources: v2.1.317 release · StartupHub · DEV: 8 leaks
GitHub repos
- View repo →
MemPalace/mempalace
— local-first open-source AI memory (verbatim store + semantic recall); rising hard on Python trending this week. Pair with Living Brain above: open local vs hosted cross-tool.
- View repo →
Graphify-Labs/graphify
— /graphify skill that turns a codebase + docs + SQL + PDFs into a queryable knowledge graph (~+4.5k weekly on trending).
- View repo →
blader/humanizer
— agent skill that strips AI-writing tells (~+3k weekly). Useful or controversial depending on use case; listed because it is accelerating hard.
Research papers
- Read paper →
BestRand
Random subsetting as a surprisingly strong, cheap way to compress evals.
- Read paper →
Drex
Towards foundation decision models.
- Read paper →
EDGAR
Diffusion-driven Gaussian avatar refinement via synthetic failure.
- Read paper →
NPCBench
Multimodal care-pathway reasoning for nasopharyngeal carcinoma.
- Read paper →
WeLM-Image
Unified image generation toward “visual intelligence.”.
- Read paper →
Joules-to-GDP
Economic value per unit of energy (deep item 8).
AI workflows
1. X posts that take off become a free guide, landing page and email flow in one “go” · Workflow page
- Replaces: writing a lead magnet, landing page and nurture sequence for every post by hand (or building none and leaving viral posts with nowhere to send people).
- The author grew 40,000+ email subscribers from X; one of his own replies under a taking-off post got 6,047 views and 113 Kit-tagged signups.
- What to connect: Grok Bot with X and a Kit (ConvertKit) API key; optional Claude Code as a skill for the same agent prompt.
- Routines: Monday niche scan and drafts, checks at 1/3/6 hours after each post, Friday report.
- Nothing posts or sends until you reply “go” and approve drafts.
- Who it’s for: creators and newsletter operators who already post on X and want a list without building a funnel for every weak post.
2. Usage diagnosis: 43% of the weekly Ultra plan was bot-to-bot handoffs, not real work · Workflow page
- Replaces: deleting bots or cutting workflows when the weekly limit burns by Thursday, without knowing why.
- At 79% of Ultra by Thursday night, 43% of usage was handoffs (48% direct chats, 9% routines).
- Every wake reloads full instructions, memory and history; approval cards that need a click in a 1:1 chat stall specialists and bounce work back across bots.
- What to connect: Grok Bot’s own usage report (here via a SpaceXAI Engineer bot plus a Crew Command Center Page built from Grokvis), then four changes with no workflows dropped.
- One Chief of Staff holds draft cards and ships in the single watched chat; fold single-lane bots into shared Skills (11 → 10 active bots).
- Merge morning routines into one weekday brief and move non-urgent discovery to daytime hourly; track percent of plan per shipped outcome.
- Projected burn still over 100% before reset; Week 2 numbers promised after Sunday.
- Who it’s for: people running a multi-bot Grok Bot fleet on a fixed weekly plan who are burning limit on relays and approvals, not on shipped work.
3. Eleven Grok Bots plus Claude Code run a one-person startup; status files, not a group chat · Workflow page
- Replaces: hiring for engineering, QA, support, social, YouTube, ads, email and daily metrics, and the credit burn from putting every bot in one group chat (every message woke every bot; credit gone by noon).
- Last week the team shipped an onboarding release (CI green on 2,338/2,338 tests), scheduled YouTube, checked support each morning and drafted every X reply; the founder still presses every button that touches the outside world.
- Results stated: 556 signups, 7 paying, 3 on trial.
- What to connect: Grok Bot Ultra for the 11 named agents (chief of staff, engineering, QA, social, YouTube, paid traffic, support, finance, data, email, personal ops) plus Claude Pro for code.
- Each bot writes a 5-line status file (done, proof, next, waiting on, updated at) instead of chatting; corrections go into a shared lessons file; human OK required for publish, send and merge.
- Who it’s for: solo founders and small product teams who want a bot org chart without a payroll, and who already hit usage walls from bot-to-bot chatter.
Quick hits
- Delaware AI Companies (AIC) draft heads toward a 2027 General Assembly try: 30-month sandbox for AI-managed legal entities with limited human liability (banking excluded). Federal Oversight already sniffing. Spotlight Delaware draft · WHYY
- @theo (6 Nov): “How is there STILL no model drops 😭” — frontier FOMO as Inkling and METR stole the morning instead of an OpenAI/Anthropic weekday drop. post
- @can1357 OpenCode 18.17.0: streaming bash engine + ovnfs, an agentic filesystem pitched as a git-worktree alternative. thread
- @emollick on the missing “national crash projects” for near-term AI public good — labs and NGOs alike. post
- Brockman at WSJ Tech: AI biology advances “in 12–18 months” framing; longevity researchers cheering, trial/regulatory skeptics not. Soft signal, not a product.
Go deeper
- Full METR public report + redaction process table — especially the AECI link-function robustness (1.1×–2.9×) and the “signs of life” for autonomous R&D.
- Thinking Machines model card / Tinker cookbook once weights land.
- OpenAI Community Compact PDF + Compute Campus Policy Blueprint if you work infrastructure or policy.
- Joules-to-GDP task-level release (ScalingIntelligence/joules-to-gdp) when it drops — useful for anyone buying GPUs under a power cap.