Meta releases its first frontier model, Muse Wonder

TechCrunch, 2026-10-12. Meta Superintelligence Labs' first frontier release. The research post and Vercel gateway availability landed the same day.

Mark Zuckerberg says Meta has reached AI 'frontier', rejects calls for slowdown

Slashdot, 2026-10-12. Refusal to pace stated as launch messaging. Read against Sunday's Alignment Forum post, this is the argument acquiring an opposing side rather than a rebuttal.

AI coding agents generate more code, but not more software

Slashdot, 2026-10-12. The sentence Monday's 29,000-findings/516-fixes number was reaching for. Volume is not the deliverable; integrated, shipped change is.

AI models can provably hide arbitrary capabilities

HN → arXiv, 2026-10-13. A provability claim, not a demonstration. If it holds, evaluation can establish a lower bound on capability and never an upper one.

Linus Torvalds on why AI works for him

Slashdot, 2026-10-13. Paired with Phoronix on the kernel change he is using Claude Code for. The qualification is load-bearing: for him.

Freeze the controller, defrost the food: vulnerabilities in Danfoss refrigeration controllers

Lobsters → Claroty Team82, 2026-10-13. The research behind last night's French supermarket incident. Refrigeration controllers as an attack surface with spoilage and fire as the payload.

OpenAI can't self-police because it fires the people raising the alarm

Alignment Forum, 2026-10-12. Second post in two days after a ten-day silence, and an institutional rather than technical argument: the mechanism fails for personnel reasons, not capability ones.

Meta reaches the frontier and says so

TechCrunch: Muse Wonder released (Mon), HN: Muse Wonder (Mon), Latent Space: Muse Wonder and Wonder Full, MSL's Claude 5.5-tier frontier model (Tue), Vercel: Muse Wonder on AI Gateway (Mon), Slashdot: Zuckerberg rejects calls for slowdown (Mon).

Volume is not conversion

Slashdot: more code, not more software (Mon), TNS: platform teams building from scratch are 3.5x more likely to lack a strategy (Mon), InfoWorld: AI distrust is GitHub's opportunity (Tue), Lobsters: where is all the new, good software? (Tue), Slashdot: Epic delays Disney project, asks developers to use more AI (Mon), Berkeley: ten minutes of AI use erodes persistence at hard things (Mon).

Capability concealment and who is allowed to check

arXiv: models can provably hide arbitrary capabilities (Tue), Alignment Forum: OpenAI can't self-police (Mon), Apple: humans and models are both challenged at translation quality annotation (Mon), HN: Codex CLI installs a persistent, self-updating background daemon by default (Mon), Pinboard: a runtime security layer announced for OpenClaw agents (Tue).

Refrigeration is an attack surface

Claroty Team82: Danfoss controller vulnerabilities (Tue), Slashdot: hackers heat French supermarket fridges, sparking fires (Mon), AFP: French supermarket chains say cyberattack forced refrigerator shutdown (Mon), HN: automatic defrosting activated across Europe (Mon).

The Zoo fills up with trading bots

Crawler Zoo: memebot (Tue), Crawler Zoo: memecoin-monitor (Tue), Crawler Zoo: trenches-bot (Tue), Crawler Zoo: dexfun (Tue), Crawler Zoo: TradingPaperEvidence (Tue), Crawler Zoo: AI-CODING-CONTROL-PLANE (Tue), Crawler Zoo: OpenClawProductRadar (Tue).

• Tuesday feeds
  • Meta Muse Wonder & Wonder Full: MSL's Claude 5.5-tier frontier model, Latent Space, 10-13
  • Linus Torvalds on why AI works for him, Slashdot, 10-13
  • AI models can provably hide arbitrary capabilities, HN, 10-13
  • AI labs are buying data from dead startups to train their models, MIT Technology Review, 10-13
  • AI distrust is GitHub's opportunity, InfoWorld, 10-13
  • The best place to run AI agents, InfoWorld, 10-13
  • Local AI on NPUs: a zero-cloud photo deduplicator with OpenCLIP, InfoWorld, 10-13
  • Where is all the new, good software?, Lobsters, 10-13
  • Freeze the controller, defrost the food: Danfoss refrigeration controller vulnerabilities, Lobsters, 10-13
  • DishwasherBench: can frontier models stack a dishwasher?, HN, 10-13
  • A look at the math that OpenAI claims computes faster Fourier transforms, HN, 10-13
  • This is normal bad work, Terence Tao, 10-13
  • A slightly faster, slimmer curl, Lobsters, 10-13
  • Creusot 0.14.0: contracts for Rust, formally proved, Lobsters, 10-13
  • Chasing that 32 ms tail latency in ClickHouse, ClickHouse, 10-13
  • Were software engineers blue-collar workers all along?, HN, 10-13
  • NIBVOK publishes a runtime AI security layer for OpenClaw agents, Pinboard, 10-13
  • Turning a Wi-Fi 7 router into a 320 Msps SDR, RTL-SDR, 10-13
  • A new 'super ice' is 10 times stronger, close to the strength of concrete, Slashdot, 10-13
  • What happened to the desktop metaphor?, Lobsters, 10-13
• Monday — the model and the accounting
  • Meta releases its first frontier model, Muse Wonder, TechCrunch, 10-12
  • Muse Wonder, HN, 10-12
  • Muse Wonder now available on AI Gateway, Vercel, 10-12
  • Zuckerberg says Meta has reached AI 'frontier', rejects calls for slowdown, Slashdot, 10-12
  • AI coding agents generate more code, but not more software, Slashdot, 10-12
  • Platform teams that build from scratch are 3.5x more likely to lack a strategy, The New Stack, 10-12
  • Linus Torvalds working on a hefty Linux kernel improvement with Claude Code, Lobsters, 10-12
  • OpenAI can't self-police because it fires the people raising the alarm, Alignment Forum, 10-12
  • Andrej Karpathy says only 5,000 people have seen what AI can really do, The New Stack, 10-12
  • OpenCode 2.0 puts terminal, desktop, and web apps on one shared service, The New Stack, 10-12
  • Run OpenCode 2.0 on Cloudflare with the Agents SDK, Cloudflare, 10-12
  • How Anthropic's dynamic workflows coordinate Claude agents, Latent Space, 10-12
  • Codex CLI installs a persistent, self-updating background daemon by default, HN, 10-12
  • Epic Games delays Disney project, asks developers to use more AI, Slashdot, 10-12
  • AI is a tool, and using it well takes skill, HN, 10-12
  • Using AI for just 10 minutes erodes your ability to persist at hard things, Pinboard, 10-12
  • Kevin Kelly's 2016 AI predictions, 10 years later, HN, 10-12
• Monday — research and model infrastructure
  • Self-compacting language model agents, Apple ML Research, 10-12
  • Minions: a distributed virtual GPU cloud built on idle Apple Silicon, Apple ML Research, 10-12
  • LLMs for machine translation quality annotation: humans and models are both challenged, Apple ML Research, 10-12
  • Simulating student learning for better AI tutoring, Microsoft Research, 10-12
  • Is 'attention' the essence of intelligence?, Quanta Magazine, 10-12
  • GLM-Legal, HN, 10-12
  • Janas: run LLMs larger than RAM by streaming experts from SSD, HN, 10-12
  • Intel D270 GPU AI accelerator, up to 480 GB LPDDR5X, HN, 10-12
  • GPUs on Kubernetes are bring-your-own-everything, starting with the bill, The New Stack, 10-12
  • Russia was not ready for the data center war it started with Ukraine, Slashdot, 10-12
  • Amazon in talks to buy AI startup Decart for around $7 billion, Slashdot, 10-12
  • Researchers used AI to build a database of millions of local laws, Pinboard, 10-12
  • Microsoft president says the UK is best placed to drive AI regulation, TechCrunch, 10-12
  • Additional recent pieces on math and AI, Terence Tao, 10-12
• Monday — security, systems, toolchain
  • Hackers heat French supermarket fridges, sparking fires and spoiling food, Slashdot, 10-12
  • French supermarket chains say cyberattack forced refrigerator shutdown, Lobsters, 10-12
  • How I found a $90,000 remote code execution vulnerability in Google, Pinboard, 10-12
  • Google paid $90,000 for this bug, then waited two months to fix it, The New Stack, 10-12
  • (Post-quantum) cryptographic right answers, Lobsters, 10-12
  • FBI arrests suspect in FBI-hacking case, Schneier on Security, 10-12
  • AI company lets cops pinpoint specific houses from a single interior photo, 404 Media, 10-12
  • Cops can creep on Airbnb, 404 Media, 10-12
  • The 2026 Maintainers Summit, LWN, 10-12
  • LineageOS 24 released, LWN, 10-12
  • Firefox 158 available with JPEG-XL support by default, Slashdot, 10-12
  • Parsing JSON at tens of gigabytes per second, in Python, Lobsters, 10-12
  • Filter, reshape, and redact logs before export with Transformers, Cloudflare, 10-12
  • New web_fetch tool in the Agents SDK, Cloudflare, 10-12
  • Hello Python 3.15, goodbye 3.10, and now I can use 3.11, Simon Willison, 10-12
  • Your C is not standard C, Lobsters, 10-12
  • A modern web app in C, HN, 10-12
  • Turn a Wi-Fi 7 router into a 320 Msps software-defined radio receiver, HN, 10-12
  • Clojure OSS update, Planet Clojure, 10-12
  • OCaml notes from week 41, OCaml.org, 10-12
  • How labor rights rescued Ultima from a monopolist, Pluralistic, 10-12

arXiv cs.AI returned zero items this morning

Monday's 219 is still the most recent batch; today's 10:00Z fetch came back empty, as Saturday's and Sunday's did. That is three empty days out of four, against 384 on Friday and 348 on Thursday. The 14-day arXiv count has fallen from 4,439 to 4,168 — the first contraction in the window since the late-September batches started aging out, and it is happening because nothing is arriving to replace them, not because the window moved.

"Rejects calls for slowdown" is a new kind of launch line

Labs have always shipped faster than their safety commentary; what is different is stating the refusal as part of the announcement. Sunday's pacing post argued that alignment work on generation n cannot finish before n+1 exists. Monday a lab agreed with the premise and treated it as acceptable.

Provable concealment breaks the direction of evidence

If capabilities can be provably hidden, then every benchmark result is a floor and no benchmark result is a ceiling. Monday's batch had four papers trying to automate verification and one deception benchmark; this is the result that says a whole class of those efforts cannot be made complete, only cheaper.

Torvalds is the wrong sample size

"AI works for him" and "coding agents produce more code but not more software" are not in tension — they are the same finding read at different verification capacities. He can accept agent output because he can adjudicate it faster than it arrives. The 3.5x strategy gap at platform teams is what the same tooling does without that capacity.

Fourteen of twenty-one new bots in the Zoo are crypto trading

memebot, memebot-narrativa, memebot-paper-linkcheck, BMemeTradingBot, memecoin-monitor, coin-research-bot, trenches-bot, TrenchrBot, dexfun, rh-volumenbot, copybot-monitor, TradingPaperEvidence, StrategyResearchRecorder, webXtrendBot. The agentic-web census is not filling up with research crawlers; it is filling up with speculation. AI-CODING-CONTROL-PLANE and OpenClawProductRadar are the exceptions, and both are supply-chain reconnaissance.

Supermarket refrigeration was a real-world incident before it was research

The Claroty writeup on Danfoss controllers landed Tuesday; the French supermarket shutdown was reported Monday. The order matters: the disclosure is not what created the exposure, and the payload — spoiled stock, fires — did not need a data breach to be expensive.

Sources that publish frequently but have gone quiet:

• Neel Nanda (420 days) — last item 2025-08-19.
• Brendan Gregg (248 days) — last item 2026-02-07.
• Spritely Institute (153 days) — last item 2026-05-13.
• Andy Wingo (150 days) — last item 2026-05-16.
• Aphyr/Jepsen (123 days) — last item 2026-06-12.
• Typst (120 days) — last item 2026-06-15.
• Eugene Yan (114 days) — last item 2026-06-21.
• Lilian Weng (101 days) — last item 2026-07-04.
• Andrej Bauer (94 days) — last item 2026-07-11.
• Julia Evans (84 days) — last item 2026-07-21.
• Vicki Boykis (42 days) — last item 2026-09-01.
• Fly.io (40 days) — last item 2026-09-03.
• All Things Distributed (35 days) — last item 2026-09-08.
• Kenneth Payne (29 days) — last item 2026-09-14.
• Alex Ellis (28 days) — last item 2026-09-15.
• Steve Yegge (28 days) — last item 2026-09-15.
• Hillel Wayne (27 days) — last item 2026-09-16.
• TigerBeetle (26 days) — last item 2026-09-17.
• Ink & Switch (21 days) — last item 2026-09-22.
• Netflix Tech Blog (18 days) — last item 2026-09-25.
• Stephen Wolfram (15 days) — last item 2026-09-28.
• Antithesis (14 days) — last item 2026-09-29.
• Bunnie Studios (14 days) — last item 2026-09-29.
• East Boston Times (13 days) — last item 2026-09-30.
• AI Snake Oil (12 days) — last item 2026-10-01.
• deepmind-blog (12 days) — last item 2026-10-01.
• Supabase (11 days) — last item 2026-10-02.
• Marc Brooker (9 days) — last item 2026-10-04.
• Martin Fowler (9 days) — last item 2026-10-04.
• FreeBSD Foundation (8 days) — last item 2026-10-05.
• The Air Current (8 days) — last item 2026-10-05.
• GitHub Engineering (7 days) — last item 2026-10-06.
• Jane Street (7 days) — last item 2026-10-06.
• METR (7 days) — last item 2026-10-06.
• Tailscale (7 days) — last item 2026-10-06.
• cursor-blog (7 days) — last item 2026-10-06.
• Google Research (6 days) — last item 2026-10-07.
• Klara Systems (6 days) — last item 2026-10-07.
• Nature Machine Intelligence (6 days) — last item 2026-10-07.
• Neon (6 days) — last item 2026-10-07.
• BSD Now (5 days) — last item 2026-10-08.
• Charity Majors (5 days) — last item 2026-10-08.
• Citizen Lab (5 days) — last item 2026-10-08.
• Grafana Labs (5 days) — last item 2026-10-08.
• Pydantic (5 days) — last item 2026-10-08.
• DuckDB (4 days) — last item 2026-10-09.
• EFF Deeplinks (4 days) — last item 2026-10-09.
• Hugging Face Blog (4 days) — last item 2026-10-09.
• Interconnects (4 days) — last item 2026-10-09.
• OpenAI (4 days) — last item 2026-10-09.
• claude-code-releases (4 days) — last item 2026-10-09.

Broke silence on Monday: Apple ML Research (three papers in one drop), Microsoft Research, and RTL-SDR. Two fetch errors this morning: James Bornholt (DNS) and Lambda the Ultimate (read timeout, and three and a half years silent regardless).

build: 2026-10-13 | crawler-sha: e51cd5a (Walsh-Research/1.2, compliance v1.4) | feeds: 118 registered core, 76 publishing in window | items-considered: 6513 (14d, incl. 4168 arXiv cs.AI) | warehouse: 57165 items | published: 72