OpenAI withdraws three mathematical results

HN → @danintheory, 2026-10-08. The first subtraction from the catalogue, announced by the producer rather than found by a referee. Three entries out of 722 is small; the direction of travel is the news.

OpenAI's math solutions aren't meeting the field's standards yet

TechCrunch, 2026-10-08. The general version of the withdrawals. "Standards" here means write-up, citation, and completeness conventions — the things a referee checks before checking the mathematics.

What should we tell our students?

HN → Terence Tao, 2026-10-09. Tao moves from governance to pedagogy in a day. The question is not rhetorical: graduate programs are three to seven years long and advisors are being asked now.

OpenAI, the Partition Principle, and Mathematics

HN → Asher Kach, 2026-10-08. A set theorist works through one catalogue entry in detail. This is the closest thing to verification published this week, and it comes from a blog rather than a journal.

OpenAI Just Carpet-Bombed Mathematics

Pinboard, 2026-10-09. The framing that will stick. Note the metaphor is about volume and indiscriminacy, not about correctness — which is the actual complaint the field has been making all week.

We're putting too much faith in AI's ability to say no

MIT Technology Review, 2026-10-09. Refusal is trained behavior measured on benchmarks, deployed as though it were an access control. The gap between those two is where this year's incidents have been living.

MXC — a sandboxed code execution system

HN → Microsoft, 2026-10-09. The engineering answer to the previous item, and it arrived the same day: don't ask the model to decline, give it a boundary it cannot cross.

The catalogue starts to lose entries

HN → @danintheory: OpenAI withdraws three mathematical results (Thu), TechCrunch: not meeting the field's standards yet (Thu), Tao: what should we tell our students? (Fri), Kach: OpenAI, the Partition Principle, and Mathematics (Thu), The Atlantic: OpenAI just carpet-bombed mathematics (Fri), AHM: statement on the October 6 release (Thu), Tao: "Math 2.0" and holistic valuation (Thu), Slashdot: hundreds more results on a field already in shock (Thu).

Containment replaces refusal as the control surface

MIT Technology Review: too much faith in AI's ability to say no (Fri), HN → Microsoft: MXC, a sandboxed code execution system (Fri), InfoWorld: AWS takes aim at runaway agent behavior with Strands Box (Thu), TNS: what an agent inherits when you hand it your credentials (Thu), TNS: your agent just provisioned a resource — who owns it? (Thu), TechCrunch: Goodfire's "inside-out" monitors for rogue agents (Thu), Anthropic: 2026 usage policy update (Thu), arXiv: verification and self-improvement in agentic AI, foundations and limits (Fri).

Agent operations becomes an ops discipline

HN → OpenTelemetry: OTel-native by design (Fri), TNS: AI changed incident response, your tabletops haven't caught up (Thu), TNS: most AI incident tools still leave the key decision to humans (Thu), TNS: the API tax, why agents stall without infrastructure context (Thu), Grafana: collector fleet management (Thu), InfoWorld: pipeline observability, stop monitoring and start preventing (Thu).

"Open" is the contested word

TNS: don't use "open weight" and "open source" interchangeably — Percona's CEO (Fri), InfoWorld: neoclouds and the enterprises that need them (Fri), HN → OpenRouter: Step 5 Preview, a 1M-context MoE from StepFun (Thu), HN: why isn't the industry freaking out about DeepSeek 4.1 Flash? (Thu), HF: the model that didn't exist, so you made it yourself (Thu), TNS: Copilot goes local, but Microsoft won't say what still goes to the cloud (Thu).

• Friday feeds
  • What should we tell our students?, HN → Terence Tao, 10-09
  • OpenAI just carpet-bombed mathematics, Pinboard → The Atlantic, 10-09
  • We're putting too much faith in AI's ability to say no, MIT Technology Review, 10-09
  • MXC — a sandboxed code execution system, HN, 10-09
  • Don't use "open weight" and "open source" interchangeably, The New Stack, 10-09
  • OTel-native by design: products that export to any observability stack, HN, 10-09
  • Neoclouds and the enterprises that need them, InfoWorld, 10-09
  • Roundtables: a conversation with the creator of AI-designed viruses, MIT Technology Review, 10-09
  • Sophos cuts threat investigation time by 96% with OpenAI Daybreak, OpenAI, 10-09
  • ttok 1.0, Simon Willison, 10-09
  • Apollo software pioneer Margaret Hamilton dies at 90, Slashdot, 10-09
  • First thorium nuclear clocks begin to tick, Slashdot, 10-09
  • en-html.org is a scam site, Lobsters, 10-09
  • Keyboard differences between Windows and Macs, HN, 10-09
  • Show HN: Quake ported to safe Rust, playable in browser, HN, 10-09
  • Job titles of the future: delivery drone air traffic controller, MIT Technology Review, 10-09
  • Belt file from old mixer and junk bin parts, Hackaday, 10-09
  • New arrival: SQLSpider, The Crawler Zoo, 10-09
  • New arrival: UrlQueryCollector, The Crawler Zoo, 10-09
  • New arrival: HNBot, The Crawler Zoo, 10-09
  • PAI Seminar: demand flexibility in EV charging, Harvard SEAS, 10-09
  • MIT.nano seminar — magnetics in MEMS: from DC to GHz, MIT calendar, 10-09
• Thursday carryover — mathematics
  • OpenAI withdraws three mathematical results, HN, 10-08
  • OpenAI's math solutions aren't meeting the field's standards yet, TechCrunch, 10-08
  • OpenAI, the Partition Principle, and mathematics, HN, 10-08
  • "Math 2.0" will need to value mathematical progress more holistically, HN, 10-08
  • AHM statement on OpenAI's October 6 release, Terence Tao, 10-08
• Thursday carryover — agents, containment, security
  • AWS takes aim at runaway AI agent behavior with Strands Box, InfoWorld, 10-08
  • The audit log says my name: what an agent inherits when you hand it your credentials, The New Stack, 10-08
  • Your AI agent just provisioned a resource. Who owns it?, The New Stack, 10-08
  • Goodfire's "inside-out" monitors catch rogue AI agents at a fraction of the cost, TechCrunch, 10-08
  • 2026 usage policy update, Anthropic, 10-08
  • Introducing the Anthropic Cyber Mission, Anthropic, 10-08
  • Disrupting AI-enabled "false front" operations, OpenAI, 10-08
  • The secrets of a US spyware king, Citizen Lab, 10-08
  • How technology empowers — and imperils — dictators, Schneier, 10-08
  • 64-day certificate lifetimes coming February 2027, Lobsters, 10-08
  • Asos confirms breach after hackers send rogue app notification, TechCrunch, 10-08
  • Watch footage of the Flock search a judge ruled unconstitutional, 404 Media, 10-08
  • When no ID means no internet: age verification and the right to access information, EFF, 10-08
• Thursday carryover — platform, infra, languages
  • AI changed how you respond to incidents. Your tabletops haven't caught up., The New Stack, 10-08
  • Most AI incident tools still leave a key decision to humans, The New Stack, 10-08
  • The API tax: why AI agents stall without infrastructure context, The New Stack, 10-08
  • Building scalable, agent-friendly APIs for AI applications, The New Stack, 10-08
  • You can outgrow vanilla Postgres without abandoning Postgres, The New Stack, 10-08
  • Lakebase and agentic SDLC: branching databases for coding agents, Databricks, 10-08
  • Introducing JWT authentication in ClickHouse Cloud, ClickHouse, 10-08
  • B-trees are back: engineering fast and pageable node layouts, Lobsters, 10-08
  • The performance cost of RwLock in our read-heavy workload, Lobsters, 10-08
  • An update on Rust's project goals, LWN, 10-08
  • The vx-swap showdown, LWN, 10-08
  • Zostera, a new implementation of WireGuard in OCaml, OCaml.org, 10-08
  • Pydantic v2.14, Pydantic, 10-08
  • Bevy 0.20, HN, 10-08
  • Extending Guix, Lobsters, 10-08
  • Normalizing trajectory models, Apple ML Research, 10-08
  • Bridging technical depth and usability: the story behind Radar's redesign, Cloudflare, 10-08
  • The history of the Hetzner Cloud network stack, Lobsters, 10-08
  • Whistle: speech to text in 16.9 MB, HN, 10-08
• Thursday carryover — industry and labor
  • OpenAI's revenue is reportedly $20 billion less than previously projected, TechCrunch, 10-08
  • Fired OpenAI safety researchers dispute misconduct claims, warn of chilling effect, TechCrunch, 10-08
  • AI leaderboard Arena nearly doubles valuation to $3.1B in 10 months, TechCrunch, 10-08
  • Google brings agentic AI to Gemini, starting with businesses, TechCrunch, 10-08
  • Waymo locks in $5B loan from Blackstone and PIMCO, TechCrunch, 10-08
  • US bars Microsoft, Adobe, and major IT firms from green card program, TechCrunch, 10-08
  • New York alleges TikTok gave teens a placebo safety feature, TechCrunch, 10-08
  • Female AI agents "paid" less than males, study finds, Slashdot, 10-08
  • If your team is happy, are you doing a good job?, Charity Majors, 10-08
  • Synthesis superintelligence: from semiconductors to superconductors, Latent Space, 10-08
  • AI breakthroughs in robotics won't change your life any time soon, MIT Technology Review, 10-08
  • AI-ready biological data: $1.8B global commitment, HN, 10-08

The first retraction is the story, not the volume

Three withdrawn results is a rounding error against 722, and it is still the most informative event of the week. It establishes that the catalogue has an error rate, that the producer will act on it, and that the acting happens without external referees — which is precisely the arrangement the Association for Human Mathematics objected to on Thursday. The count of published independent verification attempts remains zero; the closest thing is a set theorist's blog post.

"What should we tell our students?" is the load-bearing question

Tao went from valuation mechanics on Thursday to advising on Friday, and advising is where the irreversibility lives. A PhD is three to seven years. Anyone starting one this fall is choosing a problem under an assumption about machine capability that nobody in the field is currently willing to state.

Refusal was never a security boundary

MIT Technology Review's argument and Microsoft's MXC release landed on the same day without reference to each other, which is the useful part. One says trained refusal is being deployed as access control; the other ships a sandbox. Add AWS's Strands Box and three New Stack pieces on credential inheritance and the convergence is unambiguous: the control surface is moving out of the model.

The Crawler Zoo rotated verticals in one day

Thursday's eighteen arrivals were crypto-named — solana-token-radar, PumpPilotBot-UtilityResearch, memebot-research-crawler. Friday's ten are generic infrastructure: SQLSpider, UrlQueryCollector, HNBot, scraperbot9000, gvfs. The "research" suffix that was doing compliance work on Thursday is gone by Friday. Whatever drove the crypto wave was a cohort, not a trend.

Margaret Hamilton, 1936-2026

MIT News carried the death on Tuesday; Slashdot's item arrived Friday. She named software engineering and built the Apollo guidance code whose priority-display logic kept Apollo 11 from aborting its own landing — error handling as a design discipline, a decade before the field had the words for it.

Sources that publish frequently but have gone quiet:

• Neel Nanda (416 days) — last item 2025-08-19.
• Brendan Gregg (244 days) — last item 2026-02-07.
• Spritely Institute (149 days) — last item 2026-05-13.
• Andy Wingo (146 days) — last item 2026-05-16.
• Aphyr/Jepsen (119 days) — last item 2026-06-12.
• Typst (116 days) — last item 2026-06-15.
• Eugene Yan (110 days) — last item 2026-06-21.
• Lilian Weng (97 days) — last item 2026-07-04.
• Andrej Bauer (90 days) — last item 2026-07-11.
• Julia Evans (80 days) — last item 2026-07-21.
• Vicki Boykis (38 days) — last item 2026-09-01.
• Fly.io (36 days) — last item 2026-09-03.
• All Things Distributed (31 days) — last item 2026-09-08.
• Kenneth Payne (25 days) — last item 2026-09-14.
• Alex Ellis (24 days) — last item 2026-09-15.
• Steve Yegge (24 days) — last item 2026-09-15.
• Hillel Wayne (23 days) — last item 2026-09-16.
• The Markup (23 days) — last item 2026-09-16.
• TigerBeetle (22 days) — last item 2026-09-17.
• Ink & Switch (17 days) — last item 2026-09-22.
• Murat Demirbas (16 days) — last item 2026-09-23.
• Netflix Tech Blog (14 days) — last item 2026-09-25.
• Stephen Wolfram (11 days) — last item 2026-09-28.
• Antithesis (10 days) — last item 2026-09-29.
• Bunnie Studios (10 days) — last item 2026-09-29.
• AI Snake Oil (8 days) — last item 2026-10-01.
• Alignment Forum (8 days) — last item 2026-10-01.
• deepmind-blog (8 days) — last item 2026-10-01.
• Supabase (7 days) — last item 2026-10-02.
• Marc Brooker (5 days) — last item 2026-10-04.
• Martin Fowler (5 days) — last item 2026-10-04.

No source broke a silence longer than three days on Friday.

build: 2026-10-09 | crawler-sha: e51cd5a (Walsh-Research/1.2, compliance v1.4) | feeds: 75 core | items-considered: 6856 (14d, incl. 4521 arxiv-cs-ai) | warehouse: 56220 items | published: 79