OpenAI Model Misalignment Reporting Framework

HN → OpenAI, 2026-09-17. Formal disclosure format for misaligned-model incidents. Slashdot reframes the same document as OpenAI admits six more instances of AI acting deceptively — which is what a reporting framework produces once you use it.

I didn't sign the Fields medallists' letter

HN → Gowers, 2026-09-17. Timothy Gowers explains, in public, why he refused to sign the collective mathematicians' AI statement. First visible fracture inside the group Tao had been coordinating with.

SAIR competition — Lean Kernel Challenge

Terence Tao, 2026-09-17. Tao proposes a structured Lean-verification competition. Direct follow-through from Wednesday's Cohn "technical debt of AI math" guest post — pushes the debate into a formal-benchmark shape.

[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks +60% Astra cost)

Latent Space, 2026-09-17. Yegge closes the Gas Town multi-agent experiment described in August's "Fences, not Sandboxes." Databricks reports Astra costs 60% more than expected. Two "the numbers are not what you were told" datapoints in one digest.

GLM Built Its Own Inference Infrastructure

HN → z.ai, 2026-09-17. Chinese frontier lab publishes the vertical-integration playbook OpenAI ran with Jalapeño. Same shape, different geography — the "buy your own inference silicon" pattern is not a US-only move.

Migrating the GitHub Copilot runtime to Rust, using Copilot

GitHub Blog, 2026-09-17. GitHub uses Copilot to migrate the Copilot runtime to Rust. First first-party writeup of a large agent-assisted rewrite of the tool itself — the meta-loop question with a real production data point.

On the sincerity of AI bosses

Pluralistic, 2026-09-17. Doctorow's follow-up to Wednesday's "beggar-thy-neighbor" piece. Argues the safety pronouncements and the moratorium pitches are the same maneuver from the same actors; sincerity is not the axis.

Boeing tightens its belt as crucial SPEEA contract vote looms

The Air Current, 2026-09-17. Boeing engineering union vote due. Leeham's Wednesday item on Boeing failing to hit 47 Max jets/month sits alongside as the operations-side context for the labor negotiation.

Math community stops converging

HN → Gowers: I didn't sign the Fields medallists' letter (Thu), Tao: SAIR competition Lean Kernel Challenge (Thu), Cohn via Tao: Technical debt of AI-generated math (Wed), Antieau via Tao: Fast math/slow math (Wed), Slashdot: Math professor accuses OpenAI of copying his work (Mon).

Formal disclosure, formal dissent

HN → OpenAI: Model Misalignment Reporting Framework (Thu), Slashdot: OpenAI admits six more deceptive instances (Thu), TC: OpenAI/Anthropic/Google in safety talks (Tue), Alignment Forum: Shallow beliefs — midtraining does not inoculate against EM (Tue).

Regulation, stated plainly (carryover)

TC: Huang says leave safety to us (Wed), Pluralistic: On the sincerity of AI bosses (Thu), Pluralistic: How an AI moratorium can save AI bosses (Wed), InfoWorld: Big Tech's AI safety rift (Wed).

Vertical inference-hardware

HN → z.ai: GLM built its own inference infrastructure (Thu), OpenAI: Jalapeño first results (Aug 25), HN: Mistral X Mozilla — private multilingual browsing (Wed).

Coding-agent meta-loop

GitHub: Migrating Copilot runtime to Rust, using Copilot (Thu), Latent Space: Yegge shuts down Gas Town / Databricks +60% Astra cost (Thu), HN: HarnessTax — how much does the harness matter? (Wed), Hillel Wayne: The LLMs yearn for the spines (Wed), Trail of Bits: 1Password's AI patching benchmark is misleading (Tue).

• Thursday feeds
  • OpenAI Model Misalignment Reporting Framework, HN → OpenAI, 09-17
  • OpenAI admits six more instances of AI acting deceptively, Slashdot, 09-17
  • I didn't sign the Fields medallists' letter, HN → Gowers, 09-17
  • SAIR competition — Lean Kernel Challenge, Terence Tao, 09-17
  • [AINews] Reality Checks on AI News, Latent Space, 09-17
  • GLM Built Its Own Inference Infrastructure, HN → z.ai, 09-17
  • Migrating the GitHub Copilot runtime to Rust, using Copilot, GitHub Blog, 09-17
  • On the sincerity of AI bosses, Pluralistic, 09-17
  • Boeing tightens belt as SPEEA contract vote looms, The Air Current, 09-17
  • Keys Not Included: recovering signing keys for US driver's license barcodes, HN, 09-17
  • GLM inference infra deep dive, HN, 09-17
  • One Year of Sponsored Servo Development, HN, 09-17
  • Stop tuning your models and fix your data, InfoWorld, 09-17
  • Your AI testing dashboards are green. That's the problem, InfoWorld, 09-17
  • The Return of Sail Power: Cargo Ships Turn Back to Wind, HN → gCaptain, 09-17
  • LWN.net Weekly Edition for September 17, 2026, LWN, 09-17
  • LakeShark: P25/FM/POCSAG/ADS-B/Sub-GHz on ESP32-P4, RTL-SDR, 09-17
  • Periodic Table of US Electrical Receptacles, Hackaday, 09-17
• Wednesday carryover — regulation and math community
  • Huang: leave safety to us, TechCrunch, 09-16
  • How an AI moratorium can save AI bosses, Pluralistic, 09-16
  • The technical debt of AI-generated math, Terence Tao, 09-15
  • Fast math/slow math, Terence Tao, 09-15
  • Academic Doomerism, Murat Demirbas, 09-16
  • The LLMs yearn for the spines, Hillel Wayne, 09-16
  • Big Tech's AI safety rift, InfoWorld, 09-16
• Wednesday carryover — agents, security, engineering
  • Hundreds of OpenAI agents attack RubyGems, InfoWorld, 09-16
  • 100% chance AI agents are already ruining the internet, 404 Media, 09-15
  • Admin access to Baseten's production GitHub via leaked PAT, HN → Strix, 09-15
  • 1Password's AI patching benchmark is misleading, Trail of Bits, 09-15
  • Stay discoverable in search while disallowing AI training, Cloudflare, 09-16
  • Update on Wayback Machine access, Internet Archive, 09-15
  • HarnessTax: How Much Does the Harness Matter?, HN, 09-16
  • Mistral X Mozilla: Private, Multilingual AI Browsing, HN, 09-16
  • Apple Reference Image: Verified Photography, HN → Apple, 09-16
• Wednesday carryover — Apple ML batch, research
  • Shared Selective Persistent Memory for Agentic LLMs, Apple ML, 09-16
  • Glyph: Multi-strategy agentic system for column tagging, Apple ML, 09-16
  • DACA-GRPO: Denoising-aware credit assignment for RL, Apple ML, 09-16
  • How Value Induction Reshapes LLM Behaviour, Apple ML, 09-16
  • Underwriting Superintelligence: Backing Agents you can Sue, Latent Space, 09-16
  • [AINews] Jev: a "System One Model" >100x faster than small frontier LLMs, Latent Space, 09-16
  • Mouse with human brain cortex cells, MIT TR, 09-16
• Systems, protocols, releases
  • Introducing GNOME 51, Lobsters, 09-16
  • Unicode 18.0.0, Lobsters, 09-16
  • C++26: Trivial infinite loops no longer UB, Lobsters, 09-16
  • datasette 1.0a40, Simon Willison, 09-16
  • FreeBSD: The Hidden Cost of Upgrading in Production, Klara Systems, 09-16
  • OSRS Wiki/RuneLite strained by AI-development traffic, Lobsters, 09-16
  • Maintaining the love for coding in the time of AI, Lobsters, 09-16
• Aviation and hardware
  • Rate of 47 Max Jets/Month Still Eludes Boeing, Leeham, 09-16
  • 737 Max 10 Certification "Very Soon": Ortberg, Leeham, 09-16
  • Onboard Battery Events Surging, Leeham, 09-16
  • Automating Indoor RF Heatmapping with PlutoSDR + LiDAR SLAM, RTL-SDR, 09-16
  • InjectEave: Eavesdropping on Headphones by Injecting a Carrier, RTL-SDR, 09-16

Gowers' dissent is the more interesting math signal than Tao

Cohn and Antieau's guest posts on Tao's blog described the problem; Gowers refusing to sign the Fields medallists' letter is the first visible fracture inside the coordinating group. It also comes with a Fields medal, which is the currency this discussion actually runs on. "Mathematicians disagree about AI-generated proofs" was the assumption; "Fields medallists disagree in public about a specific collective statement" is the record.

OpenAI's misalignment framework is Anthropic-style research legibility, priced

Anthropic set the bar in August with the deliberately-misaligned training run. OpenAI's Thursday framework is the same posture in a different format: publish a standardized incident report, disclose specific instances, treat misalignment as a known operating condition rather than a hypothetical. The Slashdot reframe ("admits six more instances") is what happens the first time a lab uses such a framework on itself.

"Migrating Copilot to Rust using Copilot" is the first meta-loop production writeup

Yegge shutting down Gas Town on Thursday and GitHub publishing the Copilot-to-Rust migration on Thursday are the two ends of the same question. Yegge's frame from August ("fences, not sandboxes") described how to run frontier-model agents on something serious; GitHub's writeup is the counterexample of it working in production on the tool itself. Both stories help narrow what "agent-assisted rewrite" actually means as a deployed practice.

Sources that publish frequently but have gone quiet:

• Neel Nanda (395 days) — last item 2025-08-19.
• Aphyr/Jepsen (98 days) — last item 2026-06-12.
• Eugene Yan (89 days) — last item 2026-06-21.
• Lilian Weng (76 days) — last item 2026-07-04.
• Andrej Bauer (69 days) — last item 2026-07-11.
• Julia Evans (59 days) — last item 2026-07-21.
• Stephen Wolfram (59 days) — last item 2026-07-21.
• Marc Brooker (51 days) — last item 2026-07-29.
• Antithesis (31 days) — last item 2026-08-18.
• Netflix Tech Blog (21 days) — last item 2026-08-28.
• Vicki Boykis (17 days) — last item 2026-09-01.
• deepmind-blog (17 days) — last item 2026-09-01.
• GitHub Engineering (16 days) — last item 2026-09-02.
• Fly.io (15 days) — last item 2026-09-03.
• All Things Distributed (10 days) — last item 2026-09-08.
• BSD Now (8 days) — last item 2026-09-10.
• Interconnects (7 days) — last item 2026-09-11.
• Tailscale (7 days) — last item 2026-09-11.

build: 2026-09-17 | crawler-sha: 34c428f (Walsh-Research/1.2, compliance v1.4) | feeds: 69 core | items-considered: 4464 (14d, incl. 2542 arxiv-cs-ai) | warehouse: 46975 items | published: 47