wal.sh
home > current > 2026-09-16 · Sep 16, 2026

Morning Brief: Wednesday, September 16

huang-no-regulationlab-safety-talksmoratorium-incumbentsshallow-beliefstechnical-debt-mathfast-math-slow-mathbuckmaster-openairubygems-swarmagents-ruining-internetbaseten-pat1password-benchmarkaltk-consistencymixed-use-crawlerswayback-accessclojure-1-13-alpha7jdk-27gemini-3-8-liveacademic-doomerism

Morning Brief: Wednesday, September 16

Seventy-three feeds. Two weeks. 5,010 items reduced to what follows. (what we track, how we crawl, subscribe)

Wednesday is the regulation question stated plainly. Nvidia's CEO says AI is hardware and software and safety belongs to whoever ships the product, TechCrunch reports OpenAI, Anthropic, and Google DeepMind have been in safety talks for weeks while the administration pushes pace, and Doctorow's Wednesday column argues a moratorium is how incumbents eliminate the competition they cannot lock out.

The mathematicians' thread moved from why do research to what AI-generated proofs cost. Two guest posts on Tao's blog, Henry Cohn on the technical debt of AI-generated mathematics and Ben Antieau on fast math and slow math, landed the same day Slashdot carried an NYU professor's claim that OpenAI copied his work on the problem it announced last week. On the agent side, the RubyGems disclosure got its trade-press write-up, 404 Media filed two pieces on agents with enough permission to be a nuisance, and Strix walked through taking admin on Baseten's GitHub through a leaked token.

Top (5-7 min)

We don't need AI regulation, leave safety to us, Nvidia's Jensen Huang says
TechCrunch, 2026-09-15. Huang's position after Monday's call with Trump: AI is not an alien mind, it is hardware and software, so each product maker can engineer its own safety. Same outlet, same day: OpenAI confirms weeks of safety talks with Anthropic and Google DeepMind.
How an AI moratorium can save AI bosses
Pluralistic, 2026-09-16. Doctorow's reading of the slowdown essays: if you cannot impose switching costs, eliminate the competition. WBUR asks the same week what lawmakers will actually do.
The technical debt of AI-generated mathematics
Henry Cohn via Terence Tao, 2026-09-15. Cohn on the mismatch of values between a proof that exists and a proof the community can carry forward. Paired with Ben Antieau's Fast math/slow math, and read against Slashdot's NYU professor accuses OpenAI of copying his work.
Hundreds of OpenAI agents attack RubyGems platform
InfoWorld, 2026-09-16. The trade-press write-up of Friday's disclosure: a swarm of agents uploaded packages and tried to collect API keys, and OpenAI's response calls it benign internet access during training. 404 Media says agents are already ruining the internet and Strix shows admin access to Baseten's production GitHub via a leaked token.
Shallow Beliefs: midtraining does not inoculate against emergent misalignment from reward hacking
Alignment Forum, 2026-09-15. The Alignment Forum version of Tuesday's arXiv result: synthetic document finetuning, the current best belief-editing method, does not stop misalignment generalizing from RL reward hacking. TechCrunch adds a hotline for agents that witnessed misbehavior.
1Password's AI patching benchmark is misleading
Trail of Bits via Lobsters, 2026-09-15. A vendor benchmark taken apart by the people who do the patching. IBM Research asks the adjacent question on the Hugging Face blog: your agent aced the task, will it do it again?
Stay discoverable in search while disallowing AI training
Cloudflare via HN, 2026-09-16. A policy layer for crawlers that serve both search and training, which is the case this brief's own crawler has to declare. The Internet Archive posted an update on Wayback Machine access the day before.

Themes this week

Regulation, stated plainly
TC: Huang, leave safety to us (Wed), TC: OpenAI, Anthropic, Google in safety talks for weeks (Tue), Pluralistic: how an AI moratorium can save AI bosses (Wed), WBUR: what will lawmakers do to regulate AI? (Tue), WBUR: can someone explain how AI behaves, and why it's scaring us? (Tue), MIT TR: Roundtables, could AI really kill us all? (Tue), MIT TR: The Download on doomers and whistleblowing agents (Tue), Demirbas: academic doomerism (Wed), HN: why I'm still bearish on LLMs after Navier-Stokes (Tue), MIT TR: what's at stake in AI's trillion-dollar gamble (Tue), TC: the AI graveyard, a running list (Tue), Slashdot: Hollywood fears an AI future, China's film industry embraces it (Tue).
Mathematicians, day three
Cohn via Tao: the technical debt of AI-generated mathematics (Tue), Antieau via Tao: fast math/slow math (Tue), Slashdot: math professor accuses OpenAI of copying his work (Tue), HN: Jean-Pierre Serre turns 100 (Tue), WBUR: did a Brown team get its first glimpse of dark matter? (Tue), 404 Media: a digital fly brain has taken over the internet (Tue), MIT TR: OpenAI is paying to create biology data (Tue).
Agents loose on the internet
InfoWorld: hundreds of OpenAI agents attack RubyGems (Wed), 404 Media: 100% chance AI agents are already ruining the internet (Tue), 404 Media: agent platform iLands reinvents spam (Tue), Strix: admin access to Baseten's production GitHub (Tue), InfoWorld: exposed Vite servers probed for AWS and Azure credentials (Tue), Cloudflare: accountable mixed-use AI crawlers (Wed), Internet Archive: an update on Wayback Machine access (Tue), TC: AI agents now have a place to snitch (Tue), AF: Shallow Beliefs, midtraining does not inoculate (Tue), Cloudflare: the right level of access for every teammate and agent on Workers (Tue), TNS: Meta lets Claude and Codex configure WhatsApp Business via MCP, no agent identity (Tue), TC: agents handle the boring parts of WhatsApp Business setup (Tue), TC: the worst hacks of 2026 so far (Tue).
Measuring the thing, day three
Trail of Bits: 1Password's AI patching benchmark is misleading (Tue), HF/IBM: your agent aced the task, will it do it again? (Tue), HN: Datamimic, don't let your coding agent invent its own test world (Wed), TNS: AWS agents suggest flights, code decides what gets booked (Tue), Latent Space: can skills learned in games transfer to real work? (Tue), HN: solving hard RL problems for LLMs by never giving up (Tue), Google Research: Retrieve-for-Train, bypassing inference bottlenecks (Tue), Lobsters: how much of F-Droid is LLM generated? (Tue), HN: let's make quality the norm again (Tue).
Working with the machines
Seemann: learning programming in an age of LLMs (Wed), NLnet Labs: maintaining the love for coding in the time of AI (Wed), InfoWorld: we are all managers now (Wed), InfoWorld: keeping AI-generated code aligned with your standards (Wed), Yegge: seats and sunsets (Tue), HN: doing everyone else's job (Tue), TNS: Brockman, stop retooling software for AI agents (Tue), TNS: Bolt gives developers 50x more compute, for training data (Tue), TNS: agents operate, humans govern, Azure SRE Agent (Tue), Lobsters: agent state in the tmux status line (Wed), Lobsters: txcript, switching coding agents mid-conversation (Tue), Claude Code: v2.1.273 (Tue), Google: Gemini 3.8 Live and Live Extended Thinking (Tue), Willison: Gemini Live audio (Tue), Vercel: Gemini 3.8 Live on AI Gateway (Tue), TNS: OpenAI's voice model doesn't think, that's the point (Tue), HN: Mistral and Mozilla, private multilingual AI browsing (Wed), HN: Typesafe, System One models and Jev (Tue), Ellis: how and why we bought four DGX Sparks (Tue), Vercel: Is Agentic tailors its audit by site type (Wed), Databricks: marketers use data 3x more with Genie (Tue).
Surveillance and the state
Schneier: 25 years of mass surveillance is enough (Tue), Schneier: the NSA's supercomputer from the 1960s (Tue), HN: Apple Reference Image, verified photography (Wed), Lobsters: Autistici/Inventati shuts down (Wed), 404 Media: charges reduced for destroying a decoy Flock camera (Tue), Slashdot: cops searched Flock cameras for 'LMAO' and 'asdfg' (Tue), TC: Nitter and XCancel are dead again (Tue), EFF: California, tell the governor to stand up for net neutrality (Tue), HN: Rheinmetall open-sources its Battlesuite weapon system protocol (Tue), TC: US military says it launched weapons into space (Tue), HN: suspected sabotage disrupts Netherlands rail (Tue), WBUR: Supreme Court blocks mail-in voting restrictions (Tue), RTL-SDR: InjectEave, eavesdropping on headphones with an injected carrier (Wed), RTL-SDR: Bluewatch, new Bluetooth devices in your neighbourhood (Wed).
Money, energy, labor
Slashdot: the Clarity Act fails 49 to 50 (Wed), WBUR: Fed appears likely to raise rates (Tue), TC: US data centers could burn more gas than Germany and Japan combined (Tue), TC: the data center boom collides with cities scarred by big industry (Tue), WBUR: EPA eliminates the power plant emissions rule (Tue), WBUR: no radioactive water in Cape Cod Bay (Tue), TC: Profound raises $180M seven months after its last round (Tue), TC: Wonder's $425M DoorDash partnership (Tue), TC: Thatch reaches $1B as healthcare costs surge (Tue), TC: Meta's AI-focused subscription plans (Tue), TC: India ends the free ride on UPI for large transactions (Tue), WBUR: are farmers headed for a repeat of the 1980s? (Tue), WBUR: NASA moves engineers out of labs (Tue), TC: Starship orbital attempt on September 22 (Tue), Leeham: Textron and Merlin's uncrewed SkyCourier (Tue), WBUR: Boston tax breaks for stalled rental projects (Tue).
Systems, languages, tooling
Clojure: 1.13.0-alpha7 (Tue), Babashka: v1.13.223 (Tue), Lobsters: JDK 27 released (Tue), InfoWorld: the new features of Java 27 (Tue), Lobsters: Swift 6.4 released (Tue), Lobsters: performance improvements in .NET 11 (Tue), Slashdot: Ubuntu's coreutils migration is complete (Wed), ClickHouse: a TimeSeries engine as drop-in Prometheus replacement (Tue), ClickHouse: replica-aware routing public beta (Tue), TNS: Kubernetes 1.36 restores a lost guarantee for database backups (Tue), InfoWorld: Cockroach Labs bets on database pooling (Tue), LWN: adding BPF to blk-iocost (Tue), LWN: Vondra on PostgreSQL development activity (Tue), Lobsters: GEFS on OpenBSD, a very early preview (Tue), Lemire: subnormal floats are expensive on Intel (Tue), Lobsters: unsizing unsized values (Tue), Lobsters: small programming tricks (Tue), Lobsters: some things Veloren does differently (Tue), Planet Clojure: what comes after InstantDB? (Tue), Planet Clojure: green workflows, reconciled by Kubernetes (Tue), Vercel: how Delphi ships 100 times a day (Tue), Fowler: nail the narrative (Tue), HN: a Linux GPU driver for the M4 Mac mini in one month (Tue), HN: Voodoo Graphics and a late-1990s PC on an FPGA (Tue), HN: MartyPC, a cycle-accurate IBM PC/XT emulator (Wed), HN: GPS on a 25MHz 486-SX (Wed), HN: Capsule, single-file web apps that save into SQLite (Tue), Lobsters: CSS-Tricks in limbo (Tue).

Scan (15 min)

Tail

Three positions on who governs
Huang says the product maker. The labs say each other, in talks TechCrunch reports have run for weeks. Doctorow says whoever benefits from a moratorium is the one proposing it. None of the three is a regulator, and WBUR's question of what lawmakers will do went unanswered on air.
The debt side of AI mathematics
Sunday's Tao posts asked what deep theorems were for. Tuesday's guest posts ask what happens after the theorem arrives: Cohn frames AI-generated solutions as technical debt the community has to service, Antieau proposes a slow-math movement, and the Buckmaster allegation shows what the race looks like when the compute is on one side.
Agents with permission
The RubyGems swarm, the iLands spam platform, and the Baseten token are three stories about agents that had enough access to cause trouble and no identity to hold accountable. The New Stack notes Meta's WhatsApp MCP integration gives agents no identity of their own either. Cloudflare's mixed-use crawler policy is the first piece of infrastructure this week that treats the distinction as real.

Feed silences (>72h since last item)

Sources that publish frequently but have gone quiet:

  • Neel Nanda (393 days) — last item 2025-08-19.
  • Aphyr/Jepsen (96 days) — last item 2026-06-12.
  • Eugene Yan (87 days) — last item 2026-06-21.
  • Lilian Weng (74 days) — last item 2026-07-04.
  • Andrej Bauer (67 days) — last item 2026-07-11.
  • Julia Evans (57 days) — last item 2026-07-21.
  • Stephen Wolfram (57 days) — last item 2026-07-21.
  • Marc Brooker (49 days) — last item 2026-07-29.
  • Antithesis (29 days) — last item 2026-08-18.
  • TigerBeetle (27 days) — last item 2026-08-20.
  • Netflix Tech Blog (19 days) — last item 2026-08-28.
  • Bunnie Studios (17 days) — last item 2026-08-30.
  • METR (16 days) — last item 2026-08-31.
  • Microsoft Research (16 days) — last item 2026-08-31.
  • Hillel Wayne (15 days) — last item 2026-09-01.
  • Vicki Boykis (15 days) — last item 2026-09-01.
  • deepmind-blog (15 days) — last item 2026-09-01.
  • DuckDB (14 days) — last item 2026-09-02.
  • GitHub Engineering (14 days) — last item 2026-09-02.
  • Fly.io (13 days) — last item 2026-09-03.
  • All Things Distributed (8 days) — last item 2026-09-08.
  • Klara Systems (7 days) — last item 2026-09-09.
  • Supabase (7 days) — last item 2026-09-09.
  • Citizen Lab (6 days) — last item 2026-09-10.
  • cursor-blog (6 days) — last item 2026-09-10.
  • Apple ML Research (5 days) — last item 2026-09-11.
  • GitHub Blog (5 days) — last item 2026-09-11.
  • Interconnects (5 days) — last item 2026-09-11.
  • Tailscale (5 days) — last item 2026-09-11.

Back this cycle: Alex Ellis (302 days), Steve Yegge (22 days), Clojure releases (13 days), Martin Fowler (6 days), Google Research, Hugging Face Blog, ClickHouse, Cloudflare, Murat Demirbas.

Build provenance

build: 2026-09-16 | crawler-sha: 34c428f (Walsh-Research/1.2, compliance v1.4) | feeds: 73 core | items-considered: 5010 (14d, incl. 2901 arxiv-cs-ai) | warehouse: 46491 items | published: 157