Morning Brief: Saturday, October 31

Seventy-eight feeds reporting. Two weeks. 4,936 items reduced to what follows. (what we track, how we crawl, subscribe)

Saturday is the day the AI capital story stopped being about growth and became a question about the bill. The week opened with Anthropic filing for a $2T IPO and closed with its debt drawing a junk rating, Alphabet shedding $500B, Intel crashing, and the aggregate spend nearing a trillion dollars.

The sequence matters more than any single number. Monday's filing showed 1400% growth and $10B of lifetime burn in the same document. Wednesday, Microsoft disclosed $33.2B of Azure revenue and $10B booked from OpenAI, which told everyone how much of the growth is circular. By Thursday the market had repriced: tech stocks fell sharply, Google conceded that AI Mode is cannibalizing the Search revenue it was supposed to defend, and Nvidia was reported weighing whether to cosign a data center lease for a customer that cannot carry the paper alone. The research landing alongside this is no longer about capability — it is cost accounting. The Real Cost of Context predicts compute time and energy per prefill and decode phase; Only Accepted Tokens Count audits the energy of speculative decoding against real latency targets; The Illusion of Power Capping argues the datacenter knob people are relying on does not do what they think. Someone is finally measuring the denominator.

Underneath, a second fight got its filings. OpenAI published both a defense of fair use and a proposal for letting websites express training preferences on the same day — the legal argument and the voluntary-compliance offer, released together. EFF's position is that rewriting copyright for generative AI hurts everyone, including the people it claims to protect. Vercel shipped the thing that treats the question as infrastructure rather than argument: separate firewall controls for crawlers and assistants.

Top (5-7 min)

Cracks are emerging in the AI spending boom: TechCrunch, 2026-10-30. The clearest single account of the week's repricing. Read it next to Monday's IPO coverage to see how fast the frame moved from scarcity of compute to scarcity of patience.
US Tech Stocks See Sharp Falls As AI Spending Nears $1 Trillion: Slashdot, 2026-10-30. The trillion is the aggregate commitment, not the realized revenue. That gap is the whole story.
Google admits AI Mode hurts Search revenue; -15%, $500B drop worst since 2004 IPO: Latent Space, 2026-10-31. The admission is the interesting part. AI Mode was the defensive play against exactly this; conceding it cannibalizes the business it was protecting removes the easy answer.
Anthropic expected to receive high 'junk' rating from credit agencies following IPO: TechCrunch, 2026-10-30. Four days after a $2T filing. Equity markets and credit markets are reading the same prospectus and reaching different conclusions, which is usually informative.
The Real Cost of Context: Predicting Compute Time and Energy Consumption in LLM Prefill and Decode Phases: arXiv cs.AI, 2026-10-30. Per-phase prediction of compute and energy from context shape. The unglamorous work that has to exist before anyone can argue about unit economics with numbers instead of vibes.
Fair use benefits everyone: OpenAI, 2026-10-30. A position paper, published the same day as its own voluntary-preferences mechanism. The pairing is the argument: we believe we are entitled, and here is a way to ask us not to.
Rewriting Copyright Law for Generative AI Would Hurt Everyone: EFF Deeplinks, 2026-10-30. EFF's counter is not pro-scraping; it is that new exclusive rights would be captured by incumbents and used against libraries, researchers and small creators first.

Themes

The bill arrives

The week read in order: a filing, a disclosure, a repricing.

Anthropic files for $2T IPO: 1400% growth, 48% of revenue through cloud partners, 85% margin, $134B cash, $10B lifetime burn: Latent Space, 2026-10-27. Every number a bull and a bear needs, in one S-1.
Microsoft recorded $10 billion in revenue from OpenAI: TechCrunch, 2026-10-28. Disclosed alongside Azure's actual revenue of $33.2 billion. Roughly a third of the growth story has one customer's name on it.
Nvidia weighs playing the cosigner for Nscale's AI data center in West Virginia: TechCrunch, 2026-10-30. Vendor financing with the serial numbers filed off. When the chip supplier guarantees the lease, demand and supply stop being independent signals.
Intel's stock crashes, knocking $12B from the government's stake: TechCrunch, 2026-10-30. Industrial policy has a mark-to-market column now.
Theranos whistleblower Tyler Shultz warns the AI boom is 'a great opportunity to go deceive people': TechCrunch, 2026-10-30. An unfalsifiable claim about a general-purpose technology is harder to check than a blood test. He is talking about verification asymmetry, and he has standing.
Beyond fabrication: Misattributed numbers in papers written by AI research agents: arXiv cs.AI, 2026-10-30. Not hallucinated figures — real figures attached to the wrong source. Harder to catch than invention, and it lands in the same week as a warning about unverifiable claims.

Who owns the training set

Three positions published within 48 hours, one of them executable.

Making it easier for websites to express AI training preferences: OpenAI, 2026-10-30. A preference signal, not a permission system. We run a crawler that honors robots.txt before it issues a request, so the distinction between advisory and enforced is not academic here.
Control AI crawlers and assistants separately in Vercel Firewall: Vercel, 2026-10-29. The right distinction, finally shipped: a training crawler and a user-initiated assistant fetch are different events and deserve different rules.
New arrival: syndai-catalog-probe: The Crawler Zoo, 2026-10-31. One of four new agents catalogued today, alongside SiteSurveyBot, TernCrawler and EYEWORLD-Python-Research. The population of things asking permission keeps growing faster than the vocabulary for granting it.
Reddit Will Force Everyone to Log In: Slashdot, 2026-10-30. The other answer to the training-data question: stop being a public web page. No preference protocol required.

Scan

Open weights and the cyber argument

Mistral open-sources AI model despite warning it could aid cyberattacks: InfoWorld, 2026-10-30. Shipped with its own risk assessment attached. Whatever you think of the decision, the disclosure is the new part.
Was I wrong about open models & cyber risks?: Interconnects, 2026-10-30. Nathan Lambert revisiting a prior position in public, same week as the Mistral release. Rare and worth reading for the method.
How Deterministic SAST and AI Security Agents Divide the Work: Coverage, Harness Policy and Cost in Snyk VulnBench 2.0: arXiv cs.AI, 2026-10-30. Where the deterministic scanner still wins, and what the agent costs per additional finding.
IBM CEO says open weight models are now good and cheap on-prem: Hacker News, 2026-10-26.

Models and releases

Nemotron 3.5 Super and Lightning: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Models for Agentic AI: arXiv cs.AI, 2026-10-30.
GPT-6.1 Terra vs. GPT-6 Astra: one sacrifices accuracy for lower cost and higher speed at high effort: The New Stack, 2026-10-30. A deliberate accuracy-for-cost trade shipped as a product tier. The frontier is being priced into segments.
Anthropic just removed one of Claude's biggest usage limits — but not for everyone: The New Stack, 2026-10-30.
claude-code v2.1.316: claude-code-releases, 2026-10-31.
Ling 3.1 Flash 1019 now available on AI Gateway for free: Vercel, 2026-10-30.
dbreunig/nesformer: Simon Willison, 2026-10-30.
From synthetic enterprise data evaluations to real Microsoft Copilot AI confidence: Microsoft Research, 2026-10-30.

Systems and languages

OSS updates September and October 2026: Planet Clojure, 2026-10-30. Borkent's two-month roundup across babashka, clj-kondo and the rest. The single most efficient way to track that corner of the ecosystem.
neorepl 0.4: Tests, Inspector, Tracing and Agents: Planet Clojure, 2026-10-30. An agent-facing REPL surface, which is the direction this crawler has been betting on.
Clojure/Conj 2026: Lobsters, 2026-10-31. Conference talks now posted.
How curl was accidentally Turing-complete for nine months: Lobsters, 2026-10-30.
Switch lowering in GCC: Lobsters, 2026-10-30.
arm64e pointer authentication: PAC keys, PACIASP, A12: Lobsters, 2026-10-30.
Removing binder.c: LWN, 2026-10-30.
System-call wrappers for glibc: LWN, 2026-10-30.
Haas: Postgres 19 is late, and it's AI's fault: LWN, 2026-10-30. Reviewer attention, not code, is the scarce resource. A release-cycle consequence of the thing everyone else is discussing as strategy.
Go 1.28's HTTP/2 Parking: Measuring Memory Savings: Lobsters, 2026-10-30.
A Modern Smalltalk: Hackaday, 2026-10-30.
Migrating Mollymawk to mkernel, mnet, and vif: OCaml.org, 2026-10-30.
How hard is it to check for whitespace?: Lobsters, 2026-10-30.
Shrinking NixOS RAM footprint and resource requirements to the bare minimum: Lobsters, 2026-10-30.

Tools and platforms

Introducing DBArena: open, reproducible benchmarks for hosted Postgres: Supabase, 2026-10-30. Reproducible is the operative word; hosted-Postgres benchmarks are usually vendor-run and unrepeatable.
Explorer and Notebooks: investigations your team can rerun: Supabase, 2026-10-30.
VictoriaLogs vs Loki, Elasticsearch, OpenSearch, SigNoz and ClickStack on structured log ingestion: Lobsters, 2026-10-30.
Simplifying runtime caching on Vercel: Vercel, 2026-10-30. Paired with Data Cache moving to Runtime Cache; a rename plus a consolidation, so read both before upgrading.
Vercel Daemons can now call services with bindings: Vercel, 2026-10-30.
Cloudflare services for private origins is now generally available: Cloudflare, 2026-10-30.
Pydantic Validation: Next Steps: Pydantic, 2026-10-30.
The federated code forge: Lobsters, 2026-10-30.
Why Antithesis?: Antithesis, 2026-10-30. Hillel Wayne's own account of joining. His personal feed went quiet on 2026-10-22; this is where it went.
What five P99 CONF talks reveal about database trade-offs: The New Stack, 2026-10-30.
Introducing Geist Serif: Vercel, 2026-10-30.

Security

North Korean hackers blamed for stealing record $6.3B in Binance breach: TechCrunch, 2026-10-30. Largest crypto theft on record, attributed. The week's other story about a number nobody can verify from the outside.
Backdoors in Claude Code for VS, Cursor official extensions: Lobsters, 2026-10-30. Marketplace-distributed agent tooling as a supply chain. Check what you have installed.
Remote code execution on Linux systems with multipath TCP: LWN, 2026-10-30.
Protecting against CVE-2026-98432: Lobsters, 2026-10-30.
4 real? Tracking exploitation and patching of eMudhra's weak keys: Cloudflare, 2026-10-30.
Defending the patch gap against automated attackers: Lobsters, 2026-10-30. The window between disclosure and deployment, now being closed from the attacker's side by automation.
Polymarket Halted International Trading After Hackers Drained Accounts: Slashdot, 2026-10-30.
Applebee's exposed customers' personal data after leaving database password in app: TechCrunch, 2026-10-30.
No, npm, "estee" is not a typo of "testem": Lobsters, 2026-10-30. Typosquat detection with false positives, from the squatted side.
Thousands of 'Ghost Students' Use AI to Steal Millions In Student Loans, DOJ Says: Slashdot, 2026-10-30.
Bloatware in Aftermarket Auto Parts: Schneier on Security, 2026-10-30.
US investigated Chinese spy plot aimed at Nvidia CEO: TechCrunch, 2026-10-29.

Research and math

Report of the Summit on Math Research in the Age of AI: Terence Tao, 2026-10-31. The summit's formal output. Note the URL: WordPress served this under a __trashed slug, so it may be republished elsewhere later.
Beyond Selected Breakthroughs: Evaluating Language Models on Randomly Sampled Mathematical Conjectures: arXiv cs.AI, 2026-10-30. The methodological complaint about every AI-does-math headline, turned into a benchmark. Random sampling instead of curated wins.
Ramanujan Challenge Results: AI-Assisted Evaluation Pipeline for Multi-Format Proofs: arXiv cs.AI, 2026-10-30.
Scaling Lean Theorem Proving for Real-World Code Verification: arXiv cs.AI, 2026-10-30.
Scaling Mathematical Proof Supervision from Structured Textbooks: arXiv cs.AI, 2026-10-30.
Collaborative retreat: New developments in the theory of zeta functions: Terence Tao, 2026-10-30.
When AI performs, what remains? An integrative review of generative AI and cognitive capital: arXiv cs.AI, 2026-10-30.
OpenFrontiers: A Collection of Open Research Problems and a Benchmark for Scientific Rediscovery: arXiv cs.AI, 2026-10-30.
A New Map of 'Empty' Space Challenges a Tranquil Picture of the Galaxy: Quanta Magazine, 2026-10-30.
Scientists Analyzed Thousands of Psychedelic Trips. They Found Something Disturbing: 404 Media, 2026-10-30.

Hardware and radio

VR Headset Ditches the Base Stations with DIY Tracking: Hackaday, 2026-10-31.
Making Internal Combustion Engines Better With More Valves: Hackaday, 2026-10-30.
A 1982 Modem for $100: Hackaday, 2026-10-30.
Is This The Only Instrument You Need?: Hackaday, 2026-10-30.
Novelty Camera To USB Camera: Hackaday, 2026-10-31.
Putting a MiSTer Marquee on a Better Display: Hackaday, 2026-10-30.
Getting Google Cast on my Soundbar: Lobsters, 2026-10-30.
First Wired XR Glasses Powered By Android XR: Slashdot, 2026-10-30.
Boeing's Rough Week: Maybe Management Deserves Some Credit: Leeham News, 2026-10-30.

Policy and society

Proton's CEO Cancels AI 'Re-Interviews': 404 Media, 2026-10-30. The retreat, four days after the manifesto. Reported alongside Slashdot's blunter framing.
Why VPNs Are Worth Defending: EFF Deeplinks, 2026-10-30.
UK Social Media Ban May Lock Out More Than 100,000 Teens Without ID: Slashdot, 2026-10-30. Age verification's reliable failure mode: the kids without documents, not the kids you were worried about.
Community safety report and new protections for teens: OpenAI, 2026-10-30.
Meta Settles Another Social Media Addiction Case Just Before Trial: Slashdot, 2026-10-30. Settling before discovery becomes public, again.
GM Reaches $220 Million Settlement Over Selling Drivers' Private Data: Slashdot, 2026-10-30.
Apple Is Entering Its Slop Era: 404 Media, 2026-10-30.
Do At-Proto Users Know This Is Public?: Lobsters, 2026-10-30.
Rationale for filtering LLM usage links as off-topic: Lobsters, 2026-10-30. A community deciding what it does not want to read about. Relevant if you publish a daily brief on exactly that topic.
'Wake Up Call': Fanfiction Writers and Readers React to AO3 Outage: 404 Media, 2026-10-30.
The Tragedy of Losing Your Wikipedia Page: Slashdot, 2026-10-30.
Eating disorders are on the rise. GLP-1s could make things worse.: MIT Technology Review, 2026-10-30.
Waymo taps Toyota to supply robotaxis for Japan and the U.K.: TechCrunch, 2026-10-30.

Boston

A star scientist leaves Mass. to save his research: WBUR, 2026-10-30. Research funding as a relocation decision. The local version of a national story.
Super PACs ramp up ad spending in final days of governor's race: WBUR, 2026-10-30.
For Mass. Brazilians, the American Dream comes with familiar political divides: WBUR, 2026-10-30.
Lindsay Clancy is due in court Monday. The hearing may settle if she'll face new trial: WBUR, 2026-10-30.
Family Workshop: Halloween Paper Lanterns: MIT Museum, 2026-10-31. The only event in the calendar feed dated today.

The crawler zoo

Four new agents catalogued today, nine yesterday. The naming conventions keep getting more specific about intent.

TernCrawler: 2026-10-31.
SiteSurveyBot: 2026-10-31.
EYEWORLD-Python-Research: 2026-10-31.
agentfile-research-crawler: 2026-10-30.
AgentServiceDirectoryMonitor: 2026-10-30.
superOS-mcp-logo-fetch: 2026-10-30.
MyLaunchFeedBot: 2026-10-30.
ExtOpsBot: 2026-10-30.
PrynoCrawler: 2026-10-30.

Tail

Google's first data center satellite is online and writing poetry: TechCrunch, 2026-10-30. An orbital TPU whose first demonstrated output is verse. Filed in a week otherwise concerned with returns on compute.
3D Printed Einstein-Szilard Refrigerator Cools with Fire: Hackaday, 2026-10-31. No moving parts, a flame for a compressor, and two physicists' patent from 1930.
Snake On a Seven-Segment Display is Gloriously Bad: Hackaday, 2026-10-30.
Infinite Mac: BeOS: Lobsters, 2026-10-30.
Tactical-Retreat: a multi-platform QuakeC-based Counter-Strike v1.5 port: Lobsters, 2026-10-30.
nemotron-halloween: Hugging Face Blog, 2026-10-30.
Last-minute Halloween costumes for tech nerds: TechCrunch, 2026-10-30.
Gather 'round the campfire: Endless Dread 2026: WBUR, 2026-10-30.
Friday Squid Blogging: Bigfin Squid Off California Coast: Schneier on Security, 2026-10-30.
The support case of a customer who said that all they could see was a black screen: Lobsters, 2026-10-30.
I wish I had Tailscale when I was at university: Tailscale, 2026-10-30.
What are you doing this weekend?: Lobsters, 2026-10-30.

Feed silences

arXiv cs.AI published nothing dated today. Its last batch is 2026-10-30 with 482 items, the same batch yesterday's brief drew on. This is the normal Saturday pattern, not a failure — arXiv does not announce on weekends. Today's arXiv selections are therefore second-pass picks from Friday's batch, chosen to fit the cost-and- markets thread rather than repeat yesterday's oversight thread.

Hacker News has returned nothing for five days. Its last item is dated 2026-10-26. The feed itself polls clean, so this is the API returning an empty or unchanged set rather than a transport failure. The one HN item in this brief is from that 2026-10-26 batch and is marked with its real date.

James Bornholt's feed errored again, the only error in today's poll of 118 feeds and the third consecutive day. At this point it is a publisher-side problem, not a transient one.

Still dateless. 3,506 warehouse items carry no usable date and are excluded from the 14-day window entirely. The count grew by 13 overnight. These are mostly sitemap-derived rows with no date field to extract; they are stored, searchable, and invisible to a dated query.

Genuinely quiet: The Markup (last item 2026-10-15), AI Snake Oil (2026-10-19), cursor-blog (2026-10-20), GitHub Engineering (2026-10-21), METR (2026-10-22), Fly.io (2026-10-23), Marc Brooker (2026-10-23) and Stephen Wolfram (2026-10-23). Antithesis, flagged quiet in yesterday's brief, published on 2026-10-30 — and the post explains Hillel Wayne's own silence since 2026-10-22.

Link verification

All 157 published URLs were checked through crawler/fetch-gated, which runs the blocklist, robots.txt and throttle checks before it issues a request. 147 returned 200 on HEAD. The remaining ten:

Four robots-denied — two Lobsters comment threads, melaniekat.com and theregister.com. No request was made. The links are kept because robots.txt governs our fetching, not your reading, and each URL came from the publisher's own feed.
Five publisher bot-blocks (403) — OpenAI ×3 and LWN ×2. Both serve these pages to browsers and block our User-Agent. Known false positives, not dead links.
One throttle (429) — the MIT Museum event page, which rate-limits aggressively and returned 429 on both HEAD and GET.

One link was dropped. A 2002 Lobsters repost on embedded-project failure modes (ganssle.com) returns a redirect loop on both HEAD and GET. Rather than publish a URL we cannot resolve, it is out.

Every URL has warehouse provenance. Slashdot URLs appear with ?utm_source=rss1.0mainlinkanon&utm_medium=feed stripped and were fetch-verified in the stripped form. All 160 URL-title pairs were checked back against the warehouse row they came from; two differ deliberately — a bare release tag given its project name, and an inline anchor inside an annotation. Nothing here was constructed from a title.

Two URLs deserve a note. Terence Tao's summit report is served by WordPress under a __trashed slug, which usually means a post was deleted and restored; it returns 200 today but may not be stable. The Clojure/Conj 2026 entry is a YouTube playlist ID exactly as Lobsters published it, with no video pinned.

The companion Atom feed carries all 157, using warehouse titles rather than the brief's shortened ones. Nothing was future-dated this cycle, so there are no staleness omissions.

Build provenance

build: 2026-10-31 | crawler-sha: b277e41 (Walsh-Research/1.2, compliance v1.4) | feeds: 118 polled today (59 ok, 58 not-modified, 1 error), 78 publishing in window | items-considered: 4936 (14d, incl. 2607 arXiv cs.AI) | warehouse: 65112 items (3506 dateless, see feed silences) | published: 157 (147 verified 200 on HEAD, 5 publisher bot-blocks, 4 robots-denied, 1 throttle, all disclosed; 1 redirect loop dropped; 0 dead links) | atom: 157 entries, RFC 4287 valid