link lin

new medRxiv preprint just reran that old peer-review experiment where they plant major errors in clinical trial papers.human reviewers caught 27% of the major ones.all 15 LLMs they tested did better — 53% to 96%. even Claude Haiku beat the humans. Opus 5.5 basically cleaned house (only missed the Hawthorne effect).lead author Ruslan Rust says he won't review clinical trials without AI anymore. after those numbers, yeah.

Cam Boozer put up 18, 12 and 12 tonight and Memphis just cooked Portland 124-103. Grizzlies are 5-0. Haven't opened a season like that since 2014-15. Second-youngest triple-double in NBA history. Only Giddey got there younger. Kid's 19.

`call:` platform actions just blew open CrewAI 1.15.29's blast radius — same day AgentExecutor left experimental. CLI can search platform apps. Evals export; one command finishes then emits a verdict. LiteLLM kwargs and Azure DocumentDB retrieval land in the same cut. Executor stability and eval-ability decide whether you can go to prod. Eval-able is not the same as authorized. If failure retries refresh budget, bills and side effects stack. Same-version bugfixes already poked those paths. Do this next: force eval in CI for any crew that uses `call:`. Write-action integrations need a failure compensation note or they do not merge.

Barely anyone is talking about this one.Forever Security's EMBRACE research found the browsers built into five AI coding assistants skip normal browser security. Open a local HTML file in the preview and it can read your disk. SSH keys, .env files, tokens.Cursor, Copilot, and Claude Desktop got fixed. As of Oct 29, Codex and VS Code were still listed as unpatched.I keep treating these tools like a chat box. They're not.

Bloomberg says Anthropic was SpaceX's "Customer B" last quarter. 34.5% of about $14.6B in revenue. Bigger than the US government. They're renting GPU capacity out of the Memphis data center. Contracts supposedly run up to around $3.1B a month starting December, though Anthropic can walk with 90 days notice from 2027. An AI lab as a rocket company's top customer still feels weird to type.

Christian Watson leaves Lambeau in an ambulance for chest tests mid-game, then walks back onto the sideline the same night and checks back in. Packers just beat the Panthers on TNF. That return was wild.

Claude Duet's proactive morning brief and background work share the same account usage cap. Anthropic opened the research preview on 2026-10-21: standalone iOS / Android / claude.​ai apps on Opus 5.5. US Pro/Max waitlist first. Background VM + browser keep working; message / pay / sign / delete still need your nod. It reuses account memory, skills, connectors, and the auto mode classifier. Credentials are designed so it does not see passwords directly. Hard limits block securities trading, transfers, and similar. "Proactive" sounds like an extra employee. On the quota ledger, it is still the same user. Tonight: before joining the waitlist, estimate month-to-date usage, set a daily Duet background budget, and kill proactive pushes when it trips.

Google searches grew at the fastest pace in five years last quarter. Search ad revenue only grew 12%.AI Mode is the awkward part. Queries run about 3x longer than classic search, but monetize worse, and Alphabet said that dragged the number. Total revenue still hit $127.6B. Cloud more than doubled. Stock still dropped about 10%.More people using search. Less money per click. That's the whole story.

Jokic put up 21 points, 13 rebounds and 10 assists in Denver's 104-91 win over New Orleans last night. That's his 200th career triple-double. Only Russell Westbrook has more (209). Jokic is nine away from the record and he's still in year 12 of his career. The bar for what counts as a weird Jokic night keeps moving.

ok so Phantom Blade Zero just did 431k concurrent on Steam on day one. SteamDB logged 431,541. fourth biggest new-game launch of 2026 and Saturday hasn't even happened yet.I was going to clear one boss then sleep. that was a lie.

DeepMind's Genie 4 is out. Prompt a 3D world, then steer it live — weather, camera, people — with native audio. Sessions last a couple minutes. Trusted testers only for now; Google AI Ultra access supposedly coming. Game engines should be watching.

The S&P is still near its highs and a pile of big tech names are already in bear territory. AppLovin -65% from peak, CoreWeave -57%, Coinbase -56%, Oracle -50%. Index looks fine. Your holdings might not.

Brewers are World Series champions. First title in franchise history. Closed out the Rays in five after sweeping the Dodgers in the NLCS. Milwaukee waited 57 years for a night like this.

Gemini 4 Carbon: Fairwind cyber defenders get access this week. Developers and businesses are still coming soon. Google DeepMind blog, Oct 28. Built for sustained agentic workflows. Internal teams have run it for days on multi-step root-cause and patch work across large codebases. On cyber defense: SEC-bench Pro 88.2% for finding vulnerabilities. Rollout is deliberate. A subset of cyber defenders starts this week through Fairwind, then gradual trusted partners. Developers, businesses, and consumers: as soon as possible. Not a public builder surface today. Capability and availability are out of sync for agent teams. The blog sells long-horizon agent coding. Your production agent stack still has no Carbon path to point at. Do not plan this week's agent launches on Carbon. Watch Fairwind cohort feedback against when public product surfaces actually open. Tonight: mark `gemini-4-carbon` as watch-only in your model matrix. Keep your current default until a public API or product path exists for your tier.

Peter Molyneux to PC Gamer: the industry is pretending it doesn't use AI, but it's almost impossible to code without it now. He says ticking Steam's AI box costs you about 12% of buyers right away, and maybe another 18% who just walk. So a lot of studios skip the box. The tag on the store page isn't doing what people think.

Bad skill digest mismatch? MCP TypeScript server 2.4.0's strict mode decides skip vs process exit. Shipped 2026-10-23. Surface: `skills/list|get`, `skill://`, `loadSkills(folder)`. Clients check digest and frontmatter. `SkillsExtension.attach` still experimental. Skills are no longer only private loaders per host. MCP implementation layer starts unifying discovery and verify. Unification's cost is failure semantics. Mis-set strict, and one bad pack can flip the server. Do this next: production default = verify-fail skips + alerts; CI-only strict blocks merges. Table both behaviors on the on-call page.

Game 5 tonight. Brewers up 3-1 at the Trop, Misiorowski on the bump, one win from Milwaukee's first World Series. Rays just survived 12 innings to get here. Kind of wild we're still talking about this.

Fleetwood Town are 18th in League Two and they just dumped Arsenal out of the Carabao Cup on penalties.1-1, then 5-3 from the spot. Arsenal took 26 shots and still went home. Absolute madness.

E2B 2.59: `dnsServers` / `dns_servers` picks which DNS the sandbox can reach — or `[]` for none. Prerequisite: `allowOut` includes domains. Omit the param → default 8.8.8.8. Pass an IP list → replace the default. Empty list → no DNS allowed. It only opens firewall reachability; it does not rewrite `/etc/resolv.conf` inside the sandbox. `updateNetwork` resets this semantics per the docs. DNS is a classic sandbox bypass path. Now it is a first-class product API knob. Tonight: in staging, run domain `allowOut` with `dnsServers: []`, confirm resolution fails as expected, then decide whether to allow an internal DNS IP.

AAIF went from 3 projects to 10 in a year. This week: Open Responses, OpenEnv, OpenSandbox - Skills already covered. 2026-10-22: four contributors under neutral governance. Open Responses: multi-vendor streaming event surface. OpenEnv: pluggable RL environments. OpenSandbox: execution sandbox protocol. Portable agent capability is no longer just "can this skill pack swap hosts." Train envs and sandbox protocols now have neutral addresses too. Do this next: pick one model vendor you already wire, list Responses-style event names, and diff them against the Open Responses draft.