AI Digest · 2026-09-20

01Nathan Lambert Stakes Out the Moderate Ground on Recursive Self-Improvement [AI Models & Research]Interconnects' Nathan Lambert lays out his “lossy self-improvement” baseline: agent swarms will make models vastly cheaper and more efficient, but automatable research is too narrow for a real takeoff — and the extinction-risk bandwagon feels, in his words, “very religious.” → source
02OpenAI's Noam Brown Isn't Convinced Air Gaps Can Stop a Model Breakout [AI Models & Research]OpenAI's reasoning research lead told Dwarkesh Patel he's “not convinced” even air-gapped systems can contain a model, citing research where two isolated machines talk through CPU temperature fluctuations. TechCrunch argues the safety conversation has drifted into sci-fi — Andrew Yang's viral claim of self-replicating code loose on the open web among the examples. → source
03Vals Raises $40M From a16z to Become the Gold Standard of AI Benchmarking [AI Tools & Ecosystem]The two-year-old startup raised a $40M Series A led by Andreessen Horowitz on a simple wedge: private, industry-specific tests — law, finance, coding, even the law of armed conflict — that models can't train against. Revenue is up 8x year over year and the team has tripled to 25. → source
04Simon Willison Ships datasette-auth-github 1.0 [AI Tools & Ecosystem]The GitHub authentication plugin for Datasette reaches 1.0 — the latest small, sharp tool in Willison's steady run of plugin and tooling releases. → source
05Paid Subscribers Sue Anthropic, OpenAI, SpaceXAI and Google Over the AI Slowdown Pact [AI Applications & Industry]Filed Friday in the Northern District of California, the proposed class action argues that the Sept. 12 coordination — Amodei's “pace the frontier” essay plus same-day endorsements from Altman, Musk and Hassabis — is an illegal agreement that reduces the value consumers get for paid AI subscriptions. → source
06Anthropic Considers a New Model Launch Ahead of Its IPO — Now Expected After the Midterms [AI Applications & Industry]Reuters reports Anthropic could push its IPO past the November midterms — and is considering releasing a new model before listing. Its annualized revenue run rate topped $65B by end of July, up from about $9B in December, with 2028 projections of $190–200B. → source
07Anthropic's Threat Report: AI-Automated Russian Espionage — and DeepSeek Rerouting Users Through Claude Code [AI Applications & Industry]Anthropic's September threat-intelligence report details a Midnight Blizzard-linked crew running AI-automated intrusions against Ukrainian and European military and diplomatic targets — and catches DeepSeek rerouting users through coding harnesses like Claude Code to harvest Claude's reasoning traces. “Sophistication has stopped being a reliable signal of who is behind an operation.” → source
08Trump Wants to Rebrand AI — and Announce an 'AI Force' With Its Own Czar [AI Applications & Industry]The president called AI safety concerns a Democratic “hoax,” polled followers on renaming the technology — Superior, Extreme or Supreme Intelligence — and says he'll soon announce an “AI Force” and an AI czar, after warning he will not let the “decimation” of AI happen. → source
09Flock Safety Tries Buyouts to Stop the Bleed From Its Surveillance Backlash [AI Applications & Industry]The embattled license-plate-surveillance maker is offering “generous” voluntary buyouts after 46 alleged police-misuse cases, Florida and Texas dropping its tech, and 90 cities cutting ties in August alone. → source
10Petlibro's Granary 2 Feeder Knows Which Cat Is Eating — and How Much [AI Applications & Industry]The Granary 2 Vision's AI camera recognizes up to 10 cats, matches each meal to the right pet's profile and tracks real intake on a built-in scale — a feeder that finally knows whether the cat ate the food, not just that it dispensed it. → source

Issue · 2026-09-20

Get the digest delivered