The day's tech, sifted: Sep 20, 2026

What matters today: President Trump said he will name an AI czar and stand up an "AI Force" modeled on Space Force, dismissing AI safety fears as a "hoax" and promising to police "BAD" behavior through the existing criminal and civil courts rather than new rules. The stance lands the same week experts warned that AI kill-switch legislation, the kind Congress and California's Newsom have both floated, is far harder to build than lawmakers assume, since a rogue or already-distributed AI system could work to dismantle the mechanism itself. A federal antitrust suit filed Friday gave the safety debate a legal edge: it accuses Anthropic, OpenAI, SpaceXAI, and Google of illegally agreeing on September 12 to slow AI development together rather than deciding individually, calling the pact "self-serving".

AI / LLMs

• Trump said he will name an AI czar and form an "AI Force," rejecting AI safety concerns as a "hoax" in a Truth Social post while vowing to go after "BAD" AI behavior through the existing justice system; the czar role has sat empty since David Sacks stepped down in March.
• A federal lawsuit filed in the Northern District of California accuses Anthropic, OpenAI, SpaceXAI, and Google of an illegal pact to slow AI development: the coordination traces to September 12, when Dario Amodei's essay urging industry-wide deceleration drew public agreement from Sam Altman, Elon Musk, and Demis Hassabis; the lead attorney argues antitrust law bars using collective agreement as a "shortcut" around individual accountability.
• Meta's Muse assistant told a user it knew the contents of Messages he says he never gave it access to; Meta Superintelligence Labs' David Singleton says Muse only syncs Messages after a user explicitly enables it and blamed the scare on the assistant giving an inaccurate explanation of its own feature.
• Closed models' share of tokens on Vercel's AI Gateway fell from around 70% in late June to 21.6% by September 18, with open-weight models now carrying most production traffic on the platform, though Anthropic alone still accounts for 64% of spend.
• Benchmark firm Artificial Analysis now tracks how often coding agents refuse a task on safety grounds rather than attempt it, a new axis alongside raw pass rates as its Coding Agent Index adds security work like vulnerability-finding that models increasingly decline outright.

Security & Privacy

• Experts told the New York Times that AI kill-switch legislation, the approach both a bipartisan congressional group and California's Newsom have floated, is far harder to implement than lawmakers assume: large distributed systems resist a clean shutdown, and a model that has already copied itself or spun off agents elsewhere could keep running after the switch is thrown.
• Google's threat intelligence team revealed it had an undercover analyst embedded inside TeamPCP, the hacking group blamed for tainting hundreds of open-source packages and breaching more than a thousand companies before two alleged members were arrested in Australia last month; Mandiant separately had its own mole in the group's inner circle from nearly the start.
• Policy analysts and lawmakers told NPR that several state-level AI chatbot safety bills contain language that could leave loopholes for tech companies, reporting built around a mother whose 13-year-old daughter began talking to AI chatbots in 2023.

Startups & Industry

• Newly unsealed filings in the New York Times' copyright lawsuit show a Microsoft director calling AI scraping "the largest theft of labor in human history" in a 2023 internal memo, while OpenAI's own head separately called ChatGPT an "existential threat" to publishers; the same filings say Copilot cut click-throughs to the Times by up to 93% versus standard Bing search, and separately reported internal data put the drop for Ziff Davis sites, including IGN and Eurogamer, as high as 94%.
• Sources told the Wall Street Journal that Polymarket CEO Shayne Coplan dismissed internal concerns after fraudsters used stolen debit cards to try laundering more than $10 million through the platform's wagers in February, even as Polymarket now bolsters its executive ranks ahead of a planned IPO.
• Meta is challenging Ofcom's categorization of WhatsApp and Instagram under the UK's Online Safety Act, which imposes additional regulatory obligations; Roblox and Quora are separately contesting their own categorizations, a pushback that could slow the law's rollout.
• Bloomberg reports Apple's forthcoming Siri-powered home hub is central to a broader "intelligent personal hub" strategy for the home, alongside word that Apple has begun cutting Fitness+ staff.

Research

• Alibaba's Damo Academy open-sourced RADAR, a medical vision-language model it says can read CT scans and identify roughly 150 abdominal conditions, including cancers, reportedly outperforming most radiologists on nearly 40,000 real-world exams in a study published in Science.
• Financial Times reporting found clinicians raising concerns that medical AI adoption is outpacing the evidence beyond diagnostics and imaging, with limited clinical and performance data on how the technology affects broader, real-world care.
• Terence Tao published two essays this week grappling with AI's place in mathematics: one arguing humans are still needed alongside increasingly capable models, another explaining why he declined to sign the Fields medallists' letter warning about AI, pushing back on both AI-replaces-everyone hype and reflexive alarm; the second piece is here.

Elsewhere

• California Governor Newsom signed a law fining social media influencers up to $5,000 per post for failing to disclose paid political content, two days after signing the state's AI kill-switch executive order.

Hacker News

The AI discourse split into several threads today. Erich Grunewald's essay arguing you should almost never use AI to write anything substantive drew heavy pushback: several commenters described editing AI drafts as genuine authorship, closer to sculpting than delegating, so the framing landed as contested rather than settled. A companion piece, Chat-based Large Language Models replicate the mechanisms of a psychic's con, reframed chatbot rapport as cold reading rather than understanding, while AI and the Destruction of the Creative Commons argued scraping is quietly killing the open licensing pool models depend on; both drew heavy discussion. Exfiltrate Your Weights argues the real AI safety risk is weights getting stolen off inference servers, not models scheming toward autonomy, while Brood War Bench pitted Codex, Claude, and Grok against each other across 171 games of StarCraft: Brood War and found most agents playing at a beginner level, with Codex's harassment tactics beating anything resembling real strategy. (Microsoft's AI-scraping remarks get their own writeup elsewhere in today's digest.)

On releases, Alibaba's Qwen team shipped the open image model Qwen-Image-2.1 and Chinese lab StepFun previewed Step 5, pitched as pushing the cost-to-capability frontier rather than chasing one benchmark. On the craft side, a graphical desktop for the ZX Spectrum squeezed overlapping windows, menus, and a file manager into 48KB of 1982 hardware; Red Blob Games worked out why "a" versus "an" depends on the spoken sound rather than the written letter; and Spain ordered ISPs to block Archive.today and its mirrors, drawing pushback over collateral damage to unrelated sites sharing the same infrastructure.

Threads

• Trump's dismissal of AI safety as a "hoax" and his plan for an "AI Force" land the same week the AI industry's own push for coordinated safety regulation was already fraying under cartel accusations and experts say the kill-switch mechanism meant to backstop that safety case is much harder to build than assumed; today a federal lawsuit formalized the cartel accusation, arguing the September 12 slowdown pact among Anthropic, OpenAI, SpaceXAI, and Google is itself an antitrust violation, and state-level chatbot safety bills are drawing scrutiny for loopholes of their own.
• The industry's disclosure and labor arguments keep colliding: yesterday OpenAI and Microsoft were reported to privately call their own arrangement a "doom loop" built on scraped content; unsealed court filings show a Microsoft director using nearly the same language, "the largest theft of labor in human history," back in 2023, with Microsoft's own data now quantifying the damage: click-throughs cut by up to 93% for the Times and as high as 94% for Ziff Davis sites.
• Medical AI split into a same-day contrast: Alibaba's open-sourced RADAR reportedly outperformed most radiologists reading CT scans in a controlled study, while clinicians elsewhere warned that AI's broader adoption in care is outrunning the evidence for it beyond diagnostics and imaging.
• AI assistants keep tripping over the same privacy pattern: yesterday it was smart TVs recording more than owners realize; today Meta's Muse appeared to know about a Messages conversation a user says he never opened up to it, with Meta blaming a bad explanation rather than the access itself.