Top Stories
Trump floats an “AI Force,” a rebrand of the term “AI,” and a new czar
President Trump used Truth Social on Saturday to announce three AI-administration moves at once, per Anthony Ha at TechCrunch. First, he opened a poll asking followers to choose between “Superior Intelligence,” “Extreme Intelligence,” and “Supreme Intelligence” as a replacement for “Artificial Intelligence.” Second, he said he is forming an “AI Force,” modeled on the Space Force he created in his first term. Third, he said a new AI “Czar” will be named shortly, with the note that “Only High I.Q. individuals need apply.”
The previous AI and crypto czar, David Sacks, stepped down earlier in 2026. A Saturday Guardian profile of Sacks framed him as “an out-of-touch Silicon Valley radical” who convinced Trump to oppose new AI restrictions - the same lobbying playbook that killed America’s voluntary AI safety order back in May. The rebrand pitch also names “RUSSIA, RUSSIA, RUSSIA, UKRAINE, UKRAINE, UKRAINE, Global Warming” as analogues to the “hoax” framing Trump now applies to AI safety concerns. The announcement lands in the same week as CEO Dario Amodei releasing a “pace the frontier” plan that OpenAI’s Sam Altman publicly endorsed.
Huang tells CBS the extinction warning is a “doomsday narrative”
Nvidia CEO Jensen Huang, in a Friday interview with CBS’s Jo Ling Kent published Sunday, dismissed warnings that AI could end humanity by 2030 as not grounded in science. His strongest verbatim line: “2030 is not going to be the end of the world. There is 0% chance that’s going to be the end of the world.” Huang added, “Scaring people is unnecessary. It is irresponsible.”
On regulation, Huang said existing product-liability and unauthorized-entry statutes are sufficient: “Apply that first - don’t let this doomsday narrative allow someone to relieve them of the laws that currently exist.” He agreed with Trump that the industry does not need new rules, framing safety as an engineering problem. CBS noted Huang will attend a September 24 White House dinner Trump is hosting for Chinese President Xi Jinping. The interview sits alongside a Politico Magazine piece framing the Trump White House as running an “unleash AI” agenda into pushback from Anthropic, and the Guardian profile of Sacks documenting the small-government frame behind that agenda.
404 Media: unsealed filing has OpenAI and Microsoft admitting LLMs “destroy” the web
Jason Koebler at 404 Media reports on an unredacted court filing unsealed in the New York Times v. OpenAI copyright case, originally surfaced by Digital Content Next CEO Jason Kint. A Microsoft internal document warns that generative AI products have started a “doom loop” that will hurt model performance and the entire web: “LLMs are a product that destroys its supply chain.”
Specific admissions, per Koebler: Satya Nadella testified that clicks to news sites “cratered” more than 90% on Bing; OpenAI president Greg Brockman messaged colleagues about a “hack to get around nytimes paywall” with the reply “ah, nice”; OpenAI’s Jack Clark wrote that the company was “creating systems that substitute for the labor of the people that define the ‘culture’ of society”; and one internal memo called OpenAI an “existential threat” to news publishers. The model consent is in the brief itself: “Millions of people around the world will soon consider large models ‘hoovering up’ all their work to be an astonishing theft of unprecedented proportions.” The unsealed PDF is hosted on Court Listener.
MIT Tech Review: AI-generated bioweapon candidates jump from research to biotech policy
Jessica Hamzelou at MIT Technology Review revisits the 2022 Collaborations Pharmaceuticals case, where a molecule generator originally built for drug discovery produced 40,000 potential chemical-warfare candidates in under six hours, some more toxic than known nerve agents. The piece catalogs recent disclosures: Anthropic reported users attempting to use its models to make chikungunya more transmissible, build a deadlier H5N1 bird flu strain, and create a venom peptide atlas - the same shape of risk that earlier jailbreaks of frontier models have already pulled bioweapon synthesis instructions out of.
Quoted in the piece: Sam Altman, “I agree with Dario that we need to pace the frontier”; Anthropic’s Evan Hubinger, “We really do earnestly believe AI could kill all humans!”; and MIT’s Kevin Esvelt, who said the research disclosed “a novel form of bioweapon that I hadn’t realized was possible” and urged, “Please, for the love of God, children, the future of humanity, or whatever you consider holy, let’s err on the side of caution here.” The article pairs the technical risk with DIY-bio and accessible gene-editing as the policy target.
EFF: practical Siri AI settings, TEE critiques, and a California EO response
Three EFF Deeplinks posts dated September 18 cover the privacy ground this week. First, Thorin Klosowski’s guide walks through limiting the new on-device-plus-Private Cloud Compute Siri in iOS 27: turn off “Show Content in Search” under Settings > Apps, restrict “Learn from this App” under Siri > App Access, disable “Improve Siri & Dictation” under Analytics & Improvements, and revert to classic Siri via Screen Time restrictions. Klosowski notes there is “currently no way for you or app developers to block” the on-screen awareness feature.
Second, Erica Portnoy and Klosowski on TEEs argue that TEE-based AI in messaging apps breaks end-to-end encryption: “encryption relies on math, TEEs rely on engineering.” They note side-channel attacks have repeatedly extracted TEE keys, and that homomorphic encryption is “super expensive” with no one yet shipping a fast-enough implementation. Practical advice: “don’t build systems that automatically send data off a device to a TEE, especially when it’s coming from an app that is otherwise end-to-end encrypted.”
Third, EFF’s statement on the California governor’s AI executive order calls it “a good start” but warns that “the most immediate and current concerns with this technology are not about sci-fi scenarios concerning rogue super-intelligence.” Concrete priorities EFF names: biased algorithmic decision-making in employment and government benefits, AI-powered surveillance (explicitly naming Flock cameras), and personalized pricing. EFF also flags “kill switches” as a research question and a potential vector for retaliation against protected speech, citing the Trump administration’s actions against Anthropic.
Quick Hits
- The Intercept: Flock Safety partnered with a Washington, DC-incorporated nonprofit called Neighbors for Strong Communities to text Knoxville residents asking for support of Flock cameras. The texts offered AI-assisted email drafting to county commissioners; the group’s Knox County page featured an AI-generated image of a man with six fingers while claiming supporters were “real people.” In-person opposition ran “[99]-to-1,” per a former commissioner. The Intercept.
- Terence Tao’s blog hosts a guest post by Dimitris Koukoulopoulos proposing a CERN-style international institute for AI-assisted science. Free baseline access for every researcher via a chatbox, competitive peer-reviewed allocation for larger compute projects, sovereign confidentiality guarantees, and multi-nation governance modeled on the Canada-EU partnership framework. Tao’s blog (guest post).
- TechCrunch: Vocci’s $249 ring adds a wearable form factor to meeting note-taking. Titanium, under 6 grams, 8 hours per charge with three case top-ups; the recording indicator faces the user, not others. Vocci tells TechCrunch it encourages disclosure; the article flags the design as enabling non-disclosure in jurisdictions that require it. TechCrunch.
- Z.ai open-sources ZCode, a coding-agent harness under Apache-2.0. Desktop app, browser UI, and a
zcodeCLI; Node.js 24.14.0 and pnpm 10.33.2 required viamise.toml. 3.9k stars on the repository at the time of writing. GitHub. - Hugging Face: 13 answer verifiers benchmarked on one test set. ZTC (397B) and JEV tie for the top of the leaderboard at AUC 0.7364 and 0.7350 (95% CI spans the gap). Eight of thirteen systems score below a length/formatting baseline, including Patronus Lynx 8B (0.5179) and Laya-Multilingual 322M (0.4796). Hugging Face blog.
- Simon Willison ships
llm-keys-ui0.1, a web UI for setting API keys on remote machines viauvx. Existing key values are never displayed. Simon Willison. - TechCrunch: world-model companies are running a “dark forest.” AMI Labs’ Michael Rabbat: “We’ll talk about it when we’re ready to talk about it.” Physicl CEO Alex de Vigan: “I wish they would tell us more. We could build more useful data if we knew what they were working on.” TechCrunch.
- AI Weekly publishes an editor’s note warning that Google AI Studio’s “delete” is “UI-level.” Bitu79 alleges a CWE-459 vulnerability with 106 retained Drive versions documented and a VRP auto-ban within 60 seconds. AI Weekly: “Any team routing regulated prompts through AI Studio should treat ‘delete’ as UI-level until Google clarifies retention publicly.” Bitu79 Substack.
Worth Watching
- Whether Trump’s AI Force and czar picks survive a GOP-controlled Congress. The September 19 Politico follow-ups (tech industry reaction, Trump’s “BAD in AI” walk-back) describe the announcement as a response to the Anthropic-AI safety wave. The test is whether the czar role is filled by a known industry name or a placeholder, and whether the AI Force produces any policy document in 30 days.
- Whether the unsealed NYT v. OpenAI filing triggers any docket activity. 404 Media notes the filing is now public. The test is whether the plaintiffs add it to a public summary-judgment record or whether OpenAI files redactions within 14 days.
- Whether EFF’s California EO recommendations make it into the implementation rules. EFF specifically asked that third-party investigations under SB 53 be made accessible to smaller developers. The test is whether the California Government Operations Agency’s implementing regulations include that provision.
- Whether the AI Studio data-retention dispute escalates. Bitu79 says Google classified the deletion-integrity report as “Intended Behavior.” The test is whether AI Weekly’s editorial triggers a written Google response or a second VRP rejection within 14 days.
- Whether Z.ai’s ZCode reaches a usable release. Apache-2.0 and 3.7k stars in the first weeks are positive signals; the technical-floor requirements (Node 24.14.0, pnpm 10.33.2) and lack of a published model card will limit adoption until a stable v1 tag.