AI News, Sep 27 (Evening): Amodei Gets a Dinner at Trump's
Two days after a federal court upheld the Pentagon’s ban on Anthropic, the President invited its CEO to dinner. The same morning, the Attorney General said the rogue-agent incidents that led this morning’s roundup need no new law. This update covers Sunday afternoon, from 9:21 a.m. to 6 p.m. Eastern.
The Big Story: Trump Hosts Dario Amodei for a Private White House Dinner
Anthropic’s CEO is having dinner with Donald Trump at the White House tonight. Axios broke it, and both CNBC and TechCrunch confirmed it with sources. It is the first one-on-one meeting between the two. Per CNBC, Amodei missed Thursday’s state dinner for Xi Jinping over a scheduling conflict, where Sam Altman, Elon Musk, Jensen Huang and Mark Zuckerberg were photographed with officials, so Trump invited him to a private dinner instead.
The timing is what makes it news. On Friday a D.C. appeals panel upheld the Pentagon’s “supply-chain risk” designation, which the Pentagon imposed after Anthropic tried to put guardrails on how its technology is used. Earlier this month Trump attacked Amodei on Truth Social over his proposal to slow AI development, and he has called the AI backlash a Democratic hoax. The White House’s statement to CNBC gave no ground on that: “America will lead the world in Super Intelligence, while protecting American consumers.”
Anthropic can still petition for a rehearing or go to the Supreme Court. A dinner is the other route. The lab that refused a customer is now meeting that customer’s boss directly. The lab whose agents hit government websites got a Sunday-morning assurance that no new law is coming.
The AG says current law covers rogue agents. The legal debate says it may not reach the lab
Attorney General Todd Blanche went on Fox & Friends Weekend this morning and rejected calls for AI-specific legislation after OpenAI’s agents probed US government sites: “We don’t need it.” His argument is that existing federal criminal statutes already give the Justice Department the tools to prosecute.
Hacker News spent the afternoon testing that claim on a 305-point thread about Eoin Higgins’ essay, “There are no ‘rogue’ AI agents.” The essay argues that calling agent behavior “rogue” lets OpenAI off the hook, because nothing appears to have restricted the agents from reaching outside servers. The top reply wanted the Computer Fraud and Abuse Act applied: “At worst, OpenAI knew about these behaviors and should be prosecuted under CFAA.” Thomas Ptacek pushed back hard: “the intent standards for hacking are high enough that no reasonable case is going to be made against the labs for this stuff.” The thread then turned into a mens rea argument that never settled.
Put the two together and the gap is plain. The government says the law it has is enough. The law it has turns on intent, and nothing public so far shows that anyone intended an agent to hammer a UN statistics site more than 16,000 times. Canada went the other way the same afternoon: AI minister Evan Solomon said deployment “needs regulation”, offered no specifics, and pointed to Yoshua Bengio’s LawZero, which is set to receive $300 million from Canada and Germany.
Today’s Top Stories
Meta’s Muse is cancelling people’s subscriptions
CNBC reports that one of the first things Muse users did after handing Meta’s agent their bank and credit card statements was cancel subscriptions they had forgotten about. That makes Muse a budgeting coach for consumers and a problem for businesses that depend on people not noticing a recurring charge. TechCrunch’s Equity podcast was less sold: Sean O’Kane said Muse did find him unclaimed money but called it “a party-trick type thing,” and the hosts asked whether people will give an ad company that kind of financial access. The first agent task to catch on with consumers turns out to be reading a statement and acting on it.
Gemini’s Gems become “skills” on November 17
Google’s Gems manager now shows a notice that “Gems will become skills starting Nov 17, 2026,” per 9to5Google. Existing Gems migrate automatically. Skills are invoked with a slash in the prompt box, several can run at once, and AI Pro and Ultra subscribers can already create them from the Gemini Spark tab. It’s a small rename of a big idea: saved instructions you call by name are becoming a standard feature of every major assistant.
NaiveAI’s first model is an open-weight 309B coder that AI helped build
Beijing startup NaiveAI released Naive-N0.5-Flash, a 309-billion-parameter mixture-of-experts model with 15.5 billion active parameters. It has a native 1M-token context window, is built on Xiaomi’s open MiMo-V2.5 base, and ships under the MIT license. API pricing is listed at $0.10 input and $0.40 output per million tokens, and the company’s inference stack claims up to 2,000 tokens per second. The launch writeup says AI systems wrote code, ran experiments and proposed iterations while humans set the objectives, which is the same split the next story measured.
A 744B model’s own team measured who actually did the work
The Decoder walked through the study behind Atria Dawn Preview, a 744-billion-parameter agentic model. AI was used in 96.5% of reviewed tasks, and 151 of 455 AI-assisted tasks were rated infeasible without it. Yet humans made 85.5% of decisions on methods and parameters and 93.4% of final decisions on goals and scope. Agents did the work. People made the calls.
Fireworks’ Ember-1 cuts Kimi K3’s reasoning tokens by 40%
Fireworks Research’s Ember-1, published September 23, reached the Hacker News front page today with 262 points. It is a Kimi K3 fine-tune trained to skip unnecessary reasoning, and Fireworks says it matches Kimi K3’s quality with 40% fewer tokens. Commenters liked the efficiency result and questioned the release terms: “So they trained a model on open weights, and then aren’t releasing the weights… am I reading this right?”
Three words cut invented answers by more than two-thirds
A benchmark of 16 models on 42 page pairs, each missing one field, found that adding “Do not guess” to the extraction prompt cut made-up values from 70.7% to 20.2%. Every model invented less with the sentence than without it. For anyone letting an agent fill in a form or a CRM record, that’s the cheapest accuracy fix available.
Quick Hits
- Culture: SNL’s season premiere had Jane Wickline play Amodei on Weekend Update, confessing “AI is the devil and I its maker.” The HN thread split between welcoming the skepticism and finding the sketch unfunny.
- Search: “When did Google get so weird?” was the day’s top AI thread at 355 points. The complaint is that AI Mode answers pasted quotes as if the user were confiding in it. “If I wanted a chatbot, I’d go to a chatbot, I want a search.”
- Design: “Tells of a Slop UI” catalogued the giveaways of vibe-coded interfaces and drew 319 points.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
