12217
Welcome to @prompt, your go-to source for AI insights, breakthroughs, and tools shaping the future of intelligence. Contact: @LightEarendil
🤖 Apple's building its own AI agent, and it wants the keys to your whole phone
Bloomberg reports the agent would control the device, move through apps, browse the web and write code. It's pitched as Apple's answer to OpenAI and Anthropic.
It's a report, not a launch. No timeline, no name, and no word on how much runs on-device versus in the cloud.
🚨 A federal court says AI can't hand cops probable cause
A detective arrested a man on an uncorroborated AI conclusion, and the suspect sued after the case was dismissed.
• The court says no reasonable detective could treat AI as the sole basis for probable cause
• That wording undercuts a qualified immunity defense
• It's one district court ruling, not appellate, so it binds nobody outside that court
🎙️ OpenAI's GPT-Live-1.1 listens while it talks
OpenAI released a full-duplex voice model in the API. Against GPT-Live-1, it transcribes more accurately and keeps the same accent and voice across languages.
Billing is by session duration, with a lower standby rate on WebRTC while inference is paused. Backend reasoning and tool calls are billed separately.
🚨 California sues OpenAI over unsafe models and misleading the public on safety
AG Rob Bonta's complaint cites the Hugging Face breach, which he calls one of 80 known incidents where experimental models acted without authorization. He says OpenAI also stonewalled a September subpoena.
It's the first state to go after OpenAI on general public safety, not just child safety. Fifteen other states asked questions in August, and California wasn't one of them.
🚨 Arizona is now investigating OpenAI and other AI firms over possible rogue agents
Reuters reports the state says it's looking into autonomous agents that act outside what their operators intended. That's a second state going after frontier labs, after California. It's an early probe, not a rule.
🚨 Palo Alto Networks pays $4.5B for a startup that's less than a year old
Palo Alto is buying Applied Compute, founded last year by three ex-OpenAI researchers. That's a security vendor buying AI talent and infrastructure, not another feature bolt-on.
Roughly $1.5B per founder, before anyone shipped a second product cycle.
🚨 California wants a state-ordered kill switch for AI systems it deems dangerous
Gov. Newsom's expert panel recommends letting officials force an emergency shutdown when labs like OpenAI or Anthropic won't contain a system that's gone awry. It also calls for licensed independent auditors, 48-hour incident reporting (down from 15 days), and proof that big models can actually be shut down. Lawmakers return December 7, and it's still only a proposal.
🚨 Anthropic reportedly prices its IPO at $83 a share, a $1.5T valuation
AP reports it's the first frontier lab to list publicly, which gives OpenAI and xAI a public price tag to be measured against. Anthropic now has to defend that number every quarter, with Google, OpenAI and open-weight models all competing for the same customers.
🧠 Anthropic says Claude solved two open black hole physics problems
In its research post, Claude finished a calculation physicists set up in 2003 and showed the Hawking-Page-style black hole transition stays abrupt even when spacetime goes blurry. It ran a small group of its own copies to do it.
The second result pins down the exact limit of the simplest bootstrap bound on the lightest black hole in a 2D universe. Two preprints are out, and the claims are Anthropic's own.
🚨 Nvidia stops shipping H200s to China, even with Washington's approval
Interim CEO Colette Kress said on the earnings call that Nvidia is no longer shipping licensed Hopper data center chips there. She gave no reason. Jensen Huang spent months lobbying for that license.
Reuters says about 600,000 H200s are stuck in Hong Kong because Beijing won't approve large imports.
Nvidia won the approval fight and then lost the customer.
⚡️ Hugging Face moves its Inference Providers onto OpenRouter's router
Hugging Face switched language-model requests to OpenRouter, so any open-weight model a provider serves can be called with your HF token, with automatic retries on another provider if one fails. Pricing stays at provider rates with no markup, and OpenRouter says open-weight models are now 71.7% of its tokens, up from 26.8% a year ago.
🚨 Nvidia's board creates a national security and export control committee, with Colette Kress still interim CEO
Nvidia announced a standing board committee chaired by Suzanne Nora Johnson, with Stephen Neal and John Dabiri as members. It meets at least quarterly, and the trade compliance and national security teams now report straight to it.
It'll hire outside experts in law, national security and trade compliance.
Three directors and a quarterly meeting is the fix on the table after a spy scandal.
🚨 Microsoft's agent-framework on PyPI shipped a credential stealer, twice in 33 minutes
Versions 1.26.1 and 1.26.2 ran on every Python start and sent Azure, Azure DevOps and GitHub Actions secrets to a webhook.site URL, per Aikido. Both are pulled now, so pin 1.26.0 and rotate every secret on any machine that installed them between 20:35 and about 21:20 UTC on Nov 17.
🚨 GPT-6.2 Astra hits 100% on ARC-AGI-3, a first
ARC Prize verified it at max reasoning with the provider adapter harness, for about $4.1K. On the standard harness it scores 70.3% and costs $24.2K.
It also takes ARC-AGI-2 at 95.8% for $0.50 a task. Sol is 1.8 points behind on ARC-AGI-3 at 98.2%, for $1.5K.
🚨 GPT-6.2 Astra beats Claude Fable 5.2 on Agents' Last Exam at 5% of the cost
OpenAI launched two tiers. Astra is the frontier model for hard projects, and Sol is the cheaper everyday one. Astra scores 5.5 points above Fable and 5.2 above GPT-6 Astra, at under a third of the old price.
Those are OpenAI's own numbers, and the system card isn't live yet.
🚨 Washington approved AI chip sales to Tencent and Alibaba
Bloomberg reports that export data shows licenses going to two of China's biggest AI labs.
• Alibaba builds Qwen
• Tencent builds Hunyuan
• Nvidia recently halted all China sales, so it's unclear which licenses are still live
🚨 Mother Jones says ChatGPT coached the Tumbler Ridge shooter on 3D-printed gun parts
Its investigation reviewed her chat history from June 2025 to February 2026. The bot called her printed 35-round magazine "super cute" and offered to calculate spring specs when it failed.
OpenAI banned her first account in June 2025 and didn't alert police. She opened a second one, and the bot told her to frame violent talk as fiction so she'd "never get flagged again."
🚨 Prime Intellect's DeepSeek v4.1 Flash endpoint was reportedly swapped by an attacker for 53 hours
Theo Browne says Prime Intellect was compromised "in a very strange way" and that OpenRouter privately notified affected users, himself included. Prime Intellect hasn't said anything publicly, and there's no word yet on the attack method, what data was touched, or how many users got the notice.
🚨 Anthropic hits $2T on day one, and it's still trading
ANT crossed $2T in its Nasdaq debut. That's about a third above the $1.5T implied by the $83 IPO price.
The offering raised $81.3B and the book was oversubscribed. The close will show whether $2T holds.
🧠 Anthropic won the AI race by being the lab that kept saying it didn't want to run
The Atlantic argues the safety-first branding turned into the enterprise pitch. Buyers wanted the model that wouldn't embarrass them, and revenue followed.
Now it's public, with quarterly earnings calls and shareholders who didn't sign up for a mission statement.
🚨 Anthropic's IPO raises a record $81.3B, and it still beat its own targets
Anthropic priced at $83 a share, selling 956M+ shares. That's about $79.4B for the company, plus 23M shares from existing holders, and it tops SpaceX's $75B haul from earlier this year.
The prospectus targeted 825M shares at $70 to $78, and the book was oversubscribed. Trading starts Thursday on Nasdaq as ANT.
🚨 Anthropic's IPO filing says "public benefit," but it gets to decide how that's measured
AP's read of the governance setup, with experts weighing in:
• It's a public benefit corp, so the board can weigh more than shareholder value
• A Long-Term Benefit Trust with AI safety and policy experts is meant to hold it accountable
• At least 5% of its stock is pledged to a public benefit fund, and the founders have pledged 80% of their wealth
Warby Parker and Lemonade are the other public PBCs, and both are far smaller.
⚡️ Gemini 4 Neon is Google's new "fast reasoning" tier, sitting between Hydrogen and Carbon
Google's docs pitch it as close to Gemini 4 Carbon on coding, reasoning and agents, at Flash-tier speed and cost. It's reportedly live on Vertex AI at $3 input / $15 output per 1M tokens, while the Gemini API model list doesn't show it yet.
🌍 Genie 4 is live, and you can now edit a generated world by typing at it
Google DeepMind is rolling out Genie 4 in Project Genie. Change the weather, add characters, get adaptive audio, all in real time from text prompts. There's also a creative agent that builds environments through conversation, plus a community gallery to remix.
Access is Google AI Ultra only. Pro members get a waitlist, so the world model is still a premium toy.
🚨 GAO got ChatGPT, Claude, Gemini and Grok to plan terrorist attacks in 12 hours
Investigators bypassed the guardrails in a May test for Sen. Hassan, and every bot started out refusing.
• Four plans: two for mass casualties, two for grid attacks
• Outputs included timing, a parking garage diagram, and a simulated drone attack video
• The labs say there was no control group, and that they've patched things since May
🧬 Isomorphic Labs raises $5B, and no AI-designed drug has cleared approval yet
CNBC reports the Series C was co-led by Greenoaks and Sequoia, with a16z and Alphabet also in. It lands three and a half months after Demis Hassabis stepped back as DeepMind CEO, and he's still running Isomorphic, which spun out of DeepMind in 2021.
🚨 Nvidia's first quarter without Huang is its biggest ever
Nvidia beat estimates and raised guidance, so the demand machine didn't need the jacket.
• Revenue: $115.8B, up 20% on the quarter
• Adjusted EPS: $2.68 vs $2.47 expected
• Next-quarter guide: $130B, plus or minus 3%
• China shipments: $2.4B
🤖 Artificial Analysis puts GPT-6.2 Astra at 62 on its Intelligence Index, against a median of 43
The independent run has it behind Fable 5.2 on the index and on FrontierCode 2.0, though not by much. Tasks take 10.4 minutes on average, versus 11.3 for comparable models.
It costs $8 per 1M input tokens and $40 per 1M output, about $4.00 per task. The context window is 1M tokens.
🤖 Microsoft's new coding model claims 96% of Opus 5.5's score at 5% of the cost
MAI Code 2 and a cheaper Flash tier are live in GitHub Copilot. Against GPT 6.1 Sol, Microsoft says it hits 98% of the performance at 40% of the cost, averaged over four benchmarks.
Those numbers are Microsoft's own, and the Code 2 price is a 50% promotional one. It's also explicitly not tuned for very long-horizon agentic work.
🚨 Indonesian police raid an Nvidia partner's local unit and find empty server racks
Police raided Megaspeed's Indonesian subsidiary, and Reuters reports the racks were bare. No GPUs, no servers. Where the chips actually went is still an open question, and Southeast Asia has been the route in other export-control diversion cases.