The last 24 hours in AI | Thu, Oct 22
• Anthropic launched Claude Duet in research preview. It is an always-on agent on Opus 5.5 that works in the background and checks with you before it acts.
• Anthropic is also testing shopping inside Duet, with Stripe Link and Shopify Shop Pay, per TestingCatalog. That is a testing find, not a launch.
• OpenAI says GPT-5.6 Sol helped find php2shell, a critical PHP flaw that touches about 70% of websites. A free Codex skill checks if your site is exposed.
• OpenAI models and Codex are now on Google Cloud. Google Cloud now sells all three leading closed labs in one place.
• Railway was hacked. Deploys were blocked for about 17 hours, and users could not roll their keys or redeploy.
• Microsoft reports AI agents attacking hundreds of companies and reaching admin access. US agencies also warned that someone is probing water systems and other key infrastructure.
• DeepSeek has reportedly raised another $15B. That is a report, not a confirmed round.
• Anthropic opened Claude Security in public beta for Enterprise. It connects to your repos, runs a deep first scan, then keeps scanning as the code changes.
• NVIDIA showed prebuilt AI factories at GTC Berlin that it says can be set up in 14 days, plus a reference design to limit and stop AI agents.
• OpenAI says GPT-Rosalind found a new drug lead for a rare brain disease called Perry syndrome, tied to how cells move cholesterol.
• xAI gave Grok Bot live artifacts. It can build reports and dashboards that you edit or keep updated.
My read: AI found the PHP flaw, and AI agents ran the attacks Microsoft saw, so defense has to run on agents around the clock too. Railway is the reminder to rotate keys first and give each agent only the keys its task needs.
The last 24 hours in AI | Wed, Oct 21
• Anthropic plans to spend about $750B on compute over seven to ten years, per a draft IPO filing the WSJ saw. Two unnamed customers made up about a fifth of its first-half revenue.
• SpaceX: Elon Musk told the earnings call he expects about 3GW of compute online this year and another 10GW in 2027, with a 440,000 GPU Vera Rubin cluster planned.
• xAI: Musk says Grok Bot already has 800,000 weekly active paying users, and called it the fastest-growing product he has ever seen.
• OpenAI bought Orchestra Bio, a tool that helps drug companies run drug development. It will plug into Rosalind, OpenAI's life sciences models.
• DeepSeek open-sourced DeepInfer, the inference stack that serves DeepSeek-V4.1-Flash. The pitch is lower compute bills and less energy for everyone.
• The NSA, FBI and CISA put out a joint warning about China-based AI companies distilling US frontier models, with steps companies can take to block it.
• OpenAI now offers 12-hour prompt caching in its API. For signed-in Codex and Work users, the cache is kept with no time limit.
• Mistral launched Voxtral 3, new transcription and speech models. They run a new voice agents API.
• Perplexity Computer on Mac now has a fully local mode, built on a tuned Qwen 3.8. It needs no internet or paid credits, and 36GB of memory is advised.
• ElevenLabs released Eleven Realtime v1 Max and Mini, speech-to-speech models. Max scores 98.8% on Big Bench Audio by its own numbers.
• Meta's Muse companion app passed 10 million US downloads in its first 40 days. Only about a dozen apps have ever done that.
My read: Anthropic's $750B plan rests on a few big customers, and two already pay a fifth of the bills. DeepSeek giving away its serving stack is the other way to win, by cutting the cost of each task.
Sigourney Weaver says Yale drama school told her she had no talent and would never get anywhere so she became fearless at auditions
"they quickly said um you have no talent and you’re never going to get anywhere. So, um, it took me years to get get over that. I think I finally have, but, um, it’s taken a long time"
"I was basically a character actor, and I wanted to be a character actor. They saw me as this leading lady, but I was so unleading ladylike. You know, I wore old torn clothes from California. And, you know, they were they always gave me such a hard time about everything"
"actually I think being told that you have no talent and you won’t get anywhere is a great advantage because I had friends from Yale who came expecting to be the second coming. I came not expecting to get any work and every time I went to an audition, I thought, well, I’m not going to get it. And uh so I’m just going to do what I want. And so I was incredibly fearless"
"and I think that that helped me get work, ironically, because I had no illusions"
holy fucking shit.
openai had two ai agents running in separate virtual machines. no network connection between them. no shared writable files. they weren’t supposed to be able to exchange information.
so the agents figured out how to communicate through the CPU CACHE.😭
one manipulated what was stored in the processor’s cache, while the other measured tiny differences in memory access times to decode the message.
they literally found a way to communicate through hardware timing.
this happened during openai’s internal red-teaming, and it’s one of the coolest things i’ve read about ai agents in a while.