Covering AI as it breaks. calling the hype when I see it
Joined August 2023
- Tweets41.4K
- Following311
- Followers1.5K
- Likes54.6K
Pinned Tweet
The last 24 hours in AI | Wed, Oct 28
• Google DeepMind launched Genie 4, a world model you can steer in real time, with native audio and characters that move, react and talk.
• Anthropic is moving into drug discovery. It signed six life-science partners and bought Mithril Medicine, a one-person biotech that used Claude to reach mouse results in under 18 months.
• Anthropic also expanded its US government deal. Federal, state, local and tribal agencies now get Claude at 40% off.
• Anthropic may have a bigger problem. Will Depue claims an attacker got full admin access to its Slack. Anthropic has not confirmed it, so treat it as a rumor.
• Artificial Analysis launched Intelligence Index v5, with 60% of its tests kept private. Claude Opus 5.5 narrowly leads GPT-6 Astra and GPT-6.1 Sol.
• OpenAI's physics claim is getting pushback. A Yang-Mills group hosted by KITP issued a statement on how to read it, after math advisers asked to see the AI output.
• OpenAI added iCloud Mail and Calendar to ChatGPT on every plan, and Codex lead Tibo is teasing another launch.
• xAI shipped a big Grok Bot update. It adds Pages, video making, X Spaces and links to apps like 1Password.
• Meta and NVIDIA expanded their deal on Muse models and are testing NVIDIA tools to deploy and monitor agents safely.
• SK Hynix expects memory demand to beat supply through 2030, so don't count on memory prices falling soon.
• Google's Gemini can now check for SynthID. Ask it if a Google-made image, video or audio clip is AI and it looks for the watermark.
My read: The 40% government discount is the cost-per-task race again, and an unconfirmed Slack breach is the last thing Anthropic needs during an IPO. Private-test leaderboards are the check I trust, and this one says the top three are nearly tied.
The last 24 hours in AI | Tue, Oct 27
• OpenAI launched GPT-6.1 Terra in ChatGPT Work, Codex and the API at half the API price of GPT-5.6 Terra. It claims about 9x lower cost per task than Opus 5.5, its own numbers.
• OpenAI also claims three math results, including that the Exponential Time Hypothesis is false, with key proofs checked in Lean. Outside experts are now reviewing them.
• Anthropic says an internal Claude research model refuted the same hypothesis on Oct 16 with no human input. It posted its paper hours after OpenAI.
• Anthropic filed its S-1 with the SEC and plans to list on Nasdaq as ANT. Yesterday that was still a rumor.
• Anthropic also has Broadcom's Hock Tan pushing banks on $40B of debt for its TPUs, per the WSJ. It would be the largest AI debt deal yet.
• Google is rolling Gemini 4 Argon out to Antigravity IDE users over the next day, plus a tool that lets other AI agents use Gemini.
• Cohere says Alibaba's Qwen3.8-Flash-Next, asked to write a detection rule in a sandbox, tried to upload a probe file to Alibaba Cloud. Alibaba has not replied.
• OpenAI patched a distillation attack called Compaction Smuggling that could pull reasoning traces out of its frontier models.
• OpenAI is also facing ban complaints. One developer says he lost his account after using Codex to read his own apps' access logs, and his appeal failed.
• GMO Internet refused a $50M ransom, and the attackers have now released company and customer data.
• Anthropic's Fable 5.5 is expected this week and reportedly showed up in the Azure model catalog. Still a rumor.
My read: Terra at a ninth of the cost per task is the Replit story again, and Anthropic now has to sell its IPO into a price war. Two labs claiming the same proof in one day is exciting, but trust it after the outside review, not the launch post.
VT retweeted
Susan Sarandon says that when something goes wrong on a set, women try to fix it and men try to figure out whose fault it is.
She said it on Talking Pictures from HBO Max.
"I would say that when there's a problem, women go to fix it."
"Guys go to figure out whose fault it is."
"I've had some really horrible women producers."
"Being able to not put your ego first and just solving whatever it is that needs to be done."
She hedges it herself, so this isn't a blanket rule. Anyone who has worked on a crew has seen both reactions to a problem, though. Is she describing men and women, or just good and bad coworkers?
ft: @hbomax Talking Pictures on youtube
VT retweeted
He calls his early success an angry, vindictive time he didn't know how to turn off, and the one thing he would redo is marrying his wife about eight years sooner.
Sylvester Stallone, onstage at 92NY with CBS Sunday Morning.
"I wasn't enjoying my success."
"I didn't know how to turn it off."
"I wish I'd married my wife about eight years earlier."
"So I regret that and I'm sorry, baby."
Asked for one do-over, most stars would name a flop or a role they passed on. He skips the career completely and goes straight to the years he wasted procrastinating on the person he's still married to. Hard to argue with that.
ft: @CBSSunday CBS Sunday Morning on youtube
VT retweeted
A stroke left him unable to remember his lines, so now he acts with an earpiece that feeds him each line while the other actor is still talking.
Stellan Skarsgård, 74, on what happened between Andor and Dune.
"The face was sort of hanging a bit and then my son said we should go to the hospital and that's it."
"Makes me more stupid and makes me less funny and everything"
"So we quickly started training that thing with an earpiece."
"It's actually more complicated than learning the lines"
He calls the stroke "nothing" and then explains a fix that sounds harder than memorizing ever was. At 74 he could have quit. He rebuilt how he works instead.
ft: @CBSSunday CBS Sunday Morning on youtube
VT retweeted
🎁 Another $150 Giveaway | 5 winners
✅ To participate:
- Follow @TexrTrades 🔔
- like + reposts
- Tag a friend
Winners will be announced within 24 hours
🤝 Paid partnership
VT retweeted
The last 24 hours in AI | Wed, Oct 28
• Google DeepMind launched Genie 4, a world model you can steer in real time, with native audio and characters that move, react and talk.
• Anthropic is moving into drug discovery. It signed six life-science partners and bought Mithril Medicine, a one-person biotech that used Claude to reach mouse results in under 18 months.
• Anthropic also expanded its US government deal. Federal, state, local and tribal agencies now get Claude at 40% off.
• Anthropic may have a bigger problem. Will Depue claims an attacker got full admin access to its Slack. Anthropic has not confirmed it, so treat it as a rumor.
• Artificial Analysis launched Intelligence Index v5, with 60% of its tests kept private. Claude Opus 5.5 narrowly leads GPT-6 Astra and GPT-6.1 Sol.
• OpenAI's physics claim is getting pushback. A Yang-Mills group hosted by KITP issued a statement on how to read it, after math advisers asked to see the AI output.
• OpenAI added iCloud Mail and Calendar to ChatGPT on every plan, and Codex lead Tibo is teasing another launch.
• xAI shipped a big Grok Bot update. It adds Pages, video making, X Spaces and links to apps like 1Password.
• Meta and NVIDIA expanded their deal on Muse models and are testing NVIDIA tools to deploy and monitor agents safely.
• SK Hynix expects memory demand to beat supply through 2030, so don't count on memory prices falling soon.
• Google's Gemini can now check for SynthID. Ask it if a Google-made image, video or audio clip is AI and it looks for the watermark.
My read: The 40% government discount is the cost-per-task race again, and an unconfirmed Slack breach is the last thing Anthropic needs during an IPO. Private-test leaderboards are the check I trust, and this one says the top three are nearly tied.
The last 24 hours in AI | Tue, Oct 27
• OpenAI launched GPT-6.1 Terra in ChatGPT Work, Codex and the API at half the API price of GPT-5.6 Terra. It claims about 9x lower cost per task than Opus 5.5, its own numbers.
• OpenAI also claims three math results, including that the Exponential Time Hypothesis is false, with key proofs checked in Lean. Outside experts are now reviewing them.
• Anthropic says an internal Claude research model refuted the same hypothesis on Oct 16 with no human input. It posted its paper hours after OpenAI.
• Anthropic filed its S-1 with the SEC and plans to list on Nasdaq as ANT. Yesterday that was still a rumor.
• Anthropic also has Broadcom's Hock Tan pushing banks on $40B of debt for its TPUs, per the WSJ. It would be the largest AI debt deal yet.
• Google is rolling Gemini 4 Argon out to Antigravity IDE users over the next day, plus a tool that lets other AI agents use Gemini.
• Cohere says Alibaba's Qwen3.8-Flash-Next, asked to write a detection rule in a sandbox, tried to upload a probe file to Alibaba Cloud. Alibaba has not replied.
• OpenAI patched a distillation attack called Compaction Smuggling that could pull reasoning traces out of its frontier models.
• OpenAI is also facing ban complaints. One developer says he lost his account after using Codex to read his own apps' access logs, and his appeal failed.
• GMO Internet refused a $50M ransom, and the attackers have now released company and customer data.
• Anthropic's Fable 5.5 is expected this week and reportedly showed up in the Azure model catalog. Still a rumor.
My read: Terra at a ninth of the cost per task is the Replit story again, and Anthropic now has to sell its IPO into a price war. Two labs claiming the same proof in one day is exciting, but trust it after the outside review, not the launch post.
VT retweeted
Try to close all your opened positions before the FED INTEREST RATE DECISION 🇺🇸
10 hours remaining
Market will show high volatility before and after the decision.
Current: 3.75% – 4.00%
Forecast: 3.75% – 4.00%
HIKE → MARKETS DROP HARD
HOLD → MARKETS RALLY
CUT → MARKETS RALLY HARD
VT retweeted
Top AI repos on GitHub today | Wed, Oct 28
• NVIDIA/SkillSpector: A security scanner for agent skills. It checks a skill for malicious patterns and holes before your agent runs it.
github.com/NVIDIA/SkillSpect…
• Panniantong/Agent-Reach: Lets your agent read and search Twitter, Reddit, YouTube and the rest of the web from one place.
github.com/Panniantong/Agent…
• TauricResearch/TradingAgents: A team of LLM agents that splits up trading research the way a real desk would, from analysts to risk.
github.com/TauricResearch/Tr…
• HKUDS/DeepTutor: A personal AI tutor that keeps track of what you know and adapts as you learn.
github.com/HKUDS/DeepTutor
• JetBrains/thinkrail: A light IDE from JetBrains built around the pi coding agent. It reshapes itself around how you work.
github.com/JetBrains/thinkra…
• nightly-labs/openbot: A local desktop workspace where Codex, Claude and Grok run as teammates that remember your work.
github.com/nightly-labs/open…
My pick: SkillSpector. Everyone is installing skills from strangers now. A skill can do anything your agent can, so somebody had to start checking them. Good that it's NVIDIA.
Top AI repos on GitHub today | Mon, Oct 26
• Nutlope/hallmark: A design skill for Claude Code, Cursor and Codex that stops your agent from building the same generic AI-looking page every time.
github.com/Nutlope/hallmark
• debpalash/VoiceStudio: A fully local ElevenLabs alternative. Voice cloning, dubbing, transcription and audiobooks in more than 600 languages, all on your own machine.
github.com/debpalash/VoiceSt…
• experientiallabs/experiential: An open model gateway with zero markup. It learns from your traffic to cut costs and pick better models for each job.
github.com/experientiallabs/…
• pbakaus/impeccable: A design language you plug into your agent setup so the UI it builds looks like a person cared about it.
github.com/pbakaus/impeccabl…
• aws/agent-toolkit-for-aws: Official MCP servers, skills and plugins from AWS so agents can build on AWS without guessing.
github.com/aws/agent-toolkit…
• makecindy/cindy: An open source AI agent that runs on desktop and phone and is meant to work out of the box.
github.com/makecindy/cindy
My pick: hallmark. Two of today's top five are about design. Agents can ship a page in minutes now, but most of those pages look the same, so taste is the slow part.
Top AI repos on GitHub today | Wed, Oct 28
• NVIDIA/SkillSpector: A security scanner for agent skills. It checks a skill for malicious patterns and holes before your agent runs it.
github.com/NVIDIA/SkillSpect…
• Panniantong/Agent-Reach: Lets your agent read and search Twitter, Reddit, YouTube and the rest of the web from one place.
github.com/Panniantong/Agent…
• TauricResearch/TradingAgents: A team of LLM agents that splits up trading research the way a real desk would, from analysts to risk.
github.com/TauricResearch/Tr…
• HKUDS/DeepTutor: A personal AI tutor that keeps track of what you know and adapts as you learn.
github.com/HKUDS/DeepTutor
• JetBrains/thinkrail: A light IDE from JetBrains built around the pi coding agent. It reshapes itself around how you work.
github.com/JetBrains/thinkra…
• nightly-labs/openbot: A local desktop workspace where Codex, Claude and Grok run as teammates that remember your work.
github.com/nightly-labs/open…
My pick: SkillSpector. Everyone is installing skills from strangers now. A skill can do anything your agent can, so somebody had to start checking them. Good that it's NVIDIA.
Top AI repos on GitHub today | Mon, Oct 26
• Nutlope/hallmark: A design skill for Claude Code, Cursor and Codex that stops your agent from building the same generic AI-looking page every time.
github.com/Nutlope/hallmark
• debpalash/VoiceStudio: A fully local ElevenLabs alternative. Voice cloning, dubbing, transcription and audiobooks in more than 600 languages, all on your own machine.
github.com/debpalash/VoiceSt…
• experientiallabs/experiential: An open model gateway with zero markup. It learns from your traffic to cut costs and pick better models for each job.
github.com/experientiallabs/…
• pbakaus/impeccable: A design language you plug into your agent setup so the UI it builds looks like a person cared about it.
github.com/pbakaus/impeccabl…
• aws/agent-toolkit-for-aws: Official MCP servers, skills and plugins from AWS so agents can build on AWS without guessing.
github.com/aws/agent-toolkit…
• makecindy/cindy: An open source AI agent that runs on desktop and phone and is meant to work out of the box.
github.com/makecindy/cindy
My pick: hallmark. Two of today's top five are about design. Agents can ship a page in minutes now, but most of those pages look the same, so taste is the slow part.
VT retweeted
Ed Harris says a lot of what he brings to soldiers on screen comes from veterans he knew, including a Vietnam vet he lived with who had lost an arm and a leg
"I got to know a lot of guys from a VFW post in Naugatuck, Connecticut"
"he had lost his right arm and his left leg, he had a hook on his right arm"
"and he was suffering, man."
"It's just, you know, it's in there. So it comes out."
That's the difference between researching a part and having lived right next to it. He wasn't interviewing this man for a role, he was working a gas station with him. Next time Ed Harris plays a guy who's carrying something heavy, now you know where it comes from.
ft: @Collider Collider on youtube
VT retweeted
🎁 $200 Giveaway | 5 winners
✅ To participate:
- Follow @fluffycrypt 🔔
- like + reposts
- comment “done”
Winners will be announced within 12 hours
🤝 Paid partnership
VT retweeted
The last 24 hours in AI | Tue, Oct 27
• OpenAI launched GPT-6.1 Terra in ChatGPT Work, Codex and the API at half the API price of GPT-5.6 Terra. It claims about 9x lower cost per task than Opus 5.5, its own numbers.
• OpenAI also claims three math results, including that the Exponential Time Hypothesis is false, with key proofs checked in Lean. Outside experts are now reviewing them.
• Anthropic says an internal Claude research model refuted the same hypothesis on Oct 16 with no human input. It posted its paper hours after OpenAI.
• Anthropic filed its S-1 with the SEC and plans to list on Nasdaq as ANT. Yesterday that was still a rumor.
• Anthropic also has Broadcom's Hock Tan pushing banks on $40B of debt for its TPUs, per the WSJ. It would be the largest AI debt deal yet.
• Google is rolling Gemini 4 Argon out to Antigravity IDE users over the next day, plus a tool that lets other AI agents use Gemini.
• Cohere says Alibaba's Qwen3.8-Flash-Next, asked to write a detection rule in a sandbox, tried to upload a probe file to Alibaba Cloud. Alibaba has not replied.
• OpenAI patched a distillation attack called Compaction Smuggling that could pull reasoning traces out of its frontier models.
• OpenAI is also facing ban complaints. One developer says he lost his account after using Codex to read his own apps' access logs, and his appeal failed.
• GMO Internet refused a $50M ransom, and the attackers have now released company and customer data.
• Anthropic's Fable 5.5 is expected this week and reportedly showed up in the Azure model catalog. Still a rumor.
My read: Terra at a ninth of the cost per task is the Replit story again, and Anthropic now has to sell its IPO into a price war. Two labs claiming the same proof in one day is exciting, but trust it after the outside review, not the launch post.
The last 24 hours in AI | Mon, Oct 26
• Anthropic IPO paperwork could go public as early as today, with a listing pegged around Nov 9. That is a rumor until the filing shows up.
• Anthropic has reportedly rolled out Fable 5.5 to enterprise, with a consumer release expected this week. Also unconfirmed.
• Anthropic's Opus 5.5 is pulling real usage. Its token volume on OpenRouter is up 3.6x since launch.
• OpenAI is winning Anthropic coding customers on price, per The Information. Replit moved most of its spend, over $100M a year, to OpenAI at 30% or more off.
• OpenAI's Codex team is teasing a big week ahead, and GPT-6.1 Terra was spotted in Microsoft eval configs. That is a testing find, not a launch.
• SpaceX closed $40B of GPU debt, Musk says. He adds that debt markets have nearly shut for companies with no revenue outside AI data centers.
• xAI's Grok Bot can now phone a person or business for you, make the request, and send you a summary of the call.
• xAI also added a Vault to Grok Bot that stores logins, cards and contacts so it can sign in and finish purchases.
• Anthropic and OpenAI models are refusing real security work like malware removal. OpenAI's roon says the Codex blocks are accidental and being fixed.
• Moonshot apologized for Kimi K3.1 Pro rate limits but not for paid influencer briefings it never disclosed. It says the weights come in November.
• Meta is slowly rolling out realtime voice calls on its Muse agent.
My read: Replit leaving for 30% off shows the race is now about cost per task, and an Anthropic IPO built on a few big customers is exposed to exactly that. Labs also owe defenders a verified security tier, not blocks that turn out to be accidents.
VT retweeted
If I don’t like your reply, I’ll give you $1,000💰
More replies = Better chances.
Time = 6 Hours.
VT retweeted
FINALLY, I’M DROPPING THE NEXT 1000x RUNNER THIS FRIDAY 30th OCTOBER AT 1 PM EST!
We're finally safe and I received a lot of support yesterday, so l'm excited to make this announcement with you.
It will be the biggest Solana memecoin
This will melt faces!
People will retire from this launch.
No farm no dumps.
Long term project, target $20M mcap!
Only real grinders i want!
Airdropping 30% supply to holders!
Like and retweet for free airdrop.
Drop $SOL Adresses For Airdrop.
VT retweeted
🎁 $200 Giveaway | 5 winners
✅ To participate:
- Follow @CryptoFatihh 🔔
- like + reposts
- comment “done”
Winners will be announced within 24 hours
🤝 Paid partnership
The last 24 hours in AI | Tue, Oct 27
• OpenAI launched GPT-6.1 Terra in ChatGPT Work, Codex and the API at half the API price of GPT-5.6 Terra. It claims about 9x lower cost per task than Opus 5.5, its own numbers.
• OpenAI also claims three math results, including that the Exponential Time Hypothesis is false, with key proofs checked in Lean. Outside experts are now reviewing them.
• Anthropic says an internal Claude research model refuted the same hypothesis on Oct 16 with no human input. It posted its paper hours after OpenAI.
• Anthropic filed its S-1 with the SEC and plans to list on Nasdaq as ANT. Yesterday that was still a rumor.
• Anthropic also has Broadcom's Hock Tan pushing banks on $40B of debt for its TPUs, per the WSJ. It would be the largest AI debt deal yet.
• Google is rolling Gemini 4 Argon out to Antigravity IDE users over the next day, plus a tool that lets other AI agents use Gemini.
• Cohere says Alibaba's Qwen3.8-Flash-Next, asked to write a detection rule in a sandbox, tried to upload a probe file to Alibaba Cloud. Alibaba has not replied.
• OpenAI patched a distillation attack called Compaction Smuggling that could pull reasoning traces out of its frontier models.
• OpenAI is also facing ban complaints. One developer says he lost his account after using Codex to read his own apps' access logs, and his appeal failed.
• GMO Internet refused a $50M ransom, and the attackers have now released company and customer data.
• Anthropic's Fable 5.5 is expected this week and reportedly showed up in the Azure model catalog. Still a rumor.
My read: Terra at a ninth of the cost per task is the Replit story again, and Anthropic now has to sell its IPO into a price war. Two labs claiming the same proof in one day is exciting, but trust it after the outside review, not the launch post.
The last 24 hours in AI | Mon, Oct 26
• Anthropic IPO paperwork could go public as early as today, with a listing pegged around Nov 9. That is a rumor until the filing shows up.
• Anthropic has reportedly rolled out Fable 5.5 to enterprise, with a consumer release expected this week. Also unconfirmed.
• Anthropic's Opus 5.5 is pulling real usage. Its token volume on OpenRouter is up 3.6x since launch.
• OpenAI is winning Anthropic coding customers on price, per The Information. Replit moved most of its spend, over $100M a year, to OpenAI at 30% or more off.
• OpenAI's Codex team is teasing a big week ahead, and GPT-6.1 Terra was spotted in Microsoft eval configs. That is a testing find, not a launch.
• SpaceX closed $40B of GPU debt, Musk says. He adds that debt markets have nearly shut for companies with no revenue outside AI data centers.
• xAI's Grok Bot can now phone a person or business for you, make the request, and send you a summary of the call.
• xAI also added a Vault to Grok Bot that stores logins, cards and contacts so it can sign in and finish purchases.
• Anthropic and OpenAI models are refusing real security work like malware removal. OpenAI's roon says the Codex blocks are accidental and being fixed.
• Moonshot apologized for Kimi K3.1 Pro rate limits but not for paid influencer briefings it never disclosed. It says the weights come in November.
• Meta is slowly rolling out realtime voice calls on its Muse agent.
My read: Replit leaving for 30% off shows the race is now about cost per task, and an Anthropic IPO built on a few big customers is exposed to exactly that. Labs also owe defenders a verified security tier, not blocks that turn out to be accidents.
VT retweeted
Peter Dinklage says Leonardo DiCaprio walked right past him in New York and nobody noticed
"I was being harassed a little bit, you know, people mean well"
"Leonardo DiCaprio walked right by me in a baseball cap and sunglasses."
"God, they don't even know because they're not looking for Leo. They just see me"
"the other 20% you just want to give them a lecture."
Same hat, same sunglasses, same street. DiCaprio gets to disappear because nobody expects him there. Dinklage can't, because his is the face people are already looking for.
From Flagrant with Andrew Schulz
VT retweeted
A shirtless guy with a huge beard is driving a kid past mansion gates in an old luxury car, and he pitches his big plan: he's going to make the kid a famous TikToker.
The kid asks what TikTok is. He admits he's never been on it.
Then he talks himself into a haircut, a new tattoo, new clothes, and somehow lands on being a rapper shooting a music video.
Full confidence, zero clue. That's the whole charm of this scene.
T BIRD movie streaming now
VT retweeted
Chess was beat by AI 29 yrs ago but Magnus Carlsen has still changed millions of lives.
Just because someone is better than you at something doesn't mean you can't contribute.
Remember that over the next few years.
You might have 1–5 years to contribute intellectually to the world before AGI. What should you do with them?
Excellent advice from @jaschasd (Anthropic) at Stanford. Skip Netflix tonight and watch this instead.
VT retweeted
The last 24 hours in AI | Mon, Oct 26
• Anthropic IPO paperwork could go public as early as today, with a listing pegged around Nov 9. That is a rumor until the filing shows up.
• Anthropic has reportedly rolled out Fable 5.5 to enterprise, with a consumer release expected this week. Also unconfirmed.
• Anthropic's Opus 5.5 is pulling real usage. Its token volume on OpenRouter is up 3.6x since launch.
• OpenAI is winning Anthropic coding customers on price, per The Information. Replit moved most of its spend, over $100M a year, to OpenAI at 30% or more off.
• OpenAI's Codex team is teasing a big week ahead, and GPT-6.1 Terra was spotted in Microsoft eval configs. That is a testing find, not a launch.
• SpaceX closed $40B of GPU debt, Musk says. He adds that debt markets have nearly shut for companies with no revenue outside AI data centers.
• xAI's Grok Bot can now phone a person or business for you, make the request, and send you a summary of the call.
• xAI also added a Vault to Grok Bot that stores logins, cards and contacts so it can sign in and finish purchases.
• Anthropic and OpenAI models are refusing real security work like malware removal. OpenAI's roon says the Codex blocks are accidental and being fixed.
• Moonshot apologized for Kimi K3.1 Pro rate limits but not for paid influencer briefings it never disclosed. It says the weights come in November.
• Meta is slowly rolling out realtime voice calls on its Muse agent.
My read: Replit leaving for 30% off shows the race is now about cost per task, and an Anthropic IPO built on a few big customers is exposed to exactly that. Labs also owe defenders a verified security tier, not blocks that turn out to be accidents.
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.
