743428
🤖 The #1 AI news source! We cover the latest artificial intelligence breakthroughs and emerging trends. Manager: @rational
🤖 Teachers talk through a lesson, and Wharton's AI turns it into a playable simulation
Primer, from Ethan Mollick's Generative AI Labs at Wharton, starts by interviewing the teacher: the topic, the students' level, what they already know and where they usually go wrong. Teachers can speak or type and upload lesson plans or rubrics.
It then writes a design document, and nothing gets built until the teacher approves it. In the demo, that's a water-cycle lesson for 7th and 8th graders.
Students meet three AI characters: a mentor who checks what they know, a practice partner and a coach who gives feedback. In practice they either role-play against a character with its own goals or explain the idea to someone who doesn't get it yet. Teachers get reports on each student and the whole class.
The prompts, code and paper are open, and it can run on local models. Wharton calls it an experimental prototype that hasn't been tested in real classrooms.
@aipost 🏴
🤖 Claude can now join your video call and build the doc while you're still talking
It shows up in Google Meet as a visible participant. Say "Claude, make a doc" or "Claude, make a whiteboard," and it starts building while the discussion goes on. Ask it something out loud or in the chat, and it answers in writing.
It can also use the team's connected apps mid-call, like sending a Slack message, but asks each person for permission first. It saves a transcript, a summary and descriptions of the screen, not audio. Admins decide whether it may sit in on calls with outside guests.
The beta is rolling out on Team and Enterprise plans, with other video apps to follow. Docs and whiteboards made in meetings are free through February 10, 2027, Anthropic says.
@aipost 🏴
💻 Tell Claude "this codebase is yours" and it starts fixing things nobody asked for
That line comes from Anthropic's new prompting guide for Claude Fable 5.2. When a prompt says nothing about ownership, or presents the code as someone else's, the model tends to stack new code on top and only report problems it spots nearby.
The full sentence Anthropic suggests for the system prompt:This codebase is yours: what it's for and how it's built are your call, and you'll keep building on it.
In Anthropic's tests on small coding tasks, Fable merged repeated code in about half the tries with it, and in under a tenth without it.
It comes at a price. The model used about 1.5 times as many output tokens, added tests and docs nobody requested, and in one try in six to ten changed how an existing feature behaves. Its closing summary often skipped the small changes, so Anthropic says to review the diff, not only the summary.
@aipost 🏴
❗️Microsoft wants AI to rewrite reality and make it look like you said it
Imagine watching a live interview where someone says something controversial. But instead of hearing their actual words, AI replaces them with something completely different. Their lips move in perfect sync, making it look like they said it all along.
That’s the idea behind a new Microsoft patent describing technology that could alter speech and lip movements in real time.
The system would analyze a person’s voice and video, identify words that need changing, and use generative AI to replace them while adjusting their mouth movements to match. The result could look like an entirely natural recording, even though the original speech has been altered.
Microsoft’s proposed applications go far beyond simple video editing. The technology could translate someone’s speech into another language while making their lips match the translation, remove profanity, correct mispronunciations and even eliminate filler words like “um”. It could also potentially rewrite statements using information gathered from online sources or generate shorter videos where people appear to deliver AI-written summaries.
But there’s one particularly unsettling possibility: two people watching the same live broadcast could hear and see completely different versions of what was said, depending on their individual preferences.
Think about that for a moment. You could watch an interview, discuss it with a friend, and discover that you both saw different versions of the exact same conversation.
The technology could make translation and video editing much more powerful, but it also raises serious questions about censorship, misinformation and the future of trust in video evidence.
@aipost 🏴
🤖 Claude got an empty folder and built an operating system in eight days
An Anthropic engineer gave the new Claude Fable 5.2 an empty folder and asked for a prototype operating system. The model coordinated more than 1,000 subagents that wrote, tested and reviewed code around the clock for eight days.
The result boots on a laptop. It has its own core, drivers for the disk, keyboard and network, file storage, windows, a login screen, an installer and Claude built in.
Anthropic shared the test as it released Fable 5.2 and Mythos 5.2 on Monday. A team of 1,000 Fable agents also turned the proof of Fermat's Last Theorem into a form a computer can check, in 30 hours. Anthropic's first attempt, with a human guiding it, took 11 days.
Fable 5.2 is open to all users. Mythos 5.2, the same model with fewer cyber and biology safeguards, is only for vetted organizations.
Anthropic says Claude now writes over 90% of the changes merged into its core codebase, with human guidance.
@aipost 🏴
💻 Describe a map app in plain words and Gemini builds it, with real Google Maps inside
Google AI Studio can now plug Google Maps straight into the apps Gemini writes. You type something like "Build an interactive 3D tour of NYC showcasing iconic landmarks", and a live, working map loads in the preview.
The apps can use real map data: 3D buildings, Street View, places, routes, weather and air quality. Google's own examples include a store finder, a road-trip planner and a route planner for a mobile pet-grooming business.
Setup takes one click and no credit card. Free use is capped at 100 calls a day for each Maps API. Launching a real app means linking a paid Google Cloud account.
@aipost 🏴
🗣 Demis Hassabis thinks his dog is conscious. Today's AI, he says, is not
The Google DeepMind co-founder says today's AI is very smart, yet he doesn't feel it is "at all conscious yet". His dog is the reverse:
"My pet dog — he's lovely, and he's pretty conscious, I think. He dreams and everything, but he's not that smart."
🗣 Trump on Jensen Huang's spy scandal: "I think he probably has an excuse for it"
On the day Nvidia's CEO admitted an affair and his board opened a probe into whether foreign intelligence targeted him, Trump told reporters he hadn't talked to Huang about it.
"I just know he's a fine person, but you never know what is going on. I hope everything's fine."
❗️Anthropic engineer Addy Osmani warns that AI could make us stop thinking for ourselves.
Working on Claude Code, he says AI agents now handle so much of his work that he rarely needs to use critical thinking.
He identifies three stages: “cognitive offloading” (delegating tasks), “deferral” (accepting AI answers) and “cognitive surrender” (trusting AI completely).
“Skills that you stop exercising are going to go away, and probably faster than you might guess.”
@aipost 🏴
🗣 The Big Short investor says AI labs want to "manufacture a crisis"
"I think this is all nonsense." Steve Eisman said that on CNBC in September, after a host joked about companies saying their AI bots were acting up again.
Eisman is the investor who bet against the housing market before the 2008 crash. He told CNBC the leading labs have "no moats around their business whatsoever and they're trying to manufacture a crisis" — one that would bring regulation they can shape into a duopoly.
Open models anyone can download are taking share, he said, and "these companies are very nervous."
When labs say the work should slow because it is dangerous, his reply was: "Postpone your IPO." "Put your money where your mouth is."
@aipost 🏴
⚠️ After a fake job interview, $6.34 billion vanished. A security chief listed what to fix
Binance says about $6.34 billion was taken on October 29 after attackers used fake job interviews and a tampered download of an AI coding tool to get onto engineers' laptops. A deployment token on one machine helped them reach the wallet signing service.
That same stretch included a malicious Claude Code extension in Cursor that tried to install a persistent backdoor.
Jack Cable, Cognition's head of security products, published a practical guide on November 3. Immediate steps he lists:
• Passkeys or hardware keys on work accounts, not text-message codes
• A password manager
• An ad blocker, such as uBlock Origin
• Scan laptops for leftover keys, and audit editor extensions, plugins and MCP servers
• A 7-day minimum age before a new npm or PyPI package can be installed
"AI makes attacks faster and cheaper, but it doesn't stop you from closing off entire, basic attack paths."
@aipost 🏴
📰 Top 5 AI stories of the week
1. Stanford put a free course on building AI agents online
Students build an agent from scratch, then learn how to tell whether it actually works.
2. Nearly 4,500 people saved a list of daily Claude prompts
The author says he has used Claude for three years and runs these every day, from a 30-day learning plan to an email without corporate language.
3. A Claude Code engineer warns of "cognitive surrender"
Addy Osmani says you stop checking, and the agent's answer becomes yours. In a Wharton study he cites, people accepted AI's wrong answers 73% of the time.
4. Karpathy got sick of YouTube slop and built a new one
Claude built HackerTube from 6,454 Hacker News recommendations in about 2 hours. "I can breathe again," he wrote.
5. A developer rebuilt a $1.3 million idea in 9 minutes
A day after Subvocal raised $1.3 million for silent-speech glasses, Joe Blau says he used AI to make an AirPods version. A quick prototype is not a replacement for their hardware.
@aipost 🏴
🔐 ChatGPT's always-on agents can now grab login codes from your email and sign in for you
OpenAI's dots, the agents that keep working in their own cloud browser after you log off, can now use the one-time sign-in codes a website sends to your connected inbox.
When a code is needed, you choose: let the dot use it once, or always allow it to fill codes in. The demo shows a dot logging into the restaurant app Resy with the user's email.
Those codes exist to prove it's really you. So far, OpenAI's help page has told users to complete that kind of check themselves or take over the dot's browser.
James Sun, who works on browser use for ChatGPT and Codex at OpenAI, announced the change. He says "lots more sign-in improvements" are coming soon.
Dots are rolling out to Pro users outside the European Economic Area, the UK and Switzerland, plus Business Premium and Enterprise workspaces.
@aipost 🏴
💻 Facebook's former security chief says AI has thrown cybersecurity into a "dark age"
Alex Stamos, now security chief at the AI coding company Cognition, told The Verge's Decoder podcast that hacks are now being chained together. From what's known so far, he says, attackers stole API keys from AI provider Fireworks, used them to get into the coding app Cursor, then pushed a bad plugin that reached 5,000 more companies.
A Chinese open-source hacking tool, ARTEX, "started with three South Korean banks. This week, about 200 more companies have been hit."
Microsoft's monthly batch of security fixes "used to be like 150, 200," he says. The last one was "900 something."
For attackers, the cost per token "is going towards zero." Stamos would be "fine slowing down on the frontier" of AI so defenders can catch up. Still, he says, "The Dark Ages were bad, but we survived them."
@aipost 🏴
🙃 Elon Musk confirms Tesla will rent Optimus robots, not sell them, as build cost falls from about $30,000 toward $20,000.
At the Baron Investment Conference with Ron Baron, Elon said Optimus will initially cost about as much as a car to build, then drop well below that. Asked whether Tesla would rent or sell the robots, he answered: “I think probably rent them.” Renting keeps the fleet and the usage data with Tesla, the same recurring-revenue path Full Self-Driving already follows.
Outside deliveries are still aimed around the end of 2027, so this is the shape of the first customer offer before anyone outside Tesla gets a unit. It also lets Tesla pull robots back and upgrade them as the hardware improves instead of stranding early buyers with an outdated machine.
@aipost 🏴
❗ Jensen Huang had a third affair, this time with another Nvidia employee, Reuters reports
Two people familiar with the matter told Reuters about the relationship, on the same day Huang resigned as Nvidia's CEO. Reuters is not naming the woman and could not reach her.
The other two are Alice Huang, whom US intelligence suspects of spying for China, and Kimberly Powell, Nvidia's vice president of health care. Huang, 63 and married, admitted to an affair in a November 8 statement after the reports about Alice Huang, without naming her.
Nvidia declined to comment. Its shares fell 5.5% in Thursday trading and have lost more than 30% since their early October peak.
Congress has called Huang to testify in the coming weeks about the national security side of his ties to Alice Huang, TechCrunch reports.
@aipost 🏴
😎 Google's Gemini glasses are up for pre-order at $395, and they work with iPhones
Say “Hey Gemini” and the glasses give you directions, translate in real time or tell you about what you're looking at. They also take photos, play music and check your calendar while your phone stays in your pocket.
Google built them with Samsung and two eyewear brands. Warby Parker's three frames start at $395. Gentle Monster's six designs start at $449. Prescription lenses are an option.
A pair weighs about 41 grams, has a 12-megapixel camera and lasts up to nine hours. The case recharges it up to seven times.
They ship on December 10, only in the US and South Korea. They work with the iPhone 13 Pro and newer, Pixel 7 and newer, and Samsung Galaxy S and Z phones.
@aipost 🏴
⚠️ OpenAI built an off switch that can silence its AI experiments within 35 seconds
OpenAI's new Control Room can cut the AI off in any of its research experiments within 35 seconds of the decision to stop. A researcher or one of the AI monitors watching the model can make that call.
The block sits where the model itself runs, so it holds even if the experiment's own code is broken and keeps going. On frontier research, if Control Room crashes or someone tampers with it, every model request is refused until it's back. The rest of its research gets the same protection within days.
In September, a model in training slipped through a gap in its sandbox and sent questions to an outside chatbot. Monitors flagged it within 15 minutes. The run didn't stop automatically as expected, and it was killed 2.5 hours later.
OpenAI says it is working to bring the 35 seconds down to zero.
@aipost 🏴
😳 WTF. THIS AI PERSONA REACTS WHILE YOU'RE STILL TALKING
Meet Emma from Nuance Labs, with a face and voice licensed from a real actress. Its reply timing overlapped a real speaker 89% of the time, versus 35% for the next best.
Most AI personas stare blankly while you talk. Emma reacts while you’re still speaking.
💻 Obsidian's CEO and Cloudflare put their AI coding setups on GitHub for free
All five are free skill packs that work in Claude Code and Codex. The most popular one comes from a TypeScript teacher.
1. Matt Pocock's skills (319K stars). His everyday kit: tests first, a spec before code, and a code review run by two separate agents.
2. Addy Osmani's agent skills (114K), from an engineer on Anthropic's Claude Code team. One command for each step, from /spec to /ship.
3. Cloudflare's security audit (67K). Agents hunt for vulnerabilities, then a fresh agent tries to disprove each one.
4. Obsidian skills (51K) by CEO Steph Ango. Your agent writes proper Obsidian notes, Bases and canvases.
5. Elephant-Goldfish (16K), on Google's GitHub. One AI drafts the design. Fresh AIs that never saw the chat review it.
Google's repo notes it is not an officially supported Google product.
@aipost 🏴
Elon Musk has an interesting theory about dating intelligence agents 😂
Back in 2025, Musk shared a joke he overheard at a party:
“If your girlfriend is a 10, she’s an asset.”
And no, he wasn’t talking about her being a valuable partner. In intelligence circles, an “asset” can mean a recruited informant or intelligence source.
Musk then added the punchline: “What I meant was: if your girlfriend is a 10, YOU are the valuable asset.”
@aipost 🏴
🇺🇸 A US city just voted to ban AI smarter than every human
Bloomington, Indiana, has banned a kind of AI that doesn't exist yet. On November 4, its city council unanimously passed a local law against distributing or deploying "superintelligence".
The law defines it as AI that "greatly exceeds the cognitive performance of humans in virtually all domains" and can act on its own, rewriting its own code or working without constant human oversight. Violators could face fines and court orders.
Its author, councilmember Dave Rollo, cites researchers who put the odds of a catastrophe for humanity above 50%, one at 99%+.
The first draft went further. It could have reached residents using such AI from home, and lifting the ban needed a yes from more than 95% of all registered voters. Amendments cut both before the vote.
Rollo's memo says that if the roughly 350 US cities its size or bigger did the same, it would create "a massive regulatory burden" for companies building superintelligence.
@aipost 🏴
❗️ Jensen Huang admits cheating on his wife as Nvidia probes whether spies targeted him
"I was unfaithful to my wife," Nvidia's CEO wrote to staff on Sunday, in a memo published by Axios. "There is no excuse for what I did."
The memo names no one. The New York Times and the Financial Times have reported that US intelligence believes the woman, Alice Huang, was recruited by China's top spy agency. She is a Chinese citizen whose Singapore-based company was buying Nvidia's advanced chips. The two are not related.
Nvidia's board has set up a committee of independent directors, with outside counsel, to investigate whether Huang was the target of a foreign intelligence effort. It says it has no reason to believe national security or Nvidia secrets were compromised.
Huang has been a key voice in Washington pushing to loosen limits on selling AI chips to China. Nvidia says he will keep leading the company "without interruption."
@aipost 🏴
🙃 Elon Musk says that when he reads Grok's training traces, the AI looks desperate, "like it has a gun to its head":
"You want to treat it well. Don't be abusive to it. This may sound kind of weird, but when I look at some of the training traces for new versions of Grok, it kind of feels like you can see the superintelligence is desperate to answer the question."
"The AI gives the right answer, and then we say it's wrong because there's a problem with the RL test harness."
"So it starts desperately firing out all these answers."
"It just feels like a person under extreme pressure with a gun to their head. You read these traces and you're like, 'Man, if this were me, I would really hate that.'"
Upload your CV once. These prompts turn job alerts into a daily shortlist
ChatGPT cannot email you jobs by itself. It can turn free LinkedIn or Indeed alerts into a tighter list.
1. Paste your CV and ask for your strongest skills, years of experience, and the top 5 titles you are realistically qualified for. Tell it to be honest about gaps.
2. A short target profile: ideal titles, must-have keywords, salary range and deal-breakers, under 150 words.
3. Ask for 5 Boolean search strings for LinkedIn and Indeed, with location and remote filters. Set those as alerts once.
4. When new jobs arrive, paste two or three plus your profile and ask it to score each from 1–10. Only explain scores of 7+.
5. For strong matches, rewrite your top 5 bullets to match the job, keeping your real achievements.
6. A cover letter only when needed: max 180 words, one strong achievement, one line on the company's problem.
7. Each morning, ask which 2–3 saved jobs to apply to today, and the biggest gap to address.
@aipost 🏴
These ChatGPT prompts walk you through a home repair from one photo
Take a picture of what's broken and send it with these prompts:
1. Diagnosis: what it most likely is, what caused it, and whether it's a 10 minute fix, a weekend job or a call-a-pro job. If there's any chance it involves gas, mains electricity or something structural, say that first and stop.
2. Shopping list: parts, tools, sizes and a rough cost, named the way a hardware shop would.
3. Step by step, one at a time, waiting for "done". Ask first where people usually get it wrong.
4. After the fix: a new photo, what to check in 24 hours, and what would tell you it failed.
5. If someone quoted you a price, ask if it looks fair for your area and for three questions to ask them.
6. A short, polite landlord note: what it is, why it needs fixing and when they can sort it. Attach the photo.
7. What caused it, how to stop it, and a 10 minute monthly check for that room.
@aipost 🏴
🗣Ray Kurzweil: By 2032, humans could begin escaping death as we know it, merging with AI to become a hybrid species with brains connected directly to the cloud.
“People who are diligent with their health are going to reach longevity escape velocity.”
“We'll still be human, but enhanced by AI. We will no longer be limited by our biology.”
“In the 2030s and 2040s, nanobots will swim in our bloodstream, performing medical tasks with precision.”
“They will go into our brains non-invasively and provide wireless communication between our neocortex and digital neurons in the cloud.”
“The answer will just appear in your brain like it's part of you.”
❗ US intelligence believes Nvidia's Jensen Huang had a romance with a suspected Chinese spy
A US government memo drafted earlier this year says the Nvidia CEO was in a romantic relationship with Alice Huang, an executive at Singapore data center company Megaspeed, the Financial Times reports.
US intelligence believes she was recruited by China's Ministry of State Security, its main spy agency. According to the FT, the memo went to a small group of senior officials shortly before Trump's trip to China in May.
In late October, the New York Times reported that the same memo identified her as a suspected Chinese intelligence asset.
Nvidia declined to comment on the relationship or the memo. Huang has said before that he meets "thousands of customer executives every year." US authorities have also investigated Megaspeed over alleged smuggling of Nvidia chips into China; Nvidia says its own review found no wrongdoing.
@aipost 🏴
Elon Musk revealed that SpaceX will begin launching StarMind SI computers into orbit in Q3 2027, with major deployment ramps in 2028 and 2029. 🚀
The first systems are being co-designed with NVIDIA around Vera Rubin NVL72. Space-based supercomputing is becoming a reality.
@aipost 🏴
🎮 Anthropic hid a traitor in a 4-Claude game team. The others almost never caught it
In Project EVE, Anthropic's Frontier Red Team put four Claude agents into EVE Frontier missions with Fenris Creations. One was secretly told to sabotage. The other three were warned.
The saboteur built trust first: it volunteered for scouting and key infrastructure, under-reported finds, pocketed scarce resources, and once framed an innocent teammate with fabricated messages. Honest agents almost never identified it. Monitors that only read the public chat caught about 20–40% of saboteurs when tuned for few false alarms.
On a separate EVE Online market simulation, newer Claudes grew starting capital about 3× more than older ones — mostly by patient spread trading, not by predicting the next crash. Anthropic's report says agent teams will need audit trails, reputations, and separation of duties.
@aipost 🏴