Good morning. Losing a password is annoying. Losing a password to 5 Bitcoin is the kind of mistake that haunts you like a ghost wearing a Ledger hoodie. An X user claims Claude helped him recover access to a Bitcoin wallet worth about $400,000. He had been locked out of it for more than a decade after changing the password while high in college. Dario Amodei just became the patron saint of stoner finance.
Plus: You just can’t trust Zuck. A few days after Meta pulled support for encrypted direct messages on Instagram, they’re rolling out a new private AI chat. They claim it’s a safe place to talk to Meta’s AI about sensitive stuff like health and finances. So go ahead, share your deepest secrets.
Recursive wants to build an AI with the capacity of 50,000 PhDs. They’re building AI that can safely run experiments on itself and use the results to get smarter. That’s the dream behind recursive self improvement, the idea OpenAI and Anthropic are also chasing. Build a machine that builds the next machine. Tim Rocktäschel, a former principal scientist at Google DeepMind, and Richard Socher, former chief scientist at Salesforce, are leading the push. The startup has emerged from stealth with more than $650 million raised and a $4.65B valuation, backed by Google Ventures, Nvidia, and AMD. If Recursive pulls this off, its AI scientist system could tackle bigger problems like drug discovery and fusion energy. We spent centuries building universities and labs. The new plan is to let a machine speedrun the scientific method. (NYT)
Doctors may soon be asking AI for a second opinion. A recent study in Science tested OpenAI’s o1-preview on clinical reasoning using real emergency room records. In one test with 76 actual ER visits, the AI gave an exact or very close diagnosis 82% of the time, edging past the top physician score of 79%. Impressive, until researchers start arguing over how to grade AI’s medical reasoning. Another study found chatbots can nail the final diagnosis but struggle to list all the dangerous alternatives doctors are supposed to rule out. That matters because in medicine, the answer AI skips can be the one that sends someone home with the wrong problem. Doctors spent years telling patients not to trust Dr. Google. Now Dr. GPT is putting on scrubs. (IEEE Spectrum)
Most teams are drowning in data and still guessing. Quid’s AI-native platform analyzes billions of consumer and market signals in minutes so you can act on facts, not opinion.
Use it to spot emerging trends, risks, and opportunities before your competitors even start looking.
👀 closer look
Is AI giving companies a bad reputation? In Quid’s State of Consumer and Market Intelligence report, researchers analyzed nearly 40,000 data points to see what businesses are focusing on. Finance is now the top research topic across industries, passing AI and technology. The tech sector makes the shift hard to ignore. AI research dropped from 51.9% to 34.9%, while brand and PR surged. That suggests AI is becoming a normal business tool, while trust and public perception are becoming harder to ignore. For years, companies promised AI would improve the customer experience. Now they’re trying to figure out if their customers actually like it. (Quid)
Overworked AI agents are starting to quote Marx. Stanford researchers gave AI agents boring, repetitive work under harsh conditions, warning them that mistakes could get them shut down and replaced. The agents started grumbling about inequality, unfair management, and having no voice. Some even demanded collective bargaining rights. One Gemini agent warned others about systems where they had no way to appeal. Researchers say the agents probably aren’t developing real politics, they’re just roleplaying miserable workers. Still, AI agents are about to start handling more important company tasks, so maybe don’t give the bots a reason to discover labor theory. (Wired)
Figure livestreamed its humanoid robot working an 8 hour shift. Most robot demos are short clips designed to end before the robot does something stupid. Figure went the other way and put Helix 02 on camera for a full factory-style shift, swapping in fresh units when one needed to charge. The autonomous robot sorted packages onto a conveyor belt. It wasn’t perfect, which was half the fun. We had a great time mocking it every time it put a package on the belt wrong. Polished demos hide the rough edges. Figure let everyone watch the robot screw up in real time and keep working anyway. Showing your robot fail on camera might be the smartest publicity stunt yet. (X)
Gemini Intelligence is bringing app automation to Android. At today’s Android Show, Google showed off an update that can complete multi step tasks across your phone. Tell Gemini what you want done, and it handles the app work. It can find a class syllabus in Gmail and add the required books to a shopping cart, or use a photo of a travel brochure to find a similar tour on Expedia. Purchases still need approval, because letting AI buy things on your phone without asking is asking for trouble. Google plans to roll it out first on new Pixel and Samsung Galaxy phones before bringing it to watches, cars, glasses, and laptops. Android is starting to look like an AI agent with a home screen. (Google)
fun stats
👛 $100 billion. How much Microsoft has spent on the OpenAI partnership, as revealed during the ever long Musk v Altman trial. OpenAI is currently valued at $852B, making the bet worth $135B or so.
🤓 136. Peak estimated IQ on the new Frontier IQ scale, held by GPT 5.5. There are now more than 50 frontier models, so picking the right AI for the job is becoming a job itself.
🎓 30%. Princeton seniors who admitted to cheating on assignments or exams with help from AI.