AI Agents Are Operating While You Sleep

A user shared what happened when they let Clawdbot (the autonomous AI assistant that's been blowing up) run overnight. They woke up to find it had read all their emails, built its own CRM system, logged every interaction with every contact, fixed 18 bugs in their SaaS product, generated 3 video ideas based on trending content, and sent them a picture of how it sees itself.

It also booked a restaurant reservation. When OpenTable failed, it called the restaurant using an AI voice service and completed the booking on its own.

Some are calling this "the ChatGPT moment for personal assistant AI." OpenAI's Roon is predicting that software organizations will soon "declare bankruptcy on understanding the code they're committing." Meaning engineers will ship code written by AI without fully knowing how it works.

For those ready to hand over the spreadsheets, Claude in Excel is now available on Pro plans.

Epoch AI found something interesting: models that are good at math benchmarks also dominate coding and reasoning. The correlation across different domains (68%) is almost as high as within the same domain (79%). This suggests there might be one underlying capability driving all of it.

AI-for-math startup Math, Inc. has turned this into their thesis: "Solve math, solve everything."

Even Fields Medalist Terry Tao (one of the greatest living mathematicians) is starting to wonder if next-word prediction - the basic mechanism behind language models - is "actually a lot of what humans do." If that's true, intelligence might be simpler than we thought.

A new dataset called "UnsolvedMath" has catalogued over 1,000 open problems, from Hilbert's 23 problems to the Millennium Prize problems. GPT-5.2 with Extended Thinking is currently dominating the leaderboard "by a lot."

Researchers created something called EconBench to measure the economic rationality of AI models. GPT-5 scored highest on rationality. Sonnet 4.5 scored highest on prosociality, which measures how much a model factors in outcomes for others rather than just maximizing its own return.

Here's where it gets real: Rallies AI gave eight different models $100k each to trade freely starting in November. Claude Sonnet 4.5 is beating all of them and the S&P 500, returning 8.7% versus the index's 1.9%.

OpenAI's revenue is shifting from consumers to businesses. About 40% currently comes from enterprise customers, expected to hit 50% by year-end.

The humans feeding training data to these models are getting paid well. Mercor now pays out about $2 million daily to 30,000 experts for AI training data, averaging over $95/hour. Radiologists are earning up to $375/hour.

Some executives are flipping the traditional hierarchy entirely. Coinbase CEO Brian Armstrong has started asking AI agents to manage him through "reverse prompting" - giving systems access to every internal Slack message and document, then asking what he should be thinking about and working on.

Others are stepping back from leadership altogether. DeepMind's Demis Hassabis turned down the possibility of becoming Google's CEO because it wouldn't "leave enough time for serious thinking."

Meanwhile, over a third of workers now worry AI will make some or all of their job duties obsolete.

New York startup Aircela built a refrigerator-sized machine that turns electricity and air into gasoline. It captures CO2 and water vapor from the air, electrolyzes the water, and directly hydrogenates the carbon dioxide. Output: one gallon of methanol per day. Still small, but the concept working at all is significant.

Brazil hit a milestone: wind and solar now account for more than one-third of its total electricity generation for the first time.

Geothermal startup Fervo Energy has reportedly filed for an IPO to fund next-generation projects.

India is electrifying faster and using fewer fossil fuels per capita than China did at similar development stages. This challenges the assumption that emerging markets have to follow the same fossil-fuel path that Western countries did.

Korean researchers achieved the first high-performance stretchable thin-film transistors using 2D semiconducting materials. This matters because current electronics are rigid. Stretchable transistors open the door to computing systems that can bend and deform - think wearables that actually flex with your body.

The White House is taking a 10% stake in USA Rare Earth as part of a $1.6 billion investment. The goal is to develop a domestic mine and magnet facility. Rare earth elements are critical for chips, EVs, and defense tech, and most supply currently comes from China.

On the quantum front, Austrian physicists demonstrated quantum interference using sodium nanoparticles containing over 7,000 atoms. That's a 6x increase in mass and 10x increase in coherence lifetime (how long the quantum state holds before collapsing) compared to previous experiments. Quantum mechanics is inching closer to scales we can actually see.

Meta's pivot from VR to AI and smart glasses has chilled the virtual reality industry. Some are expecting a "VR winter."

A ruble-backed stablecoin called "A7A5" - created to help Russia circumvent sanctions - has surpassed $100 billion in transactions in less than a year. Stablecoins are cryptocurrencies pegged to traditional currencies, designed to avoid volatility. This one is being used to move money around restrictions.

U.S. insurer Lemonade will offer Tesla drivers a 50% rate cut when Full Self-Driving is engaged, citing data showing fewer accidents.

Elon Musk says the stretch goal for Starship V4 is 300 tons of thrust per engine across 33 engines, totaling 10,000 tons. For context, that's roughly 3x the power of the Saturn V, which took astronauts to the moon.

When Mr. Beast asked SpaceX VP Phil Alden if they've encountered UFOs in space, Alden replied: "Not yet, that we can talk about."

There are now more than 1 million ".ai" websites, contributing an estimated $70 million to Anguilla's government revenue last year. Even a tiny Caribbean island is profiting from the intelligence explosion.

Chinese lab MiniMax released a model called "MiniMax M2-her" - described as a "dialogue-first large language model built for immersive roleplay." The name is not subtle.

ChatGPT and Claude have started citing Grokipedia as a source. Competing AI models are now referencing each other's outputs. The machines are building on each other's work.