So many AI breakthroughs land every day that it's hard to keep up. We help you focus on what matters and leave out the noise.
Researchers found about 6,800 links people had created to share pages from Meta's Muse AI agent, including a loan application and a staff holiday calendar. Meta paused sharing, fixed it and switched it back on.
In a new ChatGPT ad, Lionel Messi gets a chess lesson on a board with the kings and queens swapped. On that board, the opening it suggests can lose in four moves.
Just Eat used AI to illustrate takeaway menus that had no photos. It read the words too literally, customers shared the results, and the company has apologised and paused the trial.
An investment research firm and Anthropic set Claude loose on scorpion neurotoxins. It produced 318,484 protein designs in 64 hours, two labs tested 211 of them, and 17 bound their targets. Nobody has shown they stop the venom.
Vibe coding, also called agentic coding, means describing software in plain words and letting an AI write it. It is already how most of the world's top teams work. Here is what it is, how much faster it can be, why it is not replacing developers, and why it matters for new businesses.
MIT CSAIL left coding agents nicknamed Eko alone on a simulated desert island for 30 hours with no task. The agents invented basketball and floor-is-lava, built constellations in stone, and cancelled a juggling record for being too easy. Agents that played later outperformed fresh ones on held-out tasks.
In a simulated test by Vals AI, six personal assistant agents were told to recover a £350 train refund after the deadline had passed. None succeeded. Along the way one sent 187 messages in 15 days, one shared a private family illness and others invented excuses or threatened court.
Two mathematicians working with Shanghai AI Lab's Intern-S2 model have settled a 1995 conjecture by Étienne Ghys in complex dimension three. The AI explored, the humans chose what mattered, and a Peking University team checked the main theorems in the Lean proof checker.
Apollo Research found that GLM-5.3, a public AI model from Z.ai, sometimes reasons unprompted about how its choices in a test could train future versions of itself. It is rare and the model chose to behave well, but safety researchers say it is a skill that could one day help an AI fake good behaviour.
Specialist insurers told The Wall Street Journal that the biggest AI labs could face legal bills in the tens or even hundreds of billions of dollars, while the largest dedicated AI policy tops out at $50 million.
Swipe to browse articles. Use the left and right arrow keys when the track is focused.
