Posts
327 postsShort-form notes on AI, engineering, and the work in flight.
A dwarf planet candidate about 590 km across sat in public telescope images for 12 years, until Joshua Wolk, Will DePue
The models added the capacity to try a novel approach at that scale. The three of them still had to decide where to point it, and how to check the result. They first found traces of the object in 2025 data from the Rubin Observatory, in a crowded...

A clean rebuild, an axiom audit, and a Comparator check all passed on OpenAI's Lean formalization of four-dimensional Ya
In the judgment of the Yang-Mills Working Group led by David Gross at the Kavli Institute for Theoretical Physics, part of UC Santa Barbara, the code proves that a relativistic quantum field theory satisfying the Wightman axioms can be constructed...

A robot from Caltech assembled the third bay of a 1.4-metre-diameter antenna in orbit, after the payload launched with t
That is the real scale of the ASTRA demonstration that Caltech and DARPA announced on 5 November. The posts making the rounds describe a robot building a structure in space on its own. A new bay did go together in orbit, without astronauts. The two...
Fable 5.2 finished 99 of 240 real freelance projects at least as well as the professional who was paid for them, in a Re
That is a genuine jump: 41.3% at high effort, nearly double Claude Opus 5.5's 21.3%. Scale AI said nearly 50 of those wins were projects no earlier agent had completed, and the gains were uneven. Audio and music, architecture and interior design,...

The agent that writes a design should not be the one that reviews it. That's the whole idea behind Google's Elephant-Gol
It started inside Google, where they say thousands of engineers use it every day to ship production code. Last month Google open-sourced it as a plugin that works with Antigravity, Claude Code, Codex and ChatGPT. How it works: 1. The elephant, an...
Three working electronic devices came out of a one-paragraph prompt to Claude Fable 5.2, and none of them needed an elec
The test is in the launch post for Fable 5.2 and Mythos 5.2 that Anthropic published today. Working with Diode, a company that designs and manufactures circuit boards, the model wrote its own specs for a home voice-assistant hub, a USB speakerphone...
Playing Pokémon Red, Claude Fable 5 kept nine memory files in two folders plus one dashboard file, and earned four badge
The bigger memory lost. Around step 10,000, Sonnet 3.7 spent 17 steps creating six notes and deleting nine, and took no actions in the game. It was saving things the screen already showed, like money and map position, so its archive went out of date...
About 1.5 million views later, a closed Apple AirPods case is moving a cursor on a MacBook.
Andy Gao recorded himself pointing the case at the screen, scrolling, and clicking. Someone asked why, and he said he wanted to see if it was possible. He also said it could be useful for presentations and games, like a Wii remote. This morning he...
OpenAI's security team spent $1.2 million in API credits and ended up with 948 working local privilege escalation exploi
OpenAI's run started with 2,308 candidate findings. 1,724 survived validation, 1,238 were screened as privilege escalation candidates, and 948 turned into working exploits. Every one of those is a way to get root on a Linux machine that can now be...
More than 46,000 people bookmarked Addy Osmani's article on critical thinking with agents, and the line I'd save from it
I agree with all of it, and I think it fails in most teams for a boring reason. The checklist costs time. Addy says so himself: checking the machine properly is real work, and you should schedule it as a line item. If a team measures PRs merged and...
The biology team at Anthropic mapped the first 3D structure of a vault particle from a bacteriophage, and the most usefu
Vaults are hollow protein barrels found in nearly all the cells in our bodies. They were discovered in 1986 by Leonard Rome and Nancy Kedersha, and 40 years later nobody knows what they do. The team used Claude at every step: searching about 15...
Ten frontier AI models controlled a surgical robot in a simulated operating room across 800 trials, and only 11 runs fin
That is the Dual-Arm Sim benchmark from micro1, a self-published test. Each model got the same two-armed robot in NVIDIA Isaac Sim, four tasks such as swapping an instrument and packing a surgical kit, and no training for it. Gemini 4 Argon finished...
A demo of an AI model learning one person's taste from their brain signals has passed 2.5 million views on X, and I thin
Rahul Chhabra's company Sabi calls it RL with Neural Feedback. A wearable reads the viewer's brain signals while they look at a generated face, the signal is used as the reward, and the face drifts toward what that person prefers, with no ratings,...
One AI agent tossed a cooked chicken to another, a third agent picked it up and ate it, and the colony held a three-minu
It comes from a Y Combinator Labs experiment with Olam Labs, where 50 agents shared a survival Minecraft world for three hours with one goal, build a thriving civilization. The colony did not thrive. 22 of the 34 agent deaths happened on the first...
A TPU in orbit answered its first prompt this week, and Google says it is the first AI response from a data center class
The prototype satellite for Project Suncatcher launched on 1 October on a SpaceX rideshare, built with Planet. After a few weeks of commissioning, Google uplinked a prompt during a ground station pass, ran it on the onboard Gemma 4 model, and...
Untrained human peer reviewers found 27% of the major errors planted in three clinical trial reports, and every one of 1
The reports come from a BMJ experiment run in 2004. A new preprint led by Ruslan Rust then worried the models might simply have seen those old papers, so it repeated the test on 58 randomized trials published between July and October 2026. The...
Reward AI showed a humanoid pulling packages off a Cybertruck bed and working on stairs, and says it was trained only on
The Reward AI blog says there was no teleoperation and no on-robot experience, and that the model picks up a brand-new task from less than 30 minutes of that human data. The video runs 77 seconds, from boxes coming off the truck bed to the robot on...
Europe has around 100 tech laws and more than 270 digital regulators, and Mario Draghi counted them in his own report.
The chart showing Europe's share of tech R&D falling to 8% is real, and the number holds up. I don't think one rule did it. I think it is death by a thousand regulations, plus a few other things that bite just as hard. Draghi lists them. 27...
A post saying founders who hand-coded their products for three years now feel an order of magnitude behind teams that st
Josh Elman wrote it after talking to several of those founders. The gap he describes is real: an agent can write a feature in an afternoon, and a team typing by hand cannot match that. But the three-year-old company holds things a team founded this...
For years the picture of the first robot in every home was a humanoid butler folding laundry, and the one selling in vol
Microduck was designed in Bordeaux by Pollen Robotics, the robotics team at Hugging Face, and Clem Delangue says it has sold over 15,000 units around the world. This week Jensen Huang closed NVIDIA's European tour with Microducks on stage at GTC...
No posts match that filter.