"tell me what you're thinking. leave nothing unsaid, and remember, make no mistakes."
Mostly Harmless CoT: on the reliability of chain-of-thought.
1
84
Three years ago, using AI at work was unusual. Four independent studies now put it between 39% and 88%, depending on who you count: organisations, businesses paying for it, euro area workers, or US workers using it weekly. Different groups, different definitions, all moving the same way.
Keeping fusion plasma stable is difficult because it is hotter than the sun and can fall apart in a few thousandths of a second. Princeton’s PACMAN system was tested on the actual DIII-D machine in San Diego. It watches the plasma and issues new commands roughly every 20 milliseconds while a shot is running. In one case it flagged a dangerous magnetic instability a fifth of a second before it would have appeared and adjusted the plasma in time to stop it. The same system also directed all six of the machine’s microwave heaters together. Researchers still choose the targets, and the machine’s own safety limits always have the final say.
Source: Rothstein, Farre-Kaga et al., Nuclear Fusion 66, 076050 (2026) · Princeton PPPL and Princeton University · five experiments at DIII-D
For decades, whale clicks looked like random Morse code. Project CETI ran nearly 9,000 Eastern Caribbean codas through machine learning and found they are built from four reusable parts (rhythm, tempo, rubato, and ornamentation) forming 156 distinct sequences. This year the team showed the clicks also contain two vowel-like patterns (a and i) that behave like human vowels.
Orcas show similar structure in 61 hours of recordings. Their first language model can already predict when a whale is about to dive.
Source: Project CETI (MIT CSAIL & UC Berkeley) · Nature Communications 2024 · Proceedings B April 2026 · orca analysis spring 2026 · WhaleLM Oct 2026






