9:16 PM

Today in brief.

Memory that works: Claude dreaming + Stanford Memorilla

Two memory systems landed the same day: Claude now reorganizes its notes offline, and Stanford packs a whole document store into 16 tokens.

OpenAI shipped a 100-task wet-lab benchmark with Ligo, and Anthropic posted how AI is starting to write and revise scientific models as code.

Village's Atlas says open-weight models are the best buy at only 3% of budgets now, down from 41% in June.

Human-guided DeepMind agents uncovered two weaknesses in old 1024-bit RSA, a concrete example of AI speeding up cryptanalysis.

Runway added GPT Image 2.6 with faster gens, GIFs, SVG export, and tighter character edits.

Underdog's Woof is a ~32MB text-to-speech model that speaks offline in a browser demo, open weights included.

MIT Tech Review watched a humanoid clean an apartment, and Claude drove a robot arm to paint a watercolor while mixing its own colors.

LocalLLaMA's weekend shelf: a 0.1B embedding model that beats much larger ones in a user test, plus Nous Arete-27B.