DailyChatThursday · Sep 24

Agents need more than a model.

5 stories · 3 minute read
01 / PERSONAL AGENTS / COMPANY ANNOUNCEMENT

Muse reaches for glasses, but hardware arrives in stages

Meta announced Muse for its AI glasses in the coming months, plus the pocket-sized Muse Charm, with more details promised later this year. Its separate VR Glasses are due in 2027; these are announcements, not products all shipping today.

Read the source ↗ about.fb.com
02 / AI FOR SCIENCE / COMPANY RESEARCH

Claude flags an enzyme system whose function remains unknown

Anthropic reports that Claude identified previously uncharacterized features around a known enzyme, revealing an associated repeat-array system. Human scientists performed the lab work. Its biological function remains unknown; the early finding does not establish a new gene-editing tool.

Read the source ↗ www.anthropic.com
03 / SOFTWARE / ENGINEERING CASE STUDY

More generated code moves the bottleneck into testing

Conceptual CI flow: parallel stateless listeners write a shared journal; per-test history feeds the test selector.
Conceptual illustration of Anthropic’s reported test-selection redesign. A separate consumer rolls the journal into per-test history. The company says this architecture costs more to run but is easier to scale.

In a September 14 engineering account, Anthropic reports a 25-fold rise in CI jobs over six months. It scaled test selection by moving history out of a single process into shared storage, allowing stateless workers to run in parallel.

Read the source ↗ claude.com
04 / AI FOR SCIENCE / FOUNDER PERSPECTIVE

Faster analysis still needs experiments and human checks

Endura cofounder Adrian Sanborn distinguishes scaling physical experiments from using AI to choose and interpret them. He reports broader disease-target screening and faster internal tooling at his company; primary-source checks and human diligence still guide selected programs.

Read the source ↗ www.latent.space
05 / OPEN MODELS / COMPETITION

Gemma 4 agents will be judged by working patches

Kaggle’s competition, launched September 23, asks entrants to turn Gemma 4 into a coding agent. Patches are scored by repository tests. Entry closes November 25; final submissions are due December 2, both at 23:59 UTC.

Read the source ↗ www.kaggle.com