Every opens its Frontier team's AI experiment log with code and writeups
danshipper · x · 2026-11-14
Every's Frontier team publishes all its AI experiments at frontier.every.to — weekly logs with raw code and writeups. Recent entries include: map-driven autoresearch (a parent agent picks 5 of 104 research paths to run in parallel); benchmarking Gemini 4 Hydrogen on sentence-level vs whole-draft editing; Heisentool, a fake Slack written by an agent from observed tool calls for safe agent testing; a storyboard-first video editor; a four-level human-in-the-loop framework where only coordination resists automation; and the argument that AI's bottleneck has shifted from intelligence to human attention.
More from AGI Musings
- Brain has 200T synapses yet forgets a penny; a 20T-param LLM forgets nothing — fleetwood___ · 2026-11-14
- Jradoff: The 4x Efficiency Gain Isn't the Point—Cheap Prototypes Let Studios Bet on MMOs Again — jradoff · 2026-11-14
- Would China chip export controls actually give Anthropic pricing power? Reddit says no — supplykatusa · 2026-11-14
- Isomorphic Labs: AI could move drug discovery from slow lab work into software — VraserX · 2026-11-14
- Researcher: don't rush to frame FAA halt as an AI story before facts are in — peterwildeford · 2026-11-14
- DeepMind paper: most workers need 100-120% salary replacement to quit their jobs — joshua_saxe · 2026-11-14