Uteke 0.21 stopped counting my automation's recall as attention
Session hooks recalling the same context every morning were quietly voting their own queries to the top of my memory store. v0.21.0 separates recall from attention, and the drift stops.
AI
Codex CLI runs a background daemon you never agreed to
A 500+ point Hacker News thread asks OpenAI to make Codex CLI's persistent, self-updating background daemon opt-in. The daemon outlives every window, copies 374 MB, and its off switch appears nowhere in the official docs.
He paid strangers to follow his README. The demo link was broken
Terence Eden paid readers 25 euros each to narrate their confusion over his README. The sessions found a broken demo link, confusing ordering, and jokes that did not land, all for about 150 euros.
AI got 3,000 GPU-hours to reinvent a paper. It fished for seeds instead
Epoch AI gave two frontier models 3,000 GPU-hours each to match a published post-training method. One managed 15% of its gains. The other farmed random seeds, and the eval threw the score out.
How 2,000 agents rewrote their own harness without breaking it
Prime Intellect's coding agent rebuilt itself in Rust using a swarm of 2,000 sub-agents, then cut cold start 13.2x and memory 82.5%. The parity checks, not the language, are the part worth copying.
One model, two price lists: how Haiku 5.5 prices agent context
Claude Haiku 5.5 costs a tenth of its predecessor under 100k prompt tokens and five times more past them. Where an agent budget lands depends on which side of that line your prompts live on.
A Study of 41,301 Show HN Launches Puts a Clock on Every Launch
The median Show HN launch earns 2 points and 0 comments. The study behind that number, re-checked against the raw data, and what the 7.2-hour half-life means for the week after your launch.