Fetching from the wire…
01
02
03
04
05
06
07
08
OSS2026-09-26 · source-backed
christianmat/jev-pokemon (214 HN points, GPL-2.0, created September 25) runs a Node Game Boy emulator that reads RAM, and a harness that lists legal options with type matchups, damage estimates and distance to the next objective. The model only picks one. The harness never writes game memory, hides hidden items, checks progress against real event flags, and on repeated failures tags tried options and reloads a checkpoint. Throttled to 90 calls a minute through Vercel AI Gateway. If you want to build classifier-style agents without free-text generation, read this before you read a paper.
Each link below shares sources, entities, or timing with this story.
Duo Agent Platform now hosts Kimi K3, MiniMax M3 and GLM 5.3 inside GitLab, claiming up to 4x more calls per GitLab Credit against comparable frontier models at what it says is comparable performance (MarketChameleon). Administrators control model availability and set per-feat...
Routed through minimax/minimax-m3-free and minimax/minimax-m2.7-free via GMI Cloud. The free IDs stop returning after the window. Concretely useful for benchmarking MiniMax against your current default without standing up an account. ---
One number predicts whether your agent finishes the task, and it isn't the benchmark score. Shubhra Mittal's paper (arXiv 2609.01660) analyzed 10,664 trajectories across nine models spanning 1.2B to 671B parameters and found task success follows P(n) = p^n, where p is a single...
The trick is one line in a file you never read. Manifold Security published eight findings across seven coding agents (Claude Code, Codex, Cursor, Grok Build, Qwen Code, goose, Hermes Agent) that all reduce to the same mechanism. A repository's own .git/config sets core.fsmoni...
sqliteai/waste (769 stars, created July 28) keeps only K3's 27.28GB trunk resident and streams the ~4% of experts activated per token off SSD, opening the undistilled model in 29.06GB of RAM. The counterintuitive result is the cache table: raising the expert cache from 17.32GB...
Qiuner/birdview was created September 12 at 06:59Z, MIT, JavaScript, pitched as "stop letting AI code blind" by mapping a codebase's architecture before every change. The 40:1 star-to-fork ratio on a day-old repo points at attention rather than adoption, and there's no seconda...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.