TheBitCast

Your Small AI Model Can Now Drive A Computer

Podcast: AI Breakdown ยท Varun Sirohi

Opublikowano: 2026-10-31

๐€๐›๐จ๐ฎ๐ญ ๐Ÿ’๐ŸŽ๐ŸŽ,๐ŸŽ๐ŸŽ๐ŸŽ ๐จ๐Ÿ ๐ญ๐จ๐ค๐ž๐ง๐ฌ ๐ ๐จ๐ญ ๐ง๐จ๐ฐ๐ก๐ž๐ซ๐ž.๐€๐›๐จ๐ฎ๐ญ๐Ÿ๐ŸŽ,๐ŸŽ๐ŸŽ๐ŸŽ ๐ ๐จ๐ญ ๐ข๐ญ ๐ฐ๐จ๐ซ๐ค๐ข๐ง๐  ๐ข๐ง ๐ญ๐ฐ๐จ ๐ฐ๐ž๐ž๐ค๐ฌ. ๐Ÿ˜ณ A developer set out to port the TypeScript compiler to Rust with AI. The same job, a twentieth of the spend, and the cheaper attempt is the one that worked. The README adds a line that should make any owner sit up. He has never read a line of this code. This week on ๐“๐ก๐ž ๐€๐ˆ ๐๐ซ๐ž๐š๐ค๐๐จ๐ฐ๐ง: ๐€๐ˆ ๐ˆ๐ฌ ๐Œ๐จ๐ฏ๐ข๐ง๐  ๐ญ๐จ ๐˜๐จ๐ฎ๐ซ ๐ƒ๐ž๐ฌ๐ค Nobody outside the project has confirmed how much human steering stood between the failure and the result, which is the honest footnote on every spectacular agent story. Inside the episode: ๐Ÿ”ด A small model scoring 72.4% on the offline OSWorld computer use benchmark, up from 15.7% for its predecessor, with the larger model still ahead at 83.9% ๐Ÿ”ด Input priced at 10 cents per million tokens against $1 for the previous version, a tenth of the cost ๐Ÿ”ด 39.2% on a command line benchmark...