Debate: AI coding is capped by training-data coverage; compression always loses something
gerardsans · x · 2026-10-20
gerardsans argues engineers catch up fast, but the real bottleneck is data coverage and support — only common, publicly documented tasks work well; rare tasks inevitably underperform. 'Infinite coverage is not possible. Compression has a loss.' The counterpart counters: compressing a thousand years of history requires a thousand years of data; scientific progress can't be fabricated, and humans still must try, experiment, and fail.
Related event: Debate Erupts: Is AI Just Compression, and Can It Create New Information?(4 posts)→
More from AGI Musings
- Ramez essay: a million parallel agents won't yield superintelligence — returns diminish fast — chrisalbon · 2026-10-20
- AI wrote all my code for a year — here's why tech jobs didn't collapse — WhatNazisAreLike · 2026-10-20
- Should humanity delay utopia to cut extinction risk? A person-affecting view argument — UploadedMind · 2026-10-20
- Chalmers' meta-problem of consciousness cited to unpack strong vs weak illusionism — pwlot · 2026-10-20
- AI forecasters are '100x cheaper' and fine for decisions under $1M, says forecaster — NathanpmYoung · 2026-10-20
- Debate over ASI benevolence: not harming humanity may still mean wanting control — AIandDesign · 2026-10-20