Can we fix AI? Peter Bloem's new essay examines data, copyright and synthetic data paths
pbloemesquire · x · 2026-11-16
Researcher Peter Bloem published "Can we fix AI? Part 1: Data", arguing that models like ChatGPT and Claude are technically strong and improving fast — the real problem is public resentment, yet safety-critical coding now arguably requires AI review to catch exploitable bugs, forcing use of a morally contested tool. The essay systematically explores whether the objectionable parts can be removed via data: model/output copyright, compliant corpora like the Common Pile and its limits, synthetic data and self-play, fine-tuning data, in-context learning, data laundering, and open routes such as Apertus and Common Crawl. His verdict: mostly "we don't know, but we can find out."
More from AGI Musings
- antirez defends Huang staying CEO: investigate real wrongdoing before demanding heads — antirez · 2026-11-16
- After Cognitive Offloading: Humanity Revalues What Intellectual Work Means — Hub_Pli · 2026-11-16
- Safety as capability: why safer AI models get released sooner, and the case for open-sourcing safety evals — anpaure · 2026-11-16
- tszzl amplifies call: forecasters shocked by AI progress should be shocked by alignment progress too — tszzl · 2026-11-16
- Richard Dawkins asks Prospect: Can a machine be conscious, and how would we know? — connoraxiotes · 2026-11-16
- 'We Probably Make It to 2028' Now Counts as the Optimist Take on AI Risk — Puzzleheaded-King584 · 2026-11-16