Oxford DPhil Thesis by Andrew Trask Argues Maximal AI Capability Runs Against Centralization
iamtrask · x · 2026-09-20
OpenMined founder Andrew Trask released his Oxford DPhil pre-print Attribution-Based Control in AI Systems, arguing that many AI risks — privacy, alignment, copyright, power concentration, hallucinations — reduce to an attribution-based control (ABC) problem rooted in the overuse of addition, copying, and branching in gradient descent. The thesis synthesizes deep learning, cryptography, and distributed systems techniques into a path to ABC, claiming it could unlock 6+ orders of magnitude more data and compute.
His thread makes a bet: today's AI is a temporary "mainframe computing era"; the most competitive systems will be routed ensembles of narrow AIs assembled per prompt, with cryptography (enclaves) and distributed systems pointing to a different incentive structure. He rejects the assumption that AI power concentration is inevitable and beneficial, disagreeing with safety camps pushing further centralization.
More from AGI Musings
- LAION researcher: pausing open frontier AI carries ~10% human extinction risk this decade — JJitsev · 2026-09-20
- Founder Bindu Reddy: the world is splitting into AI-amplified builders and AI-fearful skeptics — bindureddy · 2026-09-20
- Abnormal AI founder: AI-savvy builders gain 2-10x, fear-driven resisters will lose — bindureddy · 2026-09-20
- Calibration is all you need: trustworthy probabilities beat confident predictions — zsakib_ · 2026-09-20
- Researcher mocks doom numbers: I give 10% odds of human extinction if we pause frontier AI — JJitsev · 2026-09-20
- Pedro Domingos: The amount of compute wasted on RL is mind-boggling — pmddomingos · 2026-09-20