Oxford DPhil thesis argues Attribution-Based Control could fix hallucination, alignment and power concentration at once
iamtrask · x · 2026-10-19
Andrew Trask (OpenMined) released a DPhil thesis pre-print from Oxford arguing that AI's major risks—privacy, value alignment, copyright, power concentration, hallucinations—reduce to one technical root: overuse of addition, copying, and branching in gradient descent, i.e. lack of Attribution-Based Control (ABC), meaning users can't know which data points inform which predictions.
The thesis synthesizes recent work in deep learning, cryptography, and distributed systems into a path that reduces these operations and lets data holders communicate directly with insight seekers, replacing centralized data collection—and claims this unlocks 6+ orders of magnitude more data, compute, and capability.
It also makes a geopolitical argument: democracies must either centralize AI decision-making or cede advantage to authoritarian nations that can marshal national-scale data and compute, motivating a technical route that skips the centralized phase.
More from Safety
- Safety researcher critiques framing of 'deliberately training models' findings — BlancheMinerva · 2026-10-19
- Suhail: every org needs a frontier cyber model now; guardrails refused mid-incident — inductionheads · 2026-10-19
- Underwater AI data center planned in Alaska's Cook Inlet, home to 330 endangered belugas — anselm · 2026-10-19
- X withholds post in Germany based on local laws, screenshot sparks debate — djcows · 2026-10-19
- Ethan Mollick: don't give agents your bank account, but secure design beats bans — emollick · 2026-10-19
- Before uploading docs to AI, check hidden comments, edits and metadata — alifcoder · 2026-10-19