Even training a linear classifier is NP-hard to approximate, notes Aaron Roth
Aaroth · x · 2026-09-14
In his thread, Aaron Roth gives a concrete example: the simplest ML problem — learning a linear classifier that minimizes classification error — is NP-hard even to approximate (distinguishing 51% from 99% accuracy), and everything more general is at least as hard. This is part of his argument that worst-case complexity theory is a poor guide to what is achievable in machine learning.
Related event: Aaron Roth Slams Complexity-Based Arguments Against AI(4 posts)→
More from AGI Musings
- Yglesias to OpenAI staff: stop funding Leading The Future super PAC, or quit — AaronBergman18 · 2026-09-14
- Nikesh Arora: AI labs can't afford to slow down, coordinated pacing may be the answer — annbordetsky · 2026-09-14
- The Robotics Prisoner's Dilemma: Use Frontier Models and Your Moat Gets Absorbed — ChongZzZhang · 2026-09-14
- 1.2M STEM dissertations show government is the top funder of frontier PhD research — joshgans · 2026-09-14
- a16z's Josh Elman: AI collapsed build costs, not the cost of knowing what to build — a16z · 2026-09-14
- Big lab executives now publicly own up to AI loss-of-control scenarios — emmanuelvivier · 2026-09-14