AI Excels at Coding but Fails at True Knowledge Work, Says Statistician
kareem_carr · x · 2026-08-01
Statistician Kareem Carr points out a counterintuitive phenomenon: AI is increasingly good at coding and math because these fields reward repeating existing arguments and solutions. However, it is increasingly bad at general knowledge work.
Drawing from his own frustrating experiences trying to get LLMs to write YouTube scripts, Carr notes that the most valuable work typically departs strongly from consensus. Current models fail not only at generating novel, interesting topics—defaulting to widely reported ones—but also at producing satisfactory content even when given a specific prompt.
Related event: Tech Blogger Tests Top LLMs for Video Scripts, Finds Them Failing(4 posts)→
More from AGI Musings
- Paul Graham: Web services unusable by AI agents will go out of business — EdenEmarco177 · 2026-08-01
- Peter Diamandis: Graduates Need Portfolios, Not Just Transcripts — PeterDiamandis · 2026-08-01
- AI for Science projected to capture 20% of global R&D spending by 2030 — basedjensen · 2026-08-01
- Researcher Reflects on AI Evaluation Limits Without Frontier Model Access — sethlazar · 2026-08-01
- Frontier Model Oversight Needs a Dual Track: High Cyber Risks vs. Slow Work Disruption — _akpiper · 2026-08-01
- Berkeley's Agentic AI Summit 2026 Focuses on Ecosystem and Security — ben_burtenshaw · 2026-08-01