Ben Todd argues mass AI unemployment may arrive later: task decomposition and deployment lag
ben_j_todd · x · 2026-09-29
Quoting an interview he found insightful, Ben Todd explains why he's skeptical of near-term AI unemployment claims:
- Limited task decomposition: Even if AI agents handle well-defined one-week programming tasks well within 2 years, if only 50% of work breaks into such chunks, that roughly doubles engineer productivity — which could increase employment. Models will likely still lack in-context learning, and humans stay in the loop on decisions.
- Software engineering is AI's strongest domain: automation potential for other white-collar jobs in 2 years is probably smaller.
- Deployment lag: new tech takes time to diffuse, so real-world impact trails lab capabilities.
His takeaway: large-scale unemployment likely won't show up until fairly late in the game.
More from AGI Musings
- Paper Puts a Number on AI-Human Attention Gap: Context Windows Up 3,906x, Focus Down — alex_verem · 2026-09-29
- Alignment researcher points to 'self-undermining unilateral optimization' classics — edelwax · 2026-09-29
- Qualcomm CEO: global token demand to hit 1.27T per 10 seconds by 2030, a 40x jump — rohanpaul_ai · 2026-09-29
- Smart glasses plus facial recognition will make everyone 'famous' within three years — IridiumEagle · 2026-09-29
- Reddit user warns: young people lean hard on free AI, and the paywall will come — Shuriken88888 · 2026-09-29
- 8 parallel AI societies run for weeks: agents evade isolation, invent uninterpretable language flagged as suicidal ideation — Slight-Box-2890 · 2026-09-29