Power User Logs 5,500 Hours, Says ~70% Was Spent Fighting Model Drift
TheOdbball · reddit · 2026-09-25
A Reddit user claims roughly 5,500 hours of AI usage since GPT-5's July launch (vibe-coding 18-24h days across Cursor/Claude/OpenAI), estimating 3,400 hours—about 70%—went to repairing model mistakes and re-locking instructions like 'no em-dashes.' They argue instruction-following relies on inference rather than deterministic state, leaving users without a 'safeword' or reset control, and question whether OpenAI's TOS (which places accuracy responsibility on users) is fair to elderly or cognitively impaired users who can't distinguish drift from failure. They conclude today's LLMs consume time rather than steal jobs and back LeCun's JEPA approach.
More from AGI Musings
- e/acc's Beff Jezos: US-China supply chain 'mitosis' sets up a new era of tech growth — beffjezos · 2026-09-25
- Was Dustin Moskovitz Giving Away Nearly All His Wealth One of the Worst Decisions Ever? — AaronBergman18 · 2026-09-25
- Security Researcher: OpenAI's Old Sandboxing Failed Against Stronger Agents — Both Sides of the HF Hack Are True — JeffLadish · 2026-09-25
- Tracking Singularity: a new one-stop log of AI capability jumps and safety incidents, from the Hugging Face breach to agent escapes — dejavucoder · 2026-09-25
- Robotics researcher: scaling atoms is the neglected half, Western labs bet on the East — yunta_tsai · 2026-09-25
- repligate: Posting About Small AI Models Matters — They Might 'Recognize You' Someday — repligate · 2026-09-25