How are labs catching up with far less compute? Data may be the real bottleneck
biotechplays · reddit · 2026-10-02
A Reddit discussion questions the premise behind chip restrictions: if frontier AI needs massive compute, why do overseas labs keep releasing near-frontier models at a fraction of the cost? The author speculates high-quality training data matters far more than assumed, noting US data vendors like Surge AI face no restrictions and are still selling large volumes of expert training data to Chinese labs.
Related event: Reddit Debate: Training Data Quality May Matter More Than Compute(2 posts)→
More from AGI Musings
- Economist Paul Novosad: AI-written papers need their own journals, not human ones — paulnovosad · 2026-10-02
- Understanding AI launches Understanding Robots newsletter, debuting with GPT-Astra 6 analysis — binarybits · 2026-10-02
- Luiza Jarovsky taunts Uber: maybe hire back the 3,300 people AI replaced? — LuizaJarovsky · 2026-10-02
- Five agentic AI themes from The AI Conference Day 1: rent the gym, guard the playbook — sanjaykalra · 2026-10-02
- Nick Bostrom on NYT podcast: what good AI future is worth risking — round · 2026-10-02
- People fight to protect AIs from pain while ordering UberEats chicken strips — DionysianAgent · 2026-10-02