Hacker News AIT2·Models
62 pts
Dwarkesh Patel has sharply revised his timeline, arguing that AI matching the best human AI researchers may be only a couple of years away, possibly months. His central argument is compute scale: OpenAI's end-of-2027 fleet could support roughly 2.8 million GPT-3-training-run-sized experiments per year, and even with a third of the fleet dedicated to experiments, 10,000 agents could each run one every four days.