Korva is a scroll of the internet’s best ideas.
AI research is bottlenecked by experiments
Why math differs
Math is bottlenecked by thinking really hard, which models are good at. AI research differs, because someone has to run experiments, and training new models or getting results takes a while. Experiments run serially, and they need GPUs.
Compute still matters
A comparison shows the weight of compute. Take OpenAI with a hundred times less compute but staffed by the most brilliant people in the world. It would make less progress than it does today, probably a lot less, though not a hundred times less.
A speed-up, not an explosion
The likely result of AI-driven research is a significant speed-up, not an overnight intelligence explosion with progress a hundred times faster. A speed-up of three times would already be massive, since the current pace is an exponential. Estimates range from fifty percent faster to ten times faster.
Why can't researchers skip the training wait?
Training updates a model's weights step by step, and each step needs the weights from the one before, so the steps cannot all run at once. Runs at frontier scale take weeks to months even on huge clusters. Meta's report on Llama 3.1 said its 405-billion-parameter model trained on about 15 trillion tokens using up to 16,000 H100 GPUs.
How can small experiments predict big ones?
Labs fit curves on small training runs and extrapolate to much larger ones, which cuts some of the guessing before an expensive run. This works because a language model's loss falls smoothly with scale. In 2020, Jared Kaplan and colleagues at OpenAI reported that loss follows power laws in compute, data, and parameters across more than seven orders of magnitude. The paper was titled Scaling Laws for Neural Language Models.
How did the intelligence explosion idea start?
British mathematician Irving John Good wrote in 1965 that an ultraintelligent machine could design even better machines, and that an intelligence explosion would follow. He called such a machine the last invention humanity would need to make. Good had worked alongside Alan Turing at Bletchley Park during the Second World War.
More from Dwarkesh Podcast
