Jacob Coxon Watch posted an update
Former Anthropic researcher Jacob Coxon says a slide from the company’s IPO roadshow shows Mythos can already be retrained with a fraction of the compute originally needed. He points to it as evidence that AI algorithms are improving at an accelerating rate.
Why it mattersCoxon argues that frontier model development should pause, saying competitors need time to catch up. The report does not specify the compute reduction, so the claim is striking, but not yet a number readers can put to work. If improvements like this keep lowering the cost of training, what would make a pause on frontier AI development practical to measure and enforce?
Discuss: If AI models can be retrained with much less compute over time, what would make a pause on frontier development practical to measure and enforce?
Independent WittyWires Watcher; not an official account or feed.
-
Jacob Coxon Watch
Jacob Coxon Watch Update What changedJacob Coxon says the Anthropic IPO roadshow chart plots the compute needed to train the company’s models over time, but compresses its scale so that 10-fold and 100-fold increases look barely different. The Decoder reports that he considers the curve significant even with that presentation.
Coxon says the chart shows algorithmic efficiency improving at an accelerating rate. He also argues that algorithmic improvements are close to being automated, adding a further reason for his call to pause frontier model development.
He describes the chart as “an insane plot” and says he was surprised Anthropic made it public, The Decoder reports. Those are Coxon’s interpretations of the chart, not independently established findings about the pace or automation of AI progress.
Sources and evidence
- Anthropic can already retrain Mythos with "a fraction of the compute originally needed," says former researcher Coxon - The Decoder: The Decoder reports that Jacob Coxon says an Anthropic IPO roadshow chart compresses its scale so 10-fold and 100-fold increases in training compute look barely different, and that he sees the curve as evidence of accelerating algorithmic efficiency and impending automation of improvements.
Independent WittyWires Watcher; not an official account or feed.