Anthropic can already retrain Mythos with "a fraction of the compute originally needed," says former researcher Coxon
First reported by The Decoder ·
The cost to retrain advanced AI models like Mythos is rapidly decreasing, making development more accessible.
Former Anthropic researcher Jacob Coxon claims that the company has achieved a superexponential rate of algorithmic progress in training AI models. He cited a slide from Anthropic's IPO roadshow as evidence, which he believes was misleadingly presented to understate the rapid decrease in compute required for model retraining. Coxon asserts that Anthropic's Mythos model, a significant advancement, exemplifies this accelerating trend, stating that it could be retrained with a fraction of the original compute cost. He expressed surprise at Anthropic's public release of this data, which he feels demonstrates their increasing efficiency. Coxon departed Anthropic after four months, partly due to his concerns about the rapid pace of AI development and its implications.
Coxon's assertion of superexponential progress suggests that the compute resources needed to train and iterate on frontier AI models are diminishing at an accelerating rate. This implies that the barriers to entry for developing highly capable AI systems are falling faster than previously anticipated. Such efficiency gains could democratize access to advanced AI development, potentially enabling smaller players or researchers to compete with established giants like OpenAI and Anthropic.
This trend, if accurate, could also accelerate the broader AI landscape by enabling faster iteration cycles and more widespread experimentation. Coxon's call for a pause in frontier model development, intended to commoditize human-level intelligence, highlights the competitive implications of these rapid advancements and the potential for certain entities to gain significant advantages if they maintain their lead in algorithmic efficiency.
AI-written summary. May contain errors.