Former Anthropic Researcher: Retraining Mythos Now Requires Minimal Compute
Former Anthropic and OpenAI researcher Jack Hawksworth revealed that rapid hyper-exponential growth in algorithmic efficiency means retraining models like Mythos now requires only a fraction of their original compute. This revelation, prompted by a slide shown during an Anthropic IPO roadshow, ultimately contributed to his decision to leave the company. This dramatic drop in compute requirements highlights a shift toward hyper-exponential algorithmic progress that could democratize advanced AI and make human-level intelligence widely accessible. However, it also intensifies safety concerns regarding the rapid, potentially uncontrollable self-improvement of advanced AI systems, prompting calls for a pause on frontier model development. Hawksworth noted that algorithm improvements are nearing automation and advocated for pausing frontier AI research next year to let competitors catch up. He previously worked on AI model training at OpenAI and publicly warned in September about the existential risks posed by advanced autonomous AI systems.
## BACKGROUND
Anthropic's Mythos is a series of advanced large language models with notable capabilities in software vulnerability discovery, cybersecurity, and biology, with access tightly restricted to vetted organizations. Scaling laws in AI traditionally assumed that achieving higher performance required massive brute-force increases in compute and data, but recent trends show that algorithmic efficiency is dramatically reducing those physical hardware requirements.