[Former Anthropic Researcher Warns Again: Algorithmic Advancements Are Accelerating, Training Stronger AI May Not Require Pile Up Computing Power] Beating AI News: In September of this year, Jacob Coxon, who previously worked on pre-training research at OpenAI and Anthropic, resigned, angrily denouncing the two companies for rushing towards self-improving superintelligence, essentially "gambling with our lives." These remarks drew calls from US lawmakers for stronger regulation, and even Musk suspected he was engaging in regulatory propaganda. Now, Coxon has posted a chart from Anthropic's IPO roadshow, claiming it was one of the key reasons for his resignation. "This chart is insane; I didn't expect them to release it publicly." Anthropic originally used this chart to showcase its technological advantages: as algorithms improve, training models with equivalent code capabilities requires less and less computing power. However, Coxon noticed that although the vertical axis of the chart uses a logarithmic scale, the curve is still getting steeper. This means that if the data is accurate, the rate of improvement in training efficiency is itself accelerating. He argues that Anthropic's algorithmic progress has exhibited "super-exponential growth." Mythos may simply be a stop on this steep curve. Anthropic can already train models of the same caliber as Mythos with far less computing power than initially could. With the same computing power, it's possible to train even stronger AI. Furthermore, improving training algorithms doesn't necessarily require brilliant researchers constantly coming up with groundbreaking ideas; this kind of work is nearing automation. Anthropic co-founder Tom Brown also mentioned in a pitch that Claude hopes to discover methods to improve training efficiency on his own in the future. New methods can be used to train stronger models, and stronger models can then continue to seek new optimization methods. However, this graph doesn't disclose specific computing power multipliers; it only shows three data points, and the locations are marked below for illustrative purposes only. Coxon advocates pausing cutting-edge AI research in 2027 to allow competitors to catch up. This would gradually make human-level AI capabilities a common commodity available to everyone, breaking the advantage that OpenAI and Anthropic have built on leading models. [Original Link]