AI Labs Are Weaponizing Machine Learning to Accelerate Their Own Research
Source: Don't Worry About the Vase
METR's research on Anthropic shows that AI systems are beginning to compress the feedback loop between capability gains and R&D velocity—the cost and time to train the next generation of models is dropping faster than the previous generation predicted. This creates an economic race where labs with superior compute and pre-trained models gain compounding advantages, pushing the marginal cost of capability improvements toward zero for market leaders. The stakes are economic: whether frontier AI development becomes a winner-take-most competition decided by whoever reaches critical self-improvement velocity first.