Rohan Anil, co-founder of Core Automation and a former member of Anthropic's pre-training team, said on the MTS podcast that frontier AI's real bottleneck is that models cannot learn from their own mistakes while running, and he expects to deliver algorithms letting models learn at test time within the next year, according to a 2026-11-05 report. He argued the industry is pouring hundreds of billions into data centers on the premise that scale is the answer, while the more consequential unlock is algorithmic: a model that keeps learning after deployment and a training method that can stitch together scattered clusters instead of demanding one mega-campus. Anil said the most valuable computation is frontier training, with the top two models mattering and everything else mattering less. He said money isn't the constraint but collocation is, with hundreds to thousands of GB300s and B300s scattered across locations; a model trained across separate "compute islands" can reach the same quality as one trained on a single large cluster if the training algorithm is redesigned. The proof point used two Blackwell-based hardware configurations with different parallelism schemes, slightly different numerics, and different batch sizes, yielding results similar to or slightly better than an efficient frontier-training scaling baseline. He said frontier training on ordinary consumer CPUs would hit memory or latency bottlenecks his work doesn't solve.