“AI labs are deviating from the compute-optimal training recipes described in the Chinchilla paper to optimize for inference budgets instead.”