Why did Kaplan et al. get a different answer from Chinc...

The discrepancy arises from the training schedule methodology. Kaplan et al. utilized truncated cosine schedules, which inadvertently made model size appear more valuable than it actually is because longer training runs benefited from schedule decay. Chinchilla corrected this by using complete cosine schedules matched to the specific length of each training run. This revealed that the exponents for parameters and tokens are roughly equal (β ≈ α), demonstrating that token count is roughly as valuable per unit of compute as parameter count.