2024-11-16T18:48:18.868458 | 🔗
We are probably at the optimum already in terms of the mean intelligence of foundation models. Probably not possible to curate data too much more without losing generalizing power. Now the problem is post-training to sample as far along the distribution as possible.