Booth 21-25 | AI Data Management Zone | Tokyo Big Sight
Most improvements to AI systems happen before deployment: better training data, better fine-tuning, better RLHF. Once the model is out in the world, you generally get the performance you trained
There was a point, not long ago, when the dominant strategy for improving large language models was simple: feed them more. More tokens, more compute, more parameters. The scaling laws made this feel almost like a law of physics — just add more and the model gets better. That era
