I really suspect this is how the plateau of productivity will look for machine learning. It will be all about building synergy with conventional algorithms, which can provide the rigour, transparency and reproducibility trained models can’t. Computational efficiency sometimes, too, although maybe not in this case.
Yeah, that’s part of why I think that. There’s also just the alignment issue that no amount of training will fix. At the end of the day, an LLM is a very smart internet simulator, you treat it like something else at your peril, and training it to be something else is very much an open problem.
I really suspect this is how the plateau of productivity will look for machine learning. It will be all about building synergy with conventional algorithms, which can provide the rigour, transparency and reproducibility trained models can’t. Computational efficiency sometimes, too, although maybe not in this case.
It seems finding more data to scale up LLMs is a bottleneck too.
Yeah, that’s part of why I think that. There’s also just the alignment issue that no amount of training will fix. At the end of the day, an LLM is a very smart internet simulator, you treat it like something else at your peril, and training it to be something else is very much an open problem.