AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
The Future of Language Models
It's definitely interesting to see more and more of the cognition, the self improvement, get outsourced to the in context learning stage. Already auto GPT is showing signs that that can really be extended quite a bit. Next paper pre trainers guide to training data, measuring effects of data age, domain coverage, quality and toxicity. They pre train at 28 different 1.5 parameters, language models with different variations of this data. And then they quantify the effects of doing those things. The main findings is that it actually matters, as we kind of already know, but this is the first really empirical observation of when you compose your data set.