On Monday, July 24, 2023, at 8:09 AM, stefan.reich.maker.of.eye wrote: > If compression is superior to language models, when are we getting a > compression based chat bot? :)
It's only a way to evaluate AI's prediction score. It doesn't change how you, or what you, or if you, or why you, or what or how your AI works or should work. Though, to do the test efficiently, it can or does change how your AI code is some amount yes, but it's only because you are constantly predicting and using the AI and not "just" training it. With Perplexity eval, you train it only, then once done, predict only a few thousand items, and not 1GB. But to do the compression test also, you need to use the arithmetic coding thing to do the compression part that measures the score, which means you need to not really predict a few thousand items and only say 256 instead, which is a byte or so long if I recall here right (lol). But that's if you can't, maybe you can find a way... ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T919d1153741947b5-M1b77e4ff2ac84928b0018c15 Delivery options: https://agi.topicbox.com/groups/agi/subscription
