2 Просмотры

  • Спасибо! Поделитесь с друзьями!

    Вам не понравилось видео. Спасибо за то что поделились своим мнением!

aii
Издатель
I read a new PNAS Nexus paper this week that finally put words to something that has annoyed me for a while. People argue about whether an LLM is "just autocomplete" or an early mind, and both sides are arguing about the wrong thing.

The author, Zhicheng Lin, separates the model from the system built around it. Training is one thing. What you actually use is a product with tools, retrieval, memory and decoding settings on top. That is why a model hooked up to biomedical databases scored 0.83 on a genomics benchmark while the same kind of model without tools landed between 0 and 0.44. The predictor did not change, the system did.

Same story with memory. Weights, the context window and external storage are three different places, and most privacy worries live in the last one.

The clinical trial example stuck with me too: same GPT-4, but a redesigned workflow lifted doctors from 75% to 82-85% accuracy.

Which of the six misconceptions did you believe until now?
#LLM #AI #MachineLearning #AIResearch #TechMindAI
Категория
AI Инструменты
Выполните вход или зарегистрируйтесь для отправки комментария.
Комментариев нет.