> replacing what use to take a specialized ML team often months to do (create a dataset, train a model, serve it themselves). with a single prompt to codex with access to the @modaicdev API, i trained and served a custom decision model that outperformed Jev by 22.2% and hardly degraded production latency as the models get better at RSI, the idea of "zero shot" will look more and more like taking a pre-trained model shaped for your task and fitting it to your data orrrrrr you can wait 5 months for a lab to release a slightly less jagged model that performs 2 percent better
回复1转发0喜欢3收藏2浏览373
入选 11/21 每日精选 · 当天按质量分排序看当天 →
互动数据为采集时快照

Erik Spock Gafni (hiring!)