Why discriminative models suit real work better than a chat interface
The author argues that forcing generative LLMs into discriminative tasks like classification and extraction is costly and unreliable, and that Jev brings discriminative models back into focus.
One reason why the economic impact of AI isn't as big as many would think it should be is that chat is just the wrong interface for a lot of work and also the wrong premise to train models for. It is super hard to automate stuff with LLMs *reliably*, because they are trained to respond in a way that sounds natural to us humans. And that is fine for something like ChatGPT and maybe for a lot of knowledge work, but for a lot of tasks, it just is needless overhead. Think about what LLMs are actually build for: generating text. But that is often not what we are actually interested in. We may be interested in classification or information extraction. But since for the last 3 years we thought AI == LLMs, we shoehorn a generative model like ChatGPT into doing discriminative tasks. And it kinda works, but just kinda. It's like putting a nail in the wall using a sledgehammer. Since we schlepp the whole generative baggage with us, we pay the price for it in the form of huge hardware costs, slow response times and output we have to artificially constrict to what we actually care about. And unfortunately, because LLMs got so much attention the last years, research around discriminate models like encoder-only transformers or classic ConvNets for computer vision has come to a halt (relatively). Maybe that is why I am so excited about Jev by @typesafeai - because it brings new attention to discriminative models, which may be arguably much more useful in the real world. thanks for coming to my ted talk.