What do you think of AI employees?

I've been doing agentic workflows for a while. With a special variant of evolutionary search on top of them. I also sinned of thinking about selling them as an "AI employee", there are a couple of issues though. 1. They can't learn fast enough. LLMs are "frozen neurons", they don't have any means of online learning, and you can't fine-tune them fast. 2. QA is nightmarish. Their outputs are non deterministic, even if you tune the temperature. This means you're not doing a 1:1 test, you're doing a 1:N test. Where you don't know N. 3. Inference is not fast enough (yet). You can read stuff with 10 tks, but you can't have an agent do "smart work" at those speeds. Of course there's providers that do many times that, so it's less of an issue at certain thresholds. --- We have to sell reliable products at the end of the day or people will inevitably complain, as we can see with the many angry Devin customers. I've been trying some techniques to solve #1 and I think we need to develop some for #2, if we can manage QA properly and deliver, I think reliable AI agents could be a thing. Some people might want to call them employees but it's just an OODA loop on an LLM.