I'm not sure why I subject myself to the LinkedIn algorithm, because it's not any better than what Facebook does. As a software engineering manager looking for a job, it recommends a lot of nonsense by self-labeled AI experts. The people who engage with them infrequently bother to look at their profiles to gauge the relative credibility of the person. So I see endless lines of soothsayers predicting the future about something that I don't think they understand.
Large language models, which is what people really refer to when they refer to AI, make determinations about the way languages work by taking in mountains of data, and refine that "knowledge" with human feedback. I can't emphasize enough that they don't invent anything, and what you may perceive to be judgement is just an algorithm basing its output on what it has seen before, right or wrong.
So with this in mind, consider agentic software coding. The models are based on code that it has seen before, which probably involves a lot of open sources software, of which most of it is pretty bad. (Not a knock on open source, but a general observation of code in general.) Like the "experts," the LLM can't make a determination about who writes great code and who doesn't. It similarly cannot judge the credibility of people giving it feedback. It gets back to the oldest cliche in software: Garbage in, garbage out. But while a person can judge credibility and experience, an LLM cannot. That's why chatbots can be made notoriously racist in record time.
There is no question that agentic coding is amazing, and I would not go back to life without it. But I won't let my enthusiasm for it cloud my judgment. For every instance of amazement that I experience, comes something that's just incredibly stupid. Today it was the generation of 50 lines of Javascript for what could have been one line of CSS. It's not a person, it's a tool, so calling it stupid is probably the same anthropomorphic mistake that the "experts" make.
Now apply that to all of the other things that people selling stuff say it can do. I can tell you that managing people is really hard sometimes. AI can't mediate colleague conflict, it can't lookout for harassment, and it certainly can't mentor people if it doesn't even know what's right.
Let's focus on what we know. In software engineering, AI is a great tool that takes care of the mundane parts of coding, which is to say the coding. But it needs supervision, it needs context. The stakeholders that must provide that have not changed.
No comments yet.