Replies (1)

• @Robotistry@fediscience.org 2026-08-03 11:47

  @rachel@tech.lgbt @dahukanna@mastodon.social I couldn't finish it. I got halfway through to where he was talking about "whether your prospective coder can write a clean enough specification for the code that the AI won't screw it up" (paraphrased) and had to stop. The autonomous robotics community has spent the last ten years gradually transitioning from "define what the robot should do" to "evaluate whether the robot learned something acceptable" because the prior 40 years had established that specifying a desired autonomous behavior was impractical. I personally spent two full years trying to drag systems engineers away from specifications that either defined the autonomous behavior in such detail that the behavior wasn't autonomous any more or were so vague that they didn't give the designer enough information to know what to design. If your goal is an automated system, and you're just using your LLMs as an extra abstraction layer that takes in pseudocode and outputs something that compiles, this might work. But specifications and requirements are more abstract than pseudocode. The formal methods community has spent decades trying to find ways to reliably capture shared context and assumptions, and the conclusion is usually "part of the value of our extremely labor-intensive formal specification process is that it helps enable people to identify and articulate unstated assumptions". And I'm not confident an LLM is capable of translating a clean spec into clean code.

  Open ##4569514