The case against conversational interfaces

This clearly elucidated a number of things I've tried to explain to people who are so excited about "conversations" with computers. The example I've used (with varying levels of effectiveness) was to get someone to think about driving their car by only talking to it. Not a self driving car that does the driving for you, but telling it things like: turn, accelerate, stop, slow down, speed up, put on the blinker, turn off the blinker, etc. It would be annoying and painful and you couldn't talk to your passenger while you were "driving" because that might make the car do something weird. My point, and I think it was the author's as well, is that you aren't "conversing" with your computer, you are making it do what you want. There are simpler, faster, and more effective ways to do that then to talk at it with natural language.

You're onto something. We've learned to make computers and electronic devices feel like extensions of ourselves. We move our bodies and they do what we expect. Having to switch now to using our voice breaks that connection. Its no longer an extension of ourselves but a thing we interact with.

Two key things that make computers useful, specificity and exactitude, are thrown out of the window by interposing NLP between the person and the computer.

I don't get it at all.

   [imprecise thinking]
         v <--- LLMs do this for you
   [specific and exact commands]
         v
   [computers]
         v
   [specific and exact output]
         v <--- LLMs do this for you
   [contextualized output]

In many cases, you don't want or need that. In some, you do. Use right tool for the job, etc.

I don't think they give a specific and exact output, considering how nondeterminism plays a role in most models.

The diagram you're replying to agrees with this.

Does it? The way I'm reading it, the first step is LLM turning human imprecise thinking into specific and exact commands

That's true, but that is the input section of the diagram, not the output section where [specific and exact output] is labeled, so I believe there was legitimate confusion I was responding to.

To your point, which I think is separate but related, that IS a case where LLMs are good at producing specific and exact commands. The models + the right prompt are pretty reliable at tool calling by themselves, because you give them a list of specific and exact things they can do. And they can be fully specific and exact at inference time with constrained output (although you may still wish it called a different tool.)

The tool may not even exist. LLMs are really terrible at admitting where the limits of the training are. They will imagine a tool into being. They will also claim the knowledge is within their realm, when it isn't.

At inference time you can constrain output to a strict json schema that only includes valid tools.