Whither AI?

I am hardly the first person to comment¹ on this – I am given to understand AI has been a topic of some interest to many for a few years now. I’m sure I’ve seen, and possibly even re-tooted things about it in fact. I’m afraid I just don’t keep up.

But recent experiences reviewing for a couple of systems/networking venues has led me to feel I need to ask: WHY? More pointedly, why does the following seem like good motivation for a research paper?

1. There is a complex and important task that currently requires considerable expertise to carry out because it is important to be precise and get it right.
2. The task in question can be described imprecisely using natural language by non-experts.
3. AI (inevitably, some large-language model) can take that natural language description and, after training, produce some output that is stochastically like unto what an expert might produce given the same underlying problem, having brought to bear their expertise.
4. Thus we build an AI that can take the non-expert’s imprecise description and show that sometimes the output it produces is not so wrong as to fail some ad hoc tests of utility that we introduce.

Based on things I’ve recently reviewed “not so wrong” above means “error rate of no more than 25—30% when taking expertly generated natural language prompts as input”. Which is to say, probably not the sorts of input prompt that a non-expert might produce.

Network configuration and management is the domain I’ve seen this argument made in most recently. Which seems quite strange to me because I always thought that a 25% error rate in configuring, e.g., your enterprise network security perimeter would be bad. But apparently not if it’s done by an AI.

More generally, why do we want to build tools that allow untrained experts to do a job when mistakes are high impact, it requires a trained expert to detect those mistakes, and those tools by design only produce statistically valid output? An error rate of once in a blue moon is categorically worse than a zero error rate if the error involved can leave your entire digital estate open to compromise.

If the big issue here is that experts sometimes make typos when editing the configuration files, maybe building some domain-specific languages or better user interfaces or verification techniques or other tooling would be a better way to help them not do that than replacing them with tools that by design are only ever probably about right.

So please stop justifying your AI application research by saying simply that it allows non-experts to carry out expert work! I’m much more likely to be convinced by uses of AI that make experts more productive – though don’t get me started on how to measure productivity because I don’t know except via means which are expensive and time consuming, and it really seems that very few people can be bothered doing that.