AI is that tool you now use in your day to day job to help you complete tasks

What do you think of when you hear AI, Artificial Intelligence? I grew up watching science fiction movies, so I immediately think of a powerful computer entity that is capable of unique thoughts and ideas. But that is not what AI currently is. AI is a set of tools built around LLMs (Large Language Models) that use probability to determine likely outcomes from a user’s input.

I often feel this internalisation of AI as it is depicted in the movies is something that affects the way people think about the tool. It means they overestimate what it is currently capable of doing.

The Analytical Engine has no pretensions whatever to originate any thing. It can do whatever we know how to order it to perform.

— Ada Lovelace, Note G, Sketch of the Analytical Engine (1843). Wikiquote · Bodleian

AI does not think

So what does that actually mean? Well, to most people not a lot.

But to me the point here is that AI, as it currently stands, does not think. It is a fancy calculator (hopefully that won’t offend everyone) able to do whatever task you request of it, if you give it enough instructions to understand the task first. But it does not think up new ideas, it does not have opinions. In fact, any opinions it does give are just the combined output of the web, something known for its lack of credibility.

Alan M. Turing thought about criteria to settle the question of whether Machines Can Think, a question of which we now know that it is about as relevant as the question of whether Submarines Can Swim.

— Edsger W. Dijkstra, The threats to computing science (EWD898, 1984). EWD898

AI doesn’t understand hints, or know what you meant

Think of Claude as a brilliant but new employee who lacks context on your norms and workflows.

Golden rule: Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they’d be confused, Claude will be too.

— Anthropic, Prompting best practices (Claude docs). platform.claude.com

This is it, the key point. You can only get out what you put in. Tell AI what you want, be specific whenever possible, go into the sort of detail you would when explaining the problem to a toddler if you need to. This increases the likelihood you will get a suitable response. Lovelace had this right in 1843: it can do whatever we tell it to, but nothing we don’t. Well, that’s not entirely true, it’s sure to do what we don’t want if we aren’t clear in the request. You don’t blame the meat when it arrives at your table burnt, that’s the chef failing at their job. So why blame AI for producing substandard output? Look instead at the user, and how they are using the tool.

When will AI be like the movies?

I already talked about this in Kyle (Reese) was right so I won’t repeat myself here. Instead I’ll leave you with the following quote:

It was like claiming that the first monkey that climbed a tree was making progress towards landing on the moon.

— Stuart Dreyfus, quoted by Melanie Mitchell under “Fallacy 1: Narrow intelligence is on a continuum with general intelligence”, Why AI is Harder Than We Think (2021). arXiv:2104.12871