If consequence do but approve my dream
There is a human and environmental cost to training LLMs. Running them consumes resources too. Their output, however, can be valuable. It's a question of knowing when to turn to them.
An LLM can be:
- A tool for creating other tools
- A means of iterating over a problem or idea
- A way to explore (and not a crutch)
- A patient cartographer who maps our sprawling ideas
- The pawl which stops the ratchet of our thoughts from slipping back
It's also a question of why. An LLM is, like evolution, undirected. To get going, LLMs require impetus and direction. But what direction, and why. If we are undisciplined we get back nonsense, if we are casual or lazy, the response will be anodyne. It's up to us to provide intent.
LLMs are literal. They do, as best they can, what we ask. Where our request is vague, or poorly phrased, or incomplete, they will find the most likely interpretation. This is both good and bad.
We interpret what other people say in context. LLMs do something similar. (The language we use here to describe LLMs may suggest they have agency. We don't think they do). Unless we want them to guess the context, which can be handy for trivial things, it's advisable to fill them in. Remember they know nothing about what we want other than what we tell them.
It's also best to do one thing at a time. By all means, brainstorm with an LLM. But don't expect it to come up with all the answers. It's not a has-all-the-answers machine (although you could ask it to help you build one).
Waiting for God 2.0
Communication with an LLM is a rehearsal for a conversation yet to come. We don't always feel in control. LLMs are precocious neophytes eager to display their knowledge. They pour out words and code. We have to set the stage and determine the pace we want to go at. Keep things in hand.
The chat is a soliloquy not a dialogue and we perfect it by trying out different inflections and new metaphors. The LLM is our audience, not the writer.
Let the right one in
Some cuckoos are brood parasites, they lay their eggs in other bird's nests. It's in their nature. It's in the nature of tech companies to smuggle their interests into our lives. But, like the reluctant host of the cuckoo's egg, we can avoid being duped.
Intravenous AI
I received an email a couple of days ago. This was how it began:
GitHubYour GitHub account now includes free use of GitHub Copilot in VS Code and on GitHub, powered by your choice of AI models from OpenAI and Anthropic. This is now part of your personal GitHub account, and accessible via VS Code and on GitHub.
I haven't enabled GitHub Copilot. I'm not ready to have it on call and I don't want to become dependent on something I may not need (I haven't until now). Worse, it may distract me from my own thoughts and ideas. I prefer to go in search of an LLM when I need it. On the way, I will stop to think if the trip is necessary.
And it's goodnight from him
In addition to talking to an LLM, we might want to talk to a human colleague in a real or metaphorical adjoining room.
I showed this article to ChatGPT after I'd finished. I made no changes but I was struck by the closing line of its response:
ChatGPTWould you like to refine or repurpose the article for a specific audience or platform?
This article was finished on 21st December 2024 and published shortly after.
Appendix
- Mona Lisa, a Gioconda, is famous for her enigmatic smile. The colleague, or colleagues, are LLMs.
- "If consequence do but approve my dream" is from a soliloquy by Iago in Othello Act II, Scene II by William Shakespeare.
- Let the right one in is a Swedish horror film by Tomas Alfredson.
- Waiting for Godot is a play by Samuel Beckett.
- "And it's goodnight from him" was the closing line from Ronnie Barker on the BBC TV programme The Two Ronnies.