People design very quickly on a monitor, and they print on some enormous industrial printer in a warehouse in a distant country, and the designer hasn’t touched a piece of fabric at any point in the process or gotten her hands dirty with ink. Computers are great for experimentation, but they’re bad for deep thinking.

— Gabrielle Zevin, Tomorrow, and Tomorrow, and Tomorrow

Like so many people, I’m trying to figure out what I think of that nebulous term, AI. Again. Good? Bad? Useful? A waste of resources? The saviour of mankind?

I should preface this note by saying that I have spent most of my career feeling fairly cynical about AI as a technology. It’s such a coarse-grained and loaded term. For years it signalled a whole host of slightly-overlapping-but-often-only-vaguely-related technologies. They’ve often been coupled with what were obviously inflated expectations. The circa-2015 IBM Watson era was the worst I can remember: everybody going mad for chatbots that were mostly quite rubbish. I want to save the AGI chats for another day and just talk about LLMs.

I’ve been enjoying some of the LLM-derived art, like this head-melting video from Poderos and Aphex Twin’s Blackbox Life Recorder 21f by Weirdcore.

But I get the sense that this aesthetic will look dated soon.

I’ve also pootled around with ChatGPT but ultimately found it quite unsatisfying: the text output tastes like cardboard to me. Not nearly as bad as Watson, but still. Plus, the fuzzy lack of accuracy and related trust concerns feel like a significant issue when everything is totally freeform. I can see that being less of a problem if responses can be somehow bounded and verified.

When I read Pivot to AI: Pay no attention to the man behind the curtain, this part stood out for me:

In practice, everything in machine learning is incredibly hand-tweaked. Before AI can find patterns in data, all that data has to be tagged. And output that might embarrass the company needs to be filtered.

Hmmm.

I was broadly in agreement with the tone of that article until yesterday evening, when I saw a compelling demo at a React Advanced meetup: Using React Server Components and OpenAI Embeddings to Create a Killer Search Feature by Jackson Gabbard.

Jackson was able to add a search feature that was better than many existing solutions I’d seen and do it pretty much in real time. What blew my mind was the ability to search for terms in a different language to the site content i.e. cross-language search with no prior translation step.

Embeddings are not the same as LLMs, but after seeing this demo I’m fairly convinced that judicious use of this and related technologies could bring huge improvements to user experience.

I would be much happier if people weren’t badging all of this as AI, though.