JP van Oosten

Embedded AI, non-agents and the AI/Poem clock

Jan 28, 2024

Matt Webb wrote something interesting the other day:

I wonder whether the reality of a world populated with AI is not so much about listening, watching, speaking, laser-projecting entities, assistants in our pockets and hanging on necklaces and our every word - not JARVIS or HAL 9000 or Samantha or Joshua - but instead a trillion extremely mundane, genius-level, nameless embedded intelligences, squirrelling away, hidden inside everything?

And how will that work, practically? How will that technology be developed, managed, maintained, secured, networked, owned, shared and made equitable?

And how will it feel to live there?

This is about how we might interact with AI appliances. Matt is working on the Poem/1, a clock that outputs a new AI-generated poem every minute. It's delightful, non-threatening and together with his thoughts above, give a peek into a possible future with lots of small, embedded AI applications.

AI agents (the ones you interact with via, e.g., a chat interface or voice) seem much more heavy, need to have a wide range of functionality to be useful and have to be protected against malicious input and output.

What other small appliances can you think of? The other day at work, we had the idea of using quotes from stoicism in a similar way (for example, for 18:04: "At eighteen-four, embrace the now, a stoic's peace in twilight's bow. // Life's fleeting dance, in each hour's glance, serenity's silent vow."), but you can think of things that give you inspiration around the house as well. Something that generates a journal prompt for you. Maybe some inspiration for a meditation, or while you're writing. And what about a small music box that can output AI-generated music for deep concentration? If AI is "small" enough to fit in a clock, it might fit in anything!

The "genius" part of the "mundade, genius-level, nameless embedded intelligences" is also interesting. Because it's purpose is very narrow, it can actually be very good at what it does. It might not even have to be a 7-billion parameter model, and still be very good at the job.

I like this kind of imagination that's a bit off the beaten path. Of course, a lot of people are using AI for productivity. Others are worried about AI taking over jobs. I understand the concerns, and I think it's awesome to use AI to improve your productivity where it makes sense. But, I also really like the light-heartedness of the Poem/1 approach.

(Also posted on my LinkedIn feed)