My personal AI research agenda, mid 2024 (and a pitch for work)

17.26, Friday 7 Jun 2024

I want to show you some new work around AI agents.

Then I want to summarise my current interests, and pitch you on a project.

First, a pointer to an AI agent paper I put out this week:

Lares smart home assistant: A toy AI agent demonstrating emergent behavior – see videos demos, get the code, and read my observations and speculations.

I define an agent as an AI system that (a) uses tools, and (b) has autonomy on how to reach its goal and when to halt.

You may remember that last year I built a proof of concept for an AI-powered smart home assistant. Award winning! Watch the Lares v1 video here (and the technical write-up).

It demonstrated problem solving, which is wild. Like, if you ask it to look in a room for you, it’ll figure out it needs to turn the lights on first.

Well, as a proof of concept it wasn’t very reliable.

The version linked in the paper is super reliable. You can download and run the code for yourself.

What’s wilder than the emergent problem solving abilities is how simple it is. You need very little code to get really sophisticated emergent behaviour.

It’s a technical write-up.

But the observations and speculations section goes further:

  • It gives me new directions in research. Like, if you had an autonomous AI agent embedded in your stove, how would you have visibility on its actions and the chance to confirm or repair its plans? An interesting UX problem.
  • It suggests possible futures. Like, if agents are this simple to build, we’ll have a million of them. So the “tools” agents use will be shared and have to be discovered somehow… we’ll need special search engines aimed at AI agents. (I wrote about that one.)

A meta point: I like building toy prototypes because it puts me in touch with the tech in a way that talking and thinking never do. And new thoughts come!

So do check that out and share it round. Some light reading for your weekend haha

When it comes to AI, this is where I like to dig

If I were putting down a personal research agenda for AI, here in the middle of 2024, these are the areas I find most fascinating:

  • Human-AI collaboration, particular in the multiplayer domain. Here’s my work with on-canvas NPCs from last year. Human-AI interaction is a specific case of multiplayer generally; you need to solve for both. (I’ve written a lot about multiplayer.)
  • Simple agents that act on the world not knowledge (i.e. not all-powerful assistants). Both how they work, as in that paper above, but also the second order effects such as the AI search engines mentioned above.
  • Text beyond prompting – can we visualise it? Can we automate hermeneutics? Can we put Photoshop tool palettes inside TextEdit? I wrote about that last week.
  • Tiny, ubiquitous, embedded intelligence or, as I called it last year, intelligence too cheap to meter.
  • Embodiment and physical AI – whether that means a clock with AI inside or building an environment with ambient intelligence, like my speculated Map Room 2.0.

Like, if I had the opportunity to really zoom in, building in these places is where I would spend my time. They’re all connected.

I think something they have in common is that they’re all very human (human scale and in the human world) and they’re all very simple (but with emergence). I can’t quite put into words what I think will happen with concerted work in this domain but… something unique.

They all share this intriguing combination of the utterly pragmatic and also brand new techniques only just being researched by the AI companies.

So that’s the pitch I guess?

I’m looking for my next really chewy projects for my product invention micro-studio.

I would love to find projects in the domain above, ambitious enough that I could pull together a small team for a 3+ month sprint, ideally with public outcomes.

My experience is that, by building focused prototypes with tangible outputs, you invent new products, expose really interesting interaction design challenges, and also drive the technical research agenda.

But maybe it’s not a regular client project, maybe there are other ways of working on this kind of stuff, or maybe I should bite the bullet and start something new.

I’m open minded. Get in touch.

More posts tagged:
Follow-up posts:

If you enjoyed this post, please consider sharing it by email or on social media. Here’s the link. Thanks, —Matt.