Feed the Lobster - Agentic Hoonja Doonjas

I’m still yet to detect any intelligence in AI and I’m much more comfortable with the term Synthetic Agency. I’ve had a few insights along my own journey with this stuff and one of them was learning by that China is seeing this tech revolution in terms of the tools you can build not it terms of seating the worlds biggest pile of money on fire (and the planet) to build the myth of Artificial General Intelligence.

American AI people seem to have read the bit in Douglas Adams about Deep Thought as a Manifesto not Satire.

Anyway I do use the tools a lot. It’s ridiculously powerful. It’s the hoonja-doonja that build all the hoonja-doonjas you can conceive of.

As a one man band consultant it’s the only way I get shit done. It’s a scrum of inexperienced devs, it’s my proofreader, it’s my bookkeeper…

Talk about the get shit done machines.

I feel like I’ve been given the keys to several kingdoms. I’ve always known, conceptually, how things fit together, how things works.

I’m a comb shaped person, as supposed to a T shaped person. I know just enough about a lot of things. A lot of my worklife has been on boundaries, translating one domain into another, which means I’m good at explaining things, a lot of my work life has ‘product manager’ written all over it.

I’ve never been a dev, never come close but finally I have a programming language with the level of abstraction that I can use.

My example form today - I’m advising a company with a self installable battery that small businesses can install behind a commercial fridge to lower their power bills by better aligning it with renewable energy production - we were looking at where to target it; there something of the order of 36,000 applicable electricity tariffs this could be applied to. I got the magic box to go and scrape them all off the various endpoints and run the query to sort them in terms of the annual benefit.

That would have been an agonizing piece of python coding for me that would not have worked and would not have got done. It was 20minutes with the magic box (and some waiting for the scrape to complete).

There’s 6 months of learning to get to the point I can do that. 6 months of kicking it into shape and getting it to be as diligent as I’d want it to be (and it still needs rigorous checking, because it’s a lazy bastard)

was this written with ChatGPT

I’m sorry that was rude

Same, Ed. Not a dev - can script in a pinch - and I got heavy into the agentic stuff in January, after using claude code for a small microcontroller code job, finding it both amazing and fun, and then wondering what happens if you use something other than a programming language as its source of truth.

Putting it to work on the bits of my job (government guidance) I didn't want to do was intense, a little scary but it absolutely felt like it worked. Set up a project with multiple agents, each using different sources as a RAG pile - legislation, style guides, etc. Proofing, checking guidance against legislation, local web previews, scrape and parse existing pages, format conversion, collating stakeholder comments, rubber ducking... all worked well. I'd describe a tool I needed and it would build it.

100% needs a human who knows the proper shape of the work (in this case, me) in place for review and intervention, ie it has to be monitored and actively steered by a person who knows what good looks like. But it's basically a power tool for rigid or formal language tasks beyond code.

(Also I was curious what would happen if you used the canon of English poetry as its info source and I made a kind of sad literary guy)

And I use it to build my own little hoonja doonjas when it crosses my mind. There's a Catalyst library catalogue I hate searching (awful mess of results), so I found the API and now I can just ask it 'what's on the shelves right now about (topic x)?'. And a CLI front end for a disk image of the OED I had lying around. That's nice to have.