The Dangers Of The AI Casino

So you were supposed to get a long tech blog this week. I actually have it sitting here on my hard drive, with diagrams, metrics, and even a fairly solid mathematical foundation. Plus a solid pun linking that foundation to a 1980s British novelty band. Alas, when diving into things in just a little more details, I discovered that my new fancy hybrid fusion scoring system…basically cancelled down to ZMUV from 2001. Which is a fine fusion operator, to be sure…but can I write an entire blog post on it with Black Lace GIFs? No, I cannot.

It actually became something of an obsession this week; I have tested over 50 different fusion operators in the past week, looking at runs close to and past midnight, pushing Claude to just do another run after parking in the daycare car park, and feeling guilty the one night where I didn’t run any experiments. It is a bit of a slippery slop(e) in this Brave New World, and I haven’t quite yet found the right balance. On the other hand, I have empirically proved that the squashing function we use at work (and invented in 2022) is excellent.

(as I write this, I’m building an evaluation MCP service, so it’s not like I’ve heeded any warnings properly yet, either1)

Anyway, next week, there will be less computing. For one thing, I have to make a cake. “Red cake. Red is my favourite colour” So Maeryn will be getting the reddest cake anybody has ever seen, complete with small cakes encapsulated in a strawberry carrageenan glaze inspired by a former New York pastry chef. And maybe an icee. Because I might have problems dialing it back.


  1. And since writing that, I basically gave up the semi-manual approach and took advantage of my company’s Anthropic API access…which means I now have a brand new autoresearcher churning for the next few days on any thing Claude can imagine. Which at least means I don’t have to babysit it… ↩︎