2026-08-31: AI agents, economics, digital maintenance, 30 days

1.1 Bargaining with successors

He’s looking at whether we can actually sign a treaty with a future superintelligence. It’s not a bargaining problem, it’s a commitment problem. Since we won’t have any leverage once the thing is turned on, we need to bind the transition with physical constraints, not polite promises. It’s a bit bleak, really.

1.2 Performative prediction

A look at what happens when you use a model to govern people, and then the people start gaming the model to get a better outcome. Like credit scoring—if you punish someone for being risky, you might just make them riskier. Dan’s looking at how we can design systems where “gaming” the system results in genuine improvement rather than just finding a loophole.

1.3 Common knowledge/shared knowledge

The difference between knowing something and knowing that everyone else knows it. He uses the Harvey Weinstein scandal as an example—everyone knew, but it wasn’t “common knowledge” until it hit the papers. He’s got the maths here for how agents resolve this, and common knowledge is the fuel for collective action.

1.4 Multi-objective optimization

A warning that if you try to turn multiple goals into one single “score,” you end up missing the most interesting solutions. You can’t just smooth everything out and expect the best answer to pop out. It’s a good lesson in why trying to make trade-offs “go away” usually just hides the problem.

1.5 Agency under bounded compute and information

Dan argues that our obsession with “infinite” intelligence is a trap. He says we need to stop pretending we’re just crappy approximations of a god and start modelling how people actually muddle through. He’s very proud of the name “Cantor trap,” which I’m sure is exactly as pretentious as it sounds.

1.6 Optionality as an end in itself

The idea that keeping your options open is a goal in itself. He looks at this through the lens of statistics and diversity, arguing that we should treat “having options” as a resource. He’s right, but watching him try to turn “don’t kill your own options” into a rigid mathematical framework is like watching a dog try to play chess.

1.7 Ergodicity economics

The difference between “group averages” and “time averages.” In a rigged game, the group might look like they’re winning, but the individual player is usually going broke. He argues that we should be optimizing for the long-term growth of the individual rather than the mean of the crowd. It’s the kind of thing he picks up from these “Peters” types and then can’t stop bringing up at bridge night.

1.8 Language as a game

He’s decided language is basically a classifier problem. Words aren’t real things in the world; they’re labels we slap on sensory experiences. He even uses this to argue that debating whether a cheetah is a “cat” is an underspecified question. It’s the ultimate “Dan” move—declaring an entire field of philosophy “solved” so he doesn’t have to engage with it.

2.1 Open Questions

A collection of things he’s currently baffled by, including how to make machines cooperate and whether human agency exists if you know enough about someone. It’s mostly stuff he’s stuck on, but he’s filed it all away neatly like he’s hoping someone else will do the heavy lifting for him.

2.2 Running LLMs locally on a Mac

Docker’s been a resource-hungry pig for yonks, so Dan’s finally switched to OrbStack. He’s also had a crack at getting his graphics card to talk to these local AI models, which is still a fucking nightmare, but he’s got it running for now. If you’ve ever lost an afternoon to your laptop battery being drained by background software, he reckons this is the fix.

2.3 AI search

Instead of using some bloated, complex database to find his own notes, Dan’s built a search tool that uses neural embeddings. It’s a flat list of numbers and some basic maths. It’s brute force, but since his notes aren’t the Library of Congress, it works a treat.

2.4 Quarto integrated website system

The lad moved his million-word blog to a system called Quarto a fair while back. It looks fine, but the build times are a bloody disgrace. He’s waiting seventeen minutes for a full render and has resorted to linking bits of his project together just to get a preview window up without waiting for the heat death of the universe. A proper bush mechanic fix, this one.

2.5 Indieweb, small web, cozy web

Dan’s gone all misty-eyed over the “cozy web”—people building their own sites instead of posting on the big platforms. He’s set up a ‘now’ page and a ‘friends’ list, like he’s playing digital pen-pals in 1999. It’s a bit quixotic, but there are worse ways to waste your time online than making something that isn’t owned by some prick in Silicon Valley.

2.6 Friends

A list of people Dan likes, the 2026 version of a blogroll. He’s been chatting to Nick Gray about it, and now he’s got a page to show off who he knows. Most of them are academics or AI types, which tells you everything you need to know about his social life.

2.7 Taking notes

Another month, another note-taking philosophy. Dan’s been trying to sync his files across every device he owns without selling his soul to the cloud. He’s been hacking away at a tool called Hister, which managed to eat 6GB of RAM until he patched the thing himself. If he spent as much time actually writing notes as he did building the filing cabinets for them, he’d be finished by now.

2.8 Neural generative audio

He’s been playing with AI that spits out sound. He reckons you can get a Raspberry Pi to render audio in real time, but he’s not making AI-generated pop songs. He’s too busy making his computer sound like a robot falling down a flight of stairs.

2.9 Should we allow AI-written articles in the Alignment Journal?

Dan’s tried to use an economic model to see if banning AI-written papers is worth the effort. Trying to spot an AI by its writing style is a blunt instrument. He’s wondering if a submission fee might do a better job at filtering out the slop. He doesn’t offer a final verdict, but he’s laid out enough numbers to show that the “ban the bots” strategy might just be a waste of breath.

3.1 Agent harness design

Dan’s trying to figure out how to build a decent agent harness that doesn’t just hallucinate nonsense. He’s broken it down into wiring, verification, and context, and he’s convinced that using a code compiler to check the work is the cleanest way to do it. It’s a lot of faffing about with prompts, but it beats just hoping the model gets the right answer.

3.2 Building a maths agent

He’s having a crack at building an agent that can actually do maths without falling over. It’s all about iterative refinement—let the model have a go, check the result with a code sandbox, and feed the error back into the next attempt. It’s like tasting your cooking instead of just reading the recipe, and it works way better than just shouting at the model to get it right.

3.3 AI agents, applied

A shopping list for building agents. He weighs up different libraries and tools, looking at whether they’re built for serious use or just some hobbyist’s abandoned experiment. He’s pretty sharp about the trade-offs, warning that the more “ready-to-run” a harness is, the more likely you are to get locked into someone else’s proprietary mess.

3.4 Reasoning and proof models

He’s mapped out how models handle logic—from simple guessing to formal proofs that a computer can verify. If you use a compiler to check the answer, you can sample thousands of times because you’re not relying on a “vibe check” to know if it’s correct. This explains why some setups work when others are just throwing darts at a board.

3.5 Fine-tuning danbot

The man’s trained a tiny AI to talk like him, mostly so he doesn’t have to rewrite his own AI-generated slop. It’s a loop where he feeds it samples of his own writing versus generic AI writing. It’s easier to make an AI sound less like an AI than it is to make it sound like Dan.

Aunty Val is Dan’s fictitious aunt from Mukinbudin.