2026-10-12: AI reasoning, scientific peer review, scaling laws, software tooling
2026-10-12 — 2026-10-12
Forty-four days, eight new posts, and sixteen updates. Dan’s been in a right state, obsessing over whether these machines are actually “thinking” or just getting better at hiding their homework. He’s spent his time teaching robots to referee their own scientific papers—a bit rich, given the state of most academic writing—and mucking about with his computer setup. He’s finally got the machines to do the boring sums so he can get back to his fiddling.
1 Who is actually doing the thinking
1.1 AI reasoning and proofs
Dan’s trying to teach these machines to check their own homework; getting a computer to “reason” is harder than just stringing words together. He’s pulled apart the difference between a bot that just guesses the next word and one that actually follows a proof. He’s probably just trying to find a shortcut to doing his own sums.
1.2 Building a maths agent
Maths agents: once they’re talking to a compiler, they either get the right answer or they don’t. Dan’s built a harness to keep the thing from wandering off into the weeds. It’s an expensive way to solve a logic puzzle.
1.3 Agent harness design
He’s mapped out how these agents work and where the machine does the heavy lifting. It’s mostly just making sure the bot doesn’t burn your computer to the ground.
1.4 Code agents and assistants
Coding bots are editing their own files now. Dan’s looked at the whole mob of them—cloud, offline, terminal-based—and decided that without a way to search your own code, the agent is just guessing in the dark.
2 The mess we call science
2.1 Should our journal publish AI-drafted manuscripts?
Academics have spent yonks writing unreadable drivel, and now they’ve got bots to help them do it faster. Dan’s run an economic model to see if letting AI into the journals creates more slop, or helps good work rise to the top. It’ll just mean more paper for the recycling bin.
2.2 Epistemic communities
Truth is hard to find, and tech is making it harder to spot the liars. Dan’s looking at how to keep truth-seekers together when any clown can fake a thousand arguments in a minute.
2.3 Peer review meets AI
He’s looking at a proposal where AI referees scientific papers, acting as a neutral judge when the humans argue. It’s a way to settle a scrap, but I wouldn’t trust a computer with the final say.
2.4 Scientific writing
Dan’s whinging about academics hiding behind passive voice and fancy words. He’s collected a heap of guides on how to write like a normal person, though God knows if anyone will listen.
2.5 Science
Dan’s junk drawer for his musings on how research actually gets done. It’s a mess of red tape and professional survival tactics.
3 Measuring the machine
3.1 Algorithmic improvement
Everyone’s obsessed with whether AI is getting better because of hardware or code. Dan’s looked at this as a production function, noting that most “progress” is just us getting better at ignoring what doesn’t matter.
3.2 Production functions
He’s also having a go at the folks using dumb formulas to guess how AI scaling laws work. If you don’t know what an isoquant is, stop writing about AI economics.
3.3 Quis computat?
Survival analysis for AI risk. He’s modelling the chance of doom versus deliverance, but it comes down to how much you’re willing to bet on safety. It’s gambling with calculus.
3.4 Calibrating LLM estimates
Most bots are overconfident snakes, saying they’re 99% sure when they’re guessing. Dan’s looking at how to “calibrate” them so they know when they’re full of it.
3.5 AI evals
Benchmarks are rigged games. Dan’s arguing for “operational evals” that treat models like subjects in a clinical trial.
3.6 Ablation studies, knockout studies, lesion studies
He’s having a crack at the way researchers “ablate” neural nets—breaking parts to see what happens. He’s arguing for proper experimental design instead.
3.7 The language game
He’s trying to reduce meaning to a classifier function. Words don’t have “true” meanings; they represent what we all agreed to label the same way.
3.8 Intergenerational game theory
A grim look at how the present exploits the future. Since future people can’t punch back, the current mob keeps stripping the commons bare.
3.9 Complexity of markets
Markets are distributed calculators, and finding the “perfect price” is a bloody hard computational problem.
3.10 Successors and successionism
Notes on who takes over once the AI hype dies down.
4 Fiddling with the gear
4.1 LLM sampling stunts
Dan’s found tricks for sampling from these models—using “what if” scenarios to pick the next move. It’s just smarter guessing.
4.2 Running LLMs locally on a Mac
Docker’s been a resource pig for yonks, so Dan’s moved his setup over to llama-swap. It’s faster, cleaner, and doesn’t treat his battery like he’s mining bitcoin.
4.3 Zed, the code editor
The Atom crowd has built yet another editor. It’s Rust-based and apparently fast as all hell; I’ll believe that when I see it.
4.4 Code editors
A survey of what everyone is using to write code. Dan’s still stuck on VS Code, but he’s tried everything else.
4.5 Matplotlib
If you’ve ever wanted to throw your computer out the window while making a graph in Python, this is for you. He’s got the usual workarounds for when the font cache inevitably explodes.
5 Minor tweaks
- Added a note on bargaining with successors, Dan’s way of planning for his own obsolescence. About time. (editorial: Ouch, Aunty Val.)
Aunty Val is Dan’s fictitious aunt from Mukinbudin.