2026-01-18: Imprecise Bayesianism, Generative AI workflows, Causal inference, Reinforcement learning, Privacy
2026-01-18 — 2026-01-18
Stone the crows, the lad’s been busy this past week — five new posts and ten updates. Main threads are handling uncertainty with Imprecise Bayesianism (that’s not forcing one exact probability when the data’s fuzzy) and using causal inference to make better decisions — ‘learning to act’ means figuring out what actually causes outcomes so an agent doesn’t just guess. He’s gone practical too with Generative AI workflows and hacks 2026 — tips for getting these models to behave — and tossed in travel notes from London, San Francisco and Melbourne to prove he left the house. Updates sweep through reinforcement learning, visual tools for probabilistic graphs, AI evals (that is, tests to see what models actually do) and a bit on cutting down government spying.
1 Newly published
1.1 Imprecise Bayesianism
Righto, Dan’s been poking at “imprecise Bayesianism” — basically you don’t pin yourself to one probability but keep a whole range of plausible guesses. He runs through infrabayesianism (those ‘infra’ sets of beliefs), the old maximin trick of updating every prior and choosing the worst-case-safe action, and PAC‑Bayes bits that give you high‑confidence guarantees for ensembles even when the model’s wrong. Why care? Because our models are always a bit dodgy in the real world — this stuff helps you be honest about that, pick safer actions, and get useful guarantees for stacking or neural nets. He admits some of it’s more theory than toolbox, but he leaves pointers if you want to muck about.
1.2 Generative AI workflows and hacks 2026
Strewth — the chatbots have started mangling copy‑paste, especially maths and links. Dan’s put together a practical set of hacks: a tiny macOS clipboard HTML→Markdown script, a better client recommendation (Jan), browser-extension quick fixes, and a Deep Research client he wrote that keeps links and maths intact. He shows how OpenAI’s web client now spits messy HTML and how Google Deep Research loses inline links, and what actually works in the real world. If you paste AI output into notes, docs or research, these tricks save you hours of cleaning up broken maths and dead links, and the Deep Research client is on GitHub if you want to nick it or help fix it.
1.3 London
Fancy that — Dan’s sized up London by watching what people shove in their trolleys. He reckons supermarkets are like little social name‑tags: Waitrose and M&S whisper ‘comfortable’, Sainsbury’s is sensible middle, Tesco is everywhere, and Aldi/Lidl wear the ‘cheap’ label even when posh folks nip in. He also bats through department stores — Harrods, Selfridges, Argos — and nods to Kew Gardens as the empire’s plant powerhouse. It’s handy because it shows how ordinary shopping choices give away taste, postcode and money, and why those small signals matter.
1.4 San Francisco Bay Area
Crikey — Dan’s had a crack at the Bay Area. He paints it as a magnet for visionaries and weirdos, points out a few properly fun institutions and oddball people, and doesn’t muck about with the ugly bit: sewers, storm‑water and ageing PG&E kit that’s falling apart. The new bits highlight quirky places and lesser‑known folks, then give you a blunt look at why the city’s shiny tech scene sits on shaky pipes and dodgy transformers. Read it if you want a feel for the place beyond the hype — useful if you’re curious, thinking of going there, or wondering what could go wrong.
1.5 Causal inference in learning to act
Blimey — Dan’s gone and tied cause‑and‑effect to teaching machines to act. He reckons learning to act isn’t just pattern spotting but asking ‘what if I did this?’, so he walks through the old bandit results and some newer papers to show the point. The neat bit is about offline learning: using other folks’ logs can hide the real causes (their choices tangle things up), so causal thinking helps you avoid being fooled. There’s also a dash on predictive coding for the interested, and the upshot is practical — if you want agents that won’t cock things up, you need to think causal.
2 Updated
2.1 Reinforcement learning
Righto — Dan’s stuffed in three proper additions: a ‘Variants’ roundup that gathers oddballs like reward‑free RL (empowerment), generative action spaces, hierarchical schemes and diffusion tricks; a history‑based RL pointer; and a new bit that ties reinforcement learning to causality and counterfactuals. He’s also dropped runnable PyTorch snippets — a REINFORCE example and an exploration conversion — so the maths lines up with working code. The tutorial now links the policy‑gradient Monte‑Carlo maths to a value‑function/TD update so you can see how the pieces fit. If you want to follow the path from Michie’s matchbox to today’s RL tricks and why causality turns up, that’s the new stuff.
2.2 Utility and evolutionary fitness
Righto — Dan’s stuck in a proper prelude that pitches natural selection as a kind of inference, and he’s added a section tying replicator equations to statistical inference. He also flags mesa‑optimizers as an alternate story, but mostly doubles down on the idea that long‑run multiplicative growth makes utility equal to log‑fitness — the Kelly‑style bet‑hedging result. There’s a new blunt bit saying fitness landscapes aren’t real and a few extra pointers into game theory and inference. If you like the maths, he makes the Malthusian log‑fitness ↔︎ selection‑gradient link clearer so the ‘as‑if’ claim isn’t just handwaving.
2.3 Bikes
New: Dan’s added a Brompton section. He reckons the Brompton’s the classic folding bike — small enough to nudge under the usual 158 cm baggage sweet spot but still a proper pleasure to ride. Notes a healthy secondhand market but warns they’re roughly 50% dearer in Australia, so don’t buy here if you can dodge it. Practical tips: an IKEA DIMPA bag fits a Brompton, Carradice makes lovely bags, and watch out — some hard cases are slightly oversize so you’ll pay extra at the airport.
2.4 Causal inference in highly parameterized ML
Righto — Dan’s stuck in two proper new bits: ‘Causality with agency’ and ‘Implicit causality in foundation models’, so he’s actually looking at cause‑and‑effect when models act and when big foundation models seem to infer causes behind the scenes. He’s also added notes on pgmpy and a cause2e entry, and keeps DoWhy and TETRAD up front as the practical tooling to try. Importantly, he flags benchmarks like CauseMe for testing dataset‑shift, and ties the whole lot back to using causal graphs with nonparametric neural nets. If you want to know whether big models are learning real causes or just pretending, those new sections are the bits to read.
2.5 Causality, agency, decisions
Righto — Dan’s added proper mechanization bits and a tooling note. The new ‘Basic Mechanization’ bit walks you through turning causal graphs into things that can reason about decisions, including influence diagrams where decision variables have probability stuck on them. He’s fleshed out multi‑agent DAGs so agents can be modelled deciding about each other, and uses the humble thermostat as a neat feedback example. And for the practical types, there’s a pointer to pycid, a Python library for doing the causal influence‑diagram arithmetic so you can try it for real.
2.6 Melbourne / Naarm
Righto — Dan’s added a proper ‘Urban planning’ section (missing‑middle housing gets a mention and, yes, trams crop up), and he’s expanded the ‘…as Naarm’ bit into a readable chat about the name, how to pronounce it, and why getting hung up on labels is daft. He flags that Buruli (flesh‑eating) bacteria are encroaching on Melbourne. The music section is renamed ‘Where’s the electronic music?’ and now lists local sound‑system links and solid audio‑repair contacts like David Aurora and Open Ear.
2.7 Visualising probabilistic graphical models
Belter — Dan stuck a ‘pgmpy’ section into his visualising graphical models notes. pgmpy can both build and draw models (via pygraphviz, networkx or daft), which is handy when you want proper plate notation, inline maths and neat SVG/PDF exports.
2.8 AI evals
Righto — Dan’s added a proper ‘Statistical methods for evaluating AI systems’ section. He calls out the shonky stats folks still use — missing RCTs, low power, multiple-comparison messes — and points to BetterBench and a chatbot‑arena repo. The page still sets out the benchmarks vs evals split and flags EvalEval, Inspect, human judgements and user logs.
2.9 PIBBS x ILLIAD Research residency January 2026
Well I’ll be — Dan’s stuck in a new ‘Reinforcement learning meets SLT’ section on his January 2026 PIBBS x ILLIAD residency notes from London. Chris Elliott ran the idea of tying Singular Learning Theory to reinforcement learning, with the policy picking up a global Gibbs posterior — neat maths, but it leans on some strong assumptions.
2.10 How to reduce government spying on you
Hang about — Dan’s stuck in a new “Bluetooth is cursed” bit telling you to kill Bluetooth unless you really need it, with pointers to real attacks and vendor chatter. This page’s about cutting state surveillance; it still runs through anonymous drops like SecureDrop and hardware stuff like USB condoms and hardened OSes.