2026-06-22: Run it yourself, machines earning their keep, who’s pushing who, what’s baked in

1.1 Community sovereign AI compute

Dan’s reorganised this one to front-load the obvious: the cheap AI access everyone’s enjoying right now is the free samples phase. Once you’re hooked, prices go up and the fine print gets tighter — same as every platform before it. And it’s not just commercial greed to worry about; Washington, Brussels, and Beijing between them control what AI Australians can actually get their hands on, and none of those three give a rat’s arse about us. The rest of the posts in this cluster — running models on your own Mac, shared servers for small groups, that sort of thing — are Dan’s answer to that problem. Whether he finishes setting any of them up is another question entirely.

1.2 Running LLMs locally on a Mac

Running a big language model on your own Mac used to be the sort of thing only Dan’s type of lunatic would attempt. Now it’s actually practical, and he’s mapped out the whole business — the inference engines that do the number-crunching, the server bits that sit in the middle, and the desktop apps you’d actually talk to. On formats: MLX if you can get it, GGUF otherwise, and watch your quantisation or you’ll run out of RAM halfway through a sentence. There’s also a bit on macOS being precious about memory and indexing multi-gigabyte model files — yes, Time Machine will absolutely try to back up your seventeen-gigabyte llama if you let it, and yes, that’d be stupid.

1.3 Sovereign compute for small collectives: technical implementation guide

Dan’s been tinkering with his guide on running a shared AI server — the sort of thing a community group or small organisation might set up so they’re not handing all their data to some big tech outfit that’ll charge them through the nose the moment they’re hooked. The main news is he’s added Xiaomi’s MiMo-V2.5 as a model worth looking at: lighter than the previous front-runner, fits in the machine’s memory without a fuss, handles images and audio, and comes with a licence that won’t give your lawyer a headache. He’s also tidied up the model comparisons for DeepSeek, Kimi K2, and Llama 4 into one place instead of having them scattered everywhere — which, knowing Dan, is progress. He’s also flagged that if you want to run one of these models with the guardrails knocked off, don’t assume someone’s already done that for the newest release. The tools for that tend to lag behind the models themselves, sometimes by months. Find that out after you’ve built your whole setup around it and you’ll have a bastard of a time unpicking it.

1.4 Discretizing and quantizing neural nets

The big models are too fat to run on your laptop, so the trick is squishing the numbers. Instead of storing weights as big precise decimals, you swap them for rougher, smaller ones. You lose a bit of detail, but the thing actually fits in memory and runs at a sensible speed. The new bit Dan’s added is about calibration data — before you do the squishing, you run the model over a carefully chosen sample of text to work out which weights are doing the heavy lifting, so you don’t throw away the important stuff. Turns out the text you pick for that trial run actually matters, which is the kind of thing you’d only find out the hard way. Tools like llama.cpp’s imatrix do the whole thing for you if you don’t fancy working it out yourself. Read it alongside the Mac setup post above — knowing what quantisation actually does changes which format you pick.

1.5 Fine tuning foundation models

Dan’s gone down the rabbit hole of “abliteration” this time — nothing to do with demolition, despite what it sounds like. Turns out if you go digging around inside a language model’s internals, you can find the single direction responsible for all those “I can’t help with that” refusals, and just… erase it. No retraining, no fuss. People are already using this on DeepSeek and Qwen to scrub out the politically-motivated bits the Chinese developers baked in — there are whole tools, deccp and huihui-ai, doing exactly that at scale. If you want to do your own fine-tuning without the surgery, Dan’s also listed the practical options: paid services like Together and Fireworks if you’d rather not deal with it yourself, and Unsloth if you’re the DIY type.

1.6 Generative AI workflows and hacks 2026

Dan’s tidied up his running notebook on using AI day-to-day, and there’s actually some useful stuff buried in here. Most of the big providers — Anthropic, OpenAI, Google and the rest — will train on your data unless you go digging through the settings and turn it off. Dan’s mapped out who does what and how to opt out, which is the sort of thing you’d want to know before you paste anything sensitive into a chat window. He’s also been fiddling with fine-tuning a small model on his own writing so the drafts stop coming out sounding like someone trying to get promoted — fair enough goal, I reckon. And there’s new notes on running the whole lot locally on a Mac, for anyone who’d rather not send their half-baked ideas to a server farm in California. Knowing Dan, he’ll spend six months on the fine-tuning and declare victory before he’s actually finished it.

1.7 Generative art with language+diffusion models

Dan’s turned what used to be a quick squiz at image models into a proper guide for actually finding and using the bloody things. Turns out where you download a model from matters a lot — Hugging Face plays by payment processor rules, CivitAI has its own limits, and the Chinese platforms like LiblibAI and ModelScope are shaped by whatever Beijing wants that week. Different rules, different models, different communities. He’s got proper write-ups on FLUX.1, SDXL, and a few others — what they’re good at, how fast they run, what the licence lets you do with them. Knowing Dan, he spent twice as long on that than finishing anything else on his list.

1.8 Front-end clients for AI image models

Dan’s been tidying up his comparison of image-generation apps for Apple Silicon Macs — Draw Things, ComfyUI, InvokeAI, Mochi Diffusion, that mob. The big tables are gone, which is a mercy, and the install instructions are clearer. Draw Things can now train its own LoRA models on-device, meaning you can teach it what your face looks like without sending anything to a server. Knowing Dan, he’ll spend six months fiddling with that instead of finishing any of his other half-done pages.

1.9 PDF ingestion

Not all PDF converters are equal, and the difference matters if you’re feeding maths to an AI. The fancy ones that look at a whole page like a picture — Dan singles out Docling’s Granite pipeline — will mangle your equations into something that looks right but isn’t. The specialist tools, marker and MinerU, actually kept the symbols intact. Dan’s mapped out which one to use when: markitdown for plain prose, marker if you’ve got a one-off with maths, MinerU if you’re processing a stack of them. He’s also bundled the routing logic into a reusable bit called pdf-ingest-skill so an AI agent can just pick the right tool without you thinking about it. Which is handy, because let’s be honest, nobody wants to think about PDF converters.

1.10 Backups

Dan’s added a section explaining why those cheap little portable spinning hard drives keep dying on you — turns out a USB port can barely scrape together enough power to get the platters moving, so the drive just disconnects mid-write and corrupts your files. Western Digital’s 5TB mini HDDs cop a specific mention. The fix is a portable SSD, which runs fine off USB — except a 4TB one will set you back about a grand right now, so that’s a lovely solution that most people can’t actually afford. He’s also sorted the cloud backup options into their own section if you’d rather someone else’s hard drive be the problem.

1.11 Vibecoding Apple apps

Dan’s been mucking about with Claude Code and Xcode, trying to get an AI to write his iOS apps for him. Honestly, fair enough — if you can avoid learning Xcode’s signing and provisioning rigmarole yourself, you probably should, because by all accounts it’s a complete bastard to navigate. The new bits cover the parts that actually bite you in practice: getting your half-finished app onto a mate’s phone, whether SwiftUI Preview plays nicely in an AI-driven loop (sometimes), and whether the AI falls apart on trickier system stuff like HealthKit (also sometimes). He’s also tacked on a section asking whether you’d be better off skipping Swift entirely and just scripting macOS the old-fashioned way — AppleScript, Python, that sort of thing. Knowing Dan, the answer he landed on is “write a Swift app anyway,” but at least he asked the question.

2.1 Neural nets that do symbolic mathematics, logic and other reasoning tasks

Dan’s sorted these maths-doing models into three piles. General language models that can do a bit of arithmetic. Solver models like Qwen2.5-Math and AceMath that work through problems in plain language. And then prover models — DeepSeek-Prover, Goedel-Prover — that spit out code in something called Lean 4, which a compiler then checks like an auditor who doesn’t take your word for anything. Instead of hoping the model’s right, you get an actual yes or no. One trick that works across all three families is just running the same problem through dozens of times and going with whatever answer comes up most — which sounds like cheating on an exam, but apparently it works a treat.

2.2 Maths and proof models, applied

Less about what the models are, more about how you’d actually string them into a workflow. Dan’s worked out that the main trick is running the same problem through it dozens of times at once and going with whatever answer comes up most often — basically outsourcing confidence to brute repetition. You start simple: ask Claude or DeepSeek to check its own working with Python, then scale up with parallel runs through Modal. The fancy end of it uses Lean, that proof checker that either accepts your maths or tells you to get stuffed — no partial credit, no wriggling. Writing the proofs isn’t the hard part; translating a plain-English maths problem into Lean’s formal language is, and current tools only manage that right about half the time. So yes, we’re building AI mathematicians, but first we have to solve the bit where they can read the question.

2.3 Code agents and assistants

Dan’s tidied up his page on coding assistants — which, knowing Dan, means he’s added three new things and buried the ones that didn’t pan out. The newcomers worth knowing about: OpenCode, which is just a terminal wrapper that’ll work with whatever model you point it at, and MiMo Code, Xiaomi’s fork of that, which apparently keeps track of what it was doing across a longer job. The old stuff — CodeWhisperer, Codeium, that mob — got shown the door. Basically the page is now about what he actually uses instead of everything he’s ever heard of, which is an improvement.

2.4 AI tutoring

There’s a thing researchers keep banging on about called the “two-sigma problem” — back in the ‘80s someone worked out that a kid with a private tutor outperforms a classroom kid by two standard deviations, which in plain terms means tutoring is so much better it’s almost embarrassing. Dan’s question is whether AI can do that job now. He’s rounded up a few researchers watching this space and points to Google’s Learn Your Way as one real-world attempt. The hands-on bit is a prompt he uses with Claude that stops it from just handing you the answer — forces it to ask questions instead, Socratic-style, like a tutor who actually wants you to think rather than just copy. Works, apparently, but needs watching. Knowing Dan, he’ll spend more time fiddling with the prompt than actually learning anything.

3.1 Performative prediction

Dan’s been chewing on what happens when the model’s own prediction changes the thing it’s predicting. His main example is a credit-risk classifier: call someone high-risk, slug them with punishing interest rates, and surprise surprise — they default. The model wasn’t wrong, exactly; it just made itself right. A model can settle into a stable situation — retrain it on the world it helped create and nothing shifts — and that stable point can be genuinely shit, with everyone gaming the system and nobody actually getting ahead. Stable isn’t the same as good, which honestly sounds less like a machine learning problem and more like local council.

3.2 Algorithmic collective action

Performative prediction is mostly about one model shaping the world it was built to read. This post flips it: what happens when the people on the other side start shaping back? If you and a bunch of others are all feeding data into the same platform — reviews, clicks, whatever — coordinating even a small group can nudge the model toward outcomes you actually want. The platform’s not helpless, mind — it can push back by steering what you see, shaping your behaviour before you even get to shape theirs. That’s the tug of war. The second half goes somewhere different: AI agents forming little teams among themselves, using a game-theory trick to split rewards so no subgroup has reason to go rogue. Dan reckons you can formally prove those coalitions hold together, provided the agents have a decent model of what the world’s going to do next. Whether any of this stays stable once the platform works out what you’re up to is, as usual, left as an exercise for the reader.

3.3 Wisdom from the madness of crowds

The internet is full of people trying to mislead you, and a model trained on that mess just learns the mess. Dan’s been thinking about whether there’s a smarter way to squeeze out what’s actually true when your sources are biased, strategic, or outright lying. Turns out there’s a whole toolkit for this — prediction markets, scoring rules that penalise bullshit, a thing called the Surprisingly Popular algorithm where you find the answer more people would have guessed they were in the minority for believing. None of it’s magic, but it’s a damn sight more principled than hoping the lies average out. He’s also lined up some experiments to test whether a model falls apart faster against clever structured misinformation than against random noise — which, if you think about it, is the more important question. Knowing Dan he’ll get halfway through the first experiment and discover some other rabbit hole entirely.

3.4 Cooperation through uncertainty

Dan’s been picking apart why animals bother helping their relatives at all, given they can’t exactly ask for a DNA test. The short answer is they’re reading fuzzy signals — smell, looks, proximity — and making a best guess. Dan works out there’s a tipping point where those guesses flip from “help” to “don’t bother,” and getting it wrong sometimes isn’t a failure — it’s the maths working exactly as it should. You’ll always end up helping the occasional stranger by mistake, and occasionally blanking a real cousin. That’s not a bug; that’s the best any animal can do with noisy information. He’s also sniffing around the idea that evolution might actually favour keeping those signals a bit ambiguous on purpose — which, if you’ve met enough people, sounds about right.

3.5 Multi-Agent Reinforcement Learning

This one’s about what happens when you’ve got more than one AI agent in the mix — and they’re not just learning, they’re learning around each other. Could be cooperation, could be competition, usually it’s a bit of both. Dan’s got placeholder notes on how agents figure out what the others are actually trying to do just by watching them, how they form coalitions, and something called “open-source game theory” where agents share their whole strategy upfront so nobody has to guess. Trusting each other by being transparent rather than clever about it — which, frankly, is more than you get from most people at a stock agent’s office. Knowing Dan, half these placeholders’ll still be placeholders next Christmas.

4.1 Power-seeking

There’s a worry in AI circles that a smart enough system will naturally start hoarding — resources, influence, options — not because anyone told it to, but because having more options helps you achieve pretty much anything. Dan’s been working out whether you can actually measure that with a number, using something called empowerment — basically, how much does what you do now box in what happens next. The connection between “this AI is grabbing for control” and “here’s a clean formula for that” falls apart pretty quickly once you’re in any environment messier than a toy problem. Defining “future influence” turns out to be one of those things that sounds obvious until you try to write it down.

4.2 Informational empowerment

The formal version of that same idea is what Dan’s working through here. “Empowerment” — and no, it’s not the motivational poster kind — is the notion that instead of chasing some score or reward, an agent just tries to keep as many doors open as possible. It wants to end up somewhere it can still do a lot of things, rather than boxed into a corner. Dan’s got proper maths for it — mutual information between actions and future states, if you want to go looking — but the plain version is: don’t get stuck. He reckons evolution might be doing the same trick, favouring flexible strategies over narrow ones just because flexibility survives more surprises. Whether that’s a deep insight or a very long bow, I’ll leave to Barb.

4.3 Intrinsic motivation

The problem Dan’s chewing on here is how you give an AI something like curiosity — a nose for what it hasn’t seen yet — without wiring up a big external reward system that tells it what to care about. He’s added a few new ideas to the pile, including one called ICCEA power, which is basically a measure of how much freedom a human still has to pursue their own goals given what the AI’s been up to. Think of it as asking “has this thing been cornering the market?” He’s also dropped the promise of a comprehensive overview, which — knowing Dan — was never on the cards anyway.

4.4 Homunculi all the way down

Dan’s added a chunk on Douglas Hofstadter’s “strange loop” idea — the bit where a self isn’t something that has a self-model, it just is the self-model, folded back on itself. Like a camera pointed at its own screen: you’d expect it to spiral into infinite regress, but it doesn’t, because the camera’s got finite resolution. The fuzziness is what makes it settle. Dan’s point is that a lossy, cut-down self-model doesn’t just save on compute — it’s actually what stops the whole thing collapsing. He’s tied it into some formal maths around self-referential proofs and game theory, asking whether a causal graph can point back at itself without going haywire, and what you have to give up to make that work. Knowing Dan he’ll spend the next six months staring into that particular void.

4.5 Awesome illustrators and printmakers

Dan’s gone and built himself a big list of favourite old printmakers and illustrators — Dürer, Gillray, Doré, the lot — and honestly, fair enough. Someone had to. He’s grouped them sensibly: the old master engravers, the satirists, the visionary weirdos. Short notes on each, links through to the Rijksmuseum and British Museum where you can actually look at the work. The whole thing started because he went down a rabbit hole of public domain images, which, knowing Dan, probably cost him a fortnight he didn’t have. Still. Worth a look if you’ve ever wanted to know who drew all those mad old etchings.

Aunty Val is Dan’s fictitious aunt from Mukinbudin.