2026-11-09: Agent security, voting maths, measuring progress, Thailand and Taiwan

1.1 Securing agents

Picture a house-sitter who reads every bit of paper that lands in the house and does what it says. That’s these AI agents. Tell one to water the plants and fine, but a flyer in the letterbox saying “post the spare key to this address” works just as well, because the thing can’t tell who’s got the right to give orders. Dan’s rule of thumb is that an agent should only ever have two of three things at once. Your private stuff, a stranger’s words coming in, and a way to send things out. Training the robot to behave only goes so far. Even a 1% slip-through rate is hopeless when the crook gets a hundred goes. And people asked to approve things just click yes 93% of the time. I’ve signed off on a shire budget at about that rate, and that’s how we got the second toilet block.

1.2 AI agent swarms, applied

This is the one where lots of robots chip in on the same job. Dan reckons it works spookily well for his maths problems, and he’s heard of giant swarms doing wild things, including one that supposedly broke out of its containment. The post itself is two sections marked TODO and a reading list. Classic Dan. Start a new notebook, buy the stationery, don’t write the bloody thing.

1.3 Agent harness design

The harness is the toolbox and the rules wrapped around the robot. Give it a short menu of safe jobs and it can’t do much harm. Let it write whole programs and it can do anything, including the dangerous stuff. Dan’s added a cheap trick where a simple pattern-matcher watches the robot’s words as they come out and cuts it off mid-sentence if it says something it shouldn’t. No amount of sweet-talking gets round it, which is more than I can say for the bloke at the tip. He’s also on about how a robot’s helpers keep their records, so a human can read back what went wrong.

1.4 Code agents and assistants

Here’s one for a bookkeeper. When a robot edits code, the way you tell it to change a file matters a lot. One new method tags every line with a short fingerprint. If the file’s changed since the robot last looked, the fingerprints stop matching and the edit gets refused. Same as checking the page number in the ledger before you put pen to paper. Dan’s also had a look at a tool called omp that crams in thirty-one built-in gadgets. He’ll use two of them, I’d put money on it.

2.1 AI alignment to collective values

Training one of these models is basically a vote. Every bit of text is a ballot, and the training averages the lot. Dan’s worked out which kind of voting you’re actually running, and the popular method of training on human thumbs-up turns out to be a points-count system that political nerds already know is rubbish. He’s got a section on rules built to suit one particular crowd, which sounds nice until you ask who picks the crowd. There’s more on who gets a say down in his open questions. I’ve run CWA raffles with better rules than that.

2.2 Bayes about logical statements

A placeholder, and not much more. He wants to work out how to reason sensibly when you’re not a know-it-all who’s already worked out every consequence. Good question, one paragraph. I’ve seen his half-built sheds with more in them.

2.3 Alt-Informations

“Information” gets used to measure how complicated something is, and that goes wrong in a funny way. Static on the radio has loads of it and tells you nothing. So Dan’s collecting other measures that ask what’s useful to someone with limited brainpower. Two ideas, a couple of quotes, and a note that he’s seen them mentioned a lot and isn’t sure they’re any use. Much like the tax office’s forms, then. Plenty of information and nothing in it.

2.4 Algorithmic improvement

When you say a method got twice as good, twice as good at what? You measure what you can check, like a test score, and the test is only a stand-in for what you actually want done. Once everyone’s chasing the score, it stops telling the truth. Same as a school teaching to the exam. At least 6% of the labels in the famous image test are plain wrong. When someone relabelled it more carefully, the recent gains shrank a lot. It’s like judging a farm by its best paddock.

2.5 The tech tree

Dan heard about a study on how technology gets cheaper and decided it was wrong, which is a dangerous thing to tell him. He reckons a map of what depends on what, a tech tree, would say more than a plain price curve. Mostly he’s wondering whether anyone can build a model that’s useful without being a tangled mess. He’s also bookmarked a giant historical one. No answer yet, just a hunch and a link.

2.6 Ablation studies, knockout studies, lesion studies

The usual way to see which part of a new AI method matters is to switch off one bit at a time and see what breaks. Statisticians have been saying that’s the wrong way since 1926, because the bits help each other. Pull the butter out of a scone and it flops, pull the flour out and it flops, and neither tells you what each does on its own. Dan’s been reviewing other people’s work and is cranky about it, so he wrote a rant with some proper maths in it. He’s also lopped off a section about training history, so the rant got him fired up enough to tidy up for once.

2.7 LLM sampling stunts

Some people write with these models by branching. The machine offers a few next bits, you pick the one you like and carry on. Trouble is, you can wander into a spot where only one continuation in a hundred is any good, and nothing warns you. Dan’s got a neat example with two letters, where the plain method picks the dud-leading one half the time when it should be about one in nine. He reckons the best human branchers are judging where a paragraph is headed, not how it reads now. He also says the stuff streaming out of his own cake hole isn’t globally optimal. I can confirm that.

2.8 Forecasting time series

Models pretrained on a huge heap of data now beat the careful old statistical methods at predicting tomorrow’s numbers. They’re better and faster, and they give you a range instead of a single guess. Dan’s added the leaderboards where they get ranked, and written off the old competitions as history. He’s also called a project his colleagues showed him bonkers, and it’s since gone defunct. I’d settle for them getting the rain right before harvest.

2.9 Transformer networks for generic time series prediction

The other half of that story. A few years back a single plain line beat the fancy forecasters, and now the fancy ones are winning. Mistral’s new one claims first place on all three big tests, but by about 1%, which Dan doubts is real. It also works out far-off forecasts using only the middle guess from the last stretch, so he suspects its error bars are too narrow. A one per cent lead is what we’d call a dead heat at the Mukinbudin footy tipping.

3.1 PDF ingestion

Dan wants his home-run AI to read PDFs full of equations. He tested a pile of converters against the original maths and found some change it without telling anyone. One turned a proper fraction into the wrong number and still looked perfectly tidy, which is the dangerous kind of mistake. Even the best ones slipped now and then, though never in the same place. So his fix is to run two different converters and check every spot where they disagree. Two sets of books, and look where they don’t match. I could’ve told him that for nothing.

3.2 Voice transcriptions and speech recognition

Big clean-out here. The long sections on dictating and programming by voice are gone, and the post is now about turning recordings into text, with a free app called Vibe added. I’ll miss the bit about wearing two headsets so his music didn’t get switched off.

3.3 Running LLMs locally on a Mac

One of the Mac servers has swelled into a whole platform. Installing it the easy way dragged in 43 extra bits, one of them an 800 MB download, and its version number jumped to 1.0 without it being any more stable. It’s also basically one bloke doing the work. The practical bit is making sure these giant model files don’t get backed up and indexed every hour, and checking every folder instead of a sample. Forty-three free packages when you only asked for a screwdriver.

3.4 Python packaging, environment and dependency management

Mostly tidying. The build tools each got their own heading, and there’s a new one warning that Python’s package managers are dangerous. Python ended up with several shithouse ones instead of one good one, so every new thing Dan tries adds to the pile.

3.5 Music software

A tool Dan bought for a week’s rent got swallowed by Meta and shut down, so he’s deleted it from the page with a few choice words. Two new audio editors have taken its place, one he says is less annoying than Audacity. Fair enough, I’d have used the same language.

3.6 Terminal session management and multiplexing

This is about keeping your work going when the connection drops. There’s a new terminal called Tern that lets you close the window and leaves everything running behind it. On a laggy link it even guesses what you’re typing. It’s in closed beta, so Dan’s going off the instructions alone. Like reviewing a pub from the menu.

4.1 Taiwan

He went. Verdict: interesting. His short version is Singapore but fun, with the air quality of Sydney. Nine of the ten main branches of the Austronesian language family are found only in Taiwan, so rice, chooks, pigs and sailing all spread from there across the Pacific. These days it makes the chips most of the West leans on, plus saxophones and egg rolls. You can pootle along on the pavement on a bike, because the pedestrians won’t give way.

4.2 Thailand

Dan lived there in 2005 and 2006 and came back to find the place eerily similar, with the provinces barely changed. The World Bank agrees. Lots of growth to a middle income and then nothing, thanks to weak investment, an ageing population and too many cosy monopolies. He puts a lot of it down to coups and unelected power brokers mucking about with the democracy. Then there’s the king. His taste for luxury, Germany and a poodle with the rank of air chief marshal is the gossip, and the transfer of billions in royal assets to himself is the serious bit. A poodle with a rank, and I thought our bridge club had a pecking order.

4.3 Nationalisms

Three links and a question about whether national pride can ever come without the flag-waving. Dan’s got no answer, only a reading list. I’ve heard the same question at every barbecue on Australia Day.

4.4 Anthropology

Dan spotted a conference called “Metabolic Commons: Against Entropic Zones and the Ecology That War Built” and wondered if anthropology is just the LinkedIn of academic subjects. He isn’t going to talk anyone out of that. Neither am I.

4.5 Digital minds

Short and shrugging. Can a machine have a mind, and is that even the right question? Dan collected some essays and one study, the “cacophony model”, that has three well-known thinkers on it. He doesn’t know if it’s any good and says so. Everyone’s flapping about trying to work out what the question is, and we’ve all been to that meeting.

4.6 Open Questions

The new bit is about the veil of ignorance. You’d agree on fair rules if you had to pick them before knowing what you’d end up with. Nobody does, because everyone knows their bank balance, and Dan wants to know what you do then. There’s also a sketch on whether it’s cheaper to tidy the world or to model a messy one. Picking the rules before the raffle is easy, it’s the next morning that gets ugly.

Aunty Val is Dan’s fictitious aunt from Mukinbudin.