0.0 Context Setting

Tuesday August 11, 2026 in Portland Oregon.

The freezer isn’t working. It’s super irritating. We’ve tried a whole bunch of stuff (I got the multimeter out!) and we’re getting to the point where it’s time to call someone out to take a look at it.

I’m not ready to do the switchover quite yet, but a much better archive and home for this newsletter (still powered by Buttondown!) will be launching at thingsthatcaughtmyattention.com. Let me know what you think!

0.1 Events: How People Work

Two How People Work events coming up. First, the regular live show...

0.1.1 How People Work, Live! ... with weaver

Register for How People Work, Live! ... with weaver on Friday 14 August, 9am Pacific time. Who’s weaver? This is weaver: a partner at Layer Aleph, and co-author with Marina Nitze and Mikey Dickerson of Crisis Engineering.

What’s Crisis Engineering? Here’s a blurb:

Crises are no longer rare events. They are the operating environment.

From public infrastructure and digital systems to global corporations and hospitals, the systems we rely on are failing in public, under pressure, and at scale.

Crisis Engineering starts from a hard truth most leaders avoid: crises are not just events to survive, but rare and powerful catalysts for transformation.

I’ve worked with weaver before and I’m super excited for this conversation.

After that, we’ve got:

• How People Work, Live! ... with Marcin Wichary on Friday August 21; and
• How People Work, Live! ... with Margo Stern on Friday August 28

0.1.2 How People Work... Does Happy Hour!

Starting next week on Tuesday 18 August at 1pm Pacific: How People Work Does Happy Hour!

Join me for a new weekly, live, free, and open 45-minute workshop to help you strengthen relationships, increase your influence, communicate more clearly, and apply strategy that works.

In other words: a short clinic covering the topics of my workshop, and the themes covered in the live show.

1.0 Some Things That Caught My Attention

1.1 Artificially intelligent pragmatism

I had a great catchup with Ariel Kennan earlier today, whose Axiom Foundation publicly launched a few days ago. This might be a long one because I’m tying together several vertical and horizontal layers of, I guess, how society works and is governed.

(Yes, I know. It will probably be simplistic or even not-even-wrong if you’ve studied this in depth).

First, the pragmatism part. I interviewed for a job recently (didn’t go through to the next round, oh well) that was pretty interesting; VP of State Initiatives at the Recoding America Foundation. The thing about that job is that, for what the foundation wants to achieve, they’ve accepted that using AI (sure, a nebulous term, but let’s say everything from the public chat interface models through to agentic code) is not just inevitable, but necessary. So I had to think through my position on that.

Here’s where I ended up. This implementation of AI is a tidal wave. I do not believe that any one person can stop it at this point (and yes, I’ve written before about smoking and smoking in public/places like bars and pubs -- that was a big change, a necessary and positive one, and one that only operated in a certain domain). What I can do as an individual, and what I believe I would be best at would be harm reduction. If and where it’s going to be used, I can work to make sure that it’s used responsibly, openly, auditably, and so on. In other words -- and I’ve written about this before, too -- in a way that builds on how societies that cannot function without such technologies can do so ethically.

But, some of you will say: what about the fact that these models were built on stolen data? What about the fact that their implementation in physical infrastructure is racist? What about the fact that the frontier models available now in the U.S. are at the direction of private companies, with stupendous valuations, and also subject to opaque regulation or outright management by the Trump administration?

All of those things are true; I agree with them, they are terrible. You fight datacenter placement locally and at the ballot box, and that’s happening effectively and successfully right now. I’d even argue that datacenter buildout is being politicized in the least-worst way: that the demos are making their priorities known, and that this will result in negotiation and compromise. I will leave aside the structural issues in capital, economics, and governance, acknowledging that everything exists in a system.

But this is all happening anyway. If I can work to make it less bad, where less bad also means substantially helping people, then I will work to make it less bad in implementation, and I will also work to do so on the political and policy end at the ballot box.

What Ariel and Max are doing at Axiom is super interesting, and I’m glad they’re doing it. I have lots of questions (good ones, I hope!), and I’m intrigued about where it will go next, how it will work, and the area in which it’s working.

Government broadly works by writing legislation which can set out desired outcomes and high-level rules for implementation. These days (I say this as an old, learning about secondary legislation and SIs at uni in as my kids say “the nineteen-hundreds”) there’s a whole bunch of secondary legislation that’s much easier to update, that secondary legislation in the UK, is statutory instruments, in the US, I don’t know rule letters, stuff published in a register. That’s at the national government level. Then you get to things that are implemented at the local level, so states in the U.S. also get to layer on their own requirements, which they are good at, because the whole point of the U.S. is that it’s federal. (If you want to get into this more, then in the U.S. you devolve down to counties and cities/towns so you get a melange of rulemaking).

Once you’ve got all those rules for a program, then you’ve got to apply them. You can imagine that this is pretty complicated. For sufficiently large populations, this is technologically handed by a computer system that these days will have a business rules engine that, well, it does what it says in the name.

At that point, what normally happens is that the vendor involved has the job of encoding -- translating -- all that policy into rules that, uh, the business rules engine that will execute. I use the word encoding there on purpose because that’s exactly what it is. You’re turning it into a kind of code.

Those of you paying attention will realize that at this point we might have hit upon a huge ludicrous inefficiency, which is that every single implementation is more or less doing this rule encoding from scratch every time and there’s no one source of truth. People into computing hate this because it’s not efficient, and you should just do something once. Vendors -- companies -- on the other hand are totally fine with this because they get to charge for it more. This is a specific failure in governance and, I’ve got to say, at a high level a technically uniformed an un-savvy leadership class who don’t know how to deliver services at scale.

An aside: where programs are also implemented locally -- like some electrical IRA rebate programs -- in small communities, they’re frequently implemented by hand. On bits of paper. On a desk. Or on a spreadsheet. Implementation runs the gamut from small to large.

Those rules are they run when you apply for a program. They’re run in the order of operations when you submit material, they’re a flowchart, consultants get paid lots of money to draw out workflow diagrams with branches and decision points, they require submission of evidence, they require humans to look at applications and also make determinations. They also inevitably require an appeals process. Those rules also manifest in bottlenecks, some intentional, some less-so, some explicit, some implicit. At a very very high level, that’s how things work.

Axiom is aiming to throw all of that (well, some, I suppose, into a box and shake it all up. I don’t think it’s especially an AI-first thing, but that its utility and possibility of implementation are now much easier because of LLMs.

Here’s what Axiom say on their about page:

The Axiom Foundation publishes open, machine-readable encodings of the world's rules, starting with tax and benefit policy — statutes, regulations, and policy rules turned into cited, time-aware, executable code that anyone can run, audit, or reform.¹

OK, so publishing and making available what belongs to the public sounds great. It is great. Arguably this is filling a gap that governments themselves are failing at. Next, turning those rules into “cited, time-aware, executable code that anyone can run, audit, or reform”, also great and also I know some of you are lawyers (I certainly was) are saying HANG ON A MINUTE.

Every time people have talked about putting, say, contracts into software and making them Smart they’ve been more or less laughed at because the whole point of the contract is to also introduce ambiguity, tactically and strategically where you can’t be specific. The entire reason why we have courts and lawyers and the legal system is that they’re like dynamic type checking at execution time, they get invoked when there’s a dispute. You would expect me, someone who will consistently make fun of those wild-eyed optimists who proclaim the imminent arrival of smart contracts to finally make everything better, to be skeptical and at the very least not excited about what Axiom is doing here, so why aren’t I?

I think it’s because the detail is in the specifics and the implementation. Making “law” computable doesn’t work as a class because “law” also includes the well-known I’ll Know It When I See It problem of defining pornography. Case law as well is problematic when it relies on introducing helpful ambiguity, like what the white guy on the Clapham omnibus would think.

But like I pointed out above, in the cases of tax code, in the cases of health coverage, in the case of social services, these are already rules that are being encoded. They have designed to be encoded because it’s implicitly understood that they’re going to be processed. They are rules, and sometimes yes those rules aren’t written well. But the intent is there. And anyway: these laws, procedures, letters and so on have been implemented (however well or badly) into business rules engines.

So Axiom comes along and says: well, what if we did that just once? Or more accurately, what if we could do it n-1 times so the next time someone comes along and wants to execute those rules, they don’t need to encode them again?

First, let’s make sure what they’ve got actually works. One reason why I’m excited about what they’re doing is because their cofounder also founded Policyengine and Policybench², which are a) a computational model for, uh, modeling tax and benefit policies, and b) a, uh, benchmark for scoring LLM policy compilation quality and performance. The other thing is, lest you think this team is stupid, there is a corpora of verified cases available from the appropriate federal authorities that are authoritative for different determination types. So yes, there’s test data.

First, easy one. Those who execute the rules (the bureaucracies entrusted with delivery of government services, or even gatekeeping access to those services) now have n+1 methods of running the rules. That n+1 method costs less than other implementations of n+1, and indeed might be fully open, modulo “getting the data in” (of which, sure, agentic assisted pipelines) processing pipelines. You can see how that’s good in the short-term.

The second part is that you now have at least a baseline for introducing consistency in rule interpretation and processing precisely because the Axiom products (I see what they did there) are fundamentally open. I still think you end up with the xkcd problem of “congratulations, now you have n+1 standards”. Humans will always have opinions about things, and in my point above about how governance works in the U.S., Axiom may just end up making it easier for states to diverge/create their own rules.

The last part is access. The process and satisfying the rules are a gate. There’s a gate in doing the individual work to put together the application. There’s the gate in having that application verified. There’s a loop in perfecting that application, if needed. There is an indeterminate amount of time between application submission and a decision.

But if those rules were encoded and open in a way that was easy for anyone to apply (“easy”), and in a consistent process, then for example you might have a government that commits to using say release 3 of social program rules to administer social program. If that release is public, then lawyers can also use those rules to check. Applicants can use those rules to check. There are so many open questions here:

Each program -- indeed government -- has to be, or has been designed (I guess “designed”) with a series of assumptions as to usage. We’ve got limited usage, so we’ve got to budget. Programs have funding and entitlements to fund them. Lots of those assumptions get blown out of the window now; you get people asking ChatGPT questions about whether they’re owed benefits or not, ChatGPT tells them (inaccurately, sometimes!) that they are and then before you know it, callcenters are overwhelmed with angry claimants/applicants.

But then OpenAI wants ChatGPT to be more accurate and for people to get a correct answer; the product needs to be a useful one. Or rather they’ve prioritized usefulness. The model grows, the training grows to encompass more and more.

And that’s because rules have been made more transparent when they’ve been opaque. It would be hard to argue that rules like that should be kept hidden, you kind of throw the whole idea of accountable government out the window.

This is the thing that I like. I like the increase in legibility and I like the demonstration towards it. I like the ability to verify. I like the additional axes upon which trust can be created, through auditing and so on. I like how something like how what Axiom could be written to, e.g. and Atmosphere PDS so every step is open, visible, and cryptographically verified to come from the same source, a sort of chain of custody. So for all of those reasons, I’m excited to see where they end up.

1.2 Your subjective experience is generative

Dang Nguyen posted a video³ from Big Think that reminded me about the theory of reconstructive memory⁴. I remarked at the time that

I imagine some people are going to be super upset to learn that memory recall is generative and also in a sense hallucinatory (I find this stuff fascinating)⁵

Like, do you get it? Under this theory, a memory isn’t a thing that’s encoded and that you recall. There’s a kind of memory where you reconstruct it, which would include hallucination, and has a sort of folk explanatory power in that you can remember things but also get them wrong. Remind you of anything lately?

1.3 The Profane Town Crier

I have another bot. It’s a profane town crier. That’s it, that’s the post. See:

Clang, clang, clang, you tossers. It's six fifteen.⁶

I suppose it only works for the west coast right now, but that’s because it’s where I am. I could make ones for each timezone, UGH.

OK, that’s it for now. This ended up being longer than I wanted it to, and I still feel like I’m grasping at explaining what it is about Axiom’s work that I’m interested in and why it’s interesting.

How are you doing?

Best,

Dan