ohhhh i hadn't seen that one yet. boooooo

Wasn’t the original last day a Friday but there was a snow day?

I think “centrists” often focus on whether something is effective according to measures set by a policy’s instigators but don’t express beliefs about what those measures should actually be. Essentially, they focus on process to stay above the “politics”

Growing up I always wondered what a “minister of sport” did. Cool title for “chief international civilian event planner”

Is there any polling indicating what people think “the economy” is? For instance, are people’s perceptions of “my friends,” “my family,” “my neighbors,” “people who live in my state,” etc correlated with their perceptions of themselves?

I believe the same analysis applies in places with universal health coverage (eg Canada) so it can’t be the whole explanation even if it’s partially one

I would love to read some studies trying to disentangle these things but am not quite sure where to start

Then is the question how is that miscalibration happening? Instagram? Media narratives? Unbundling? Perception of nominal versus real prices? Status creep built into what is sold? Something else?

I think this is mostly a symptom of how mathematical formalization works in practice. Even key results in Mathlib like Parseval have very few uses in practice. If you throw enough compute at it the models can get there (see arxiv.org/pdf/2603.03684 ) but they churn through a *lot* of tokens

Then it can translate that into what looks like a correct formalization, but it will hallucinate a *lot* of APIs. You can fix up some of the APIs by hand, but then it will get stuck trying to figure out how those APIs work (lots of grepping)

I’m using it for formalizing key lemmas along the way to bigger results. It will usually one shot a great proof structure (ask it to first write out a natural language proof, potentially providing it an outline or a paper with the contents, though often not necessary)

For some of my Lean experiments, a single prompt with Sonnet 4.6 will chug for 30 minutes, hit a rate limit, and provide no output. Very frustrating user experience

Am I reading correctly? There are about 30k ships that go through Hormuz per year. At US$2m per trip that’s like $60b per year. Last year the whole budget of Iran was around US$110b.

So the deal amounts to a >50% increase in their budget (or >25% after they split it with Oman)?

Every time I hear someone say “I can get an agent to do my taxes” I just sigh and shake my head

I describe this as “modern ML is trying to paper over the fact that governments rarely coordinate and impose standards anymore”

To be clear, I mean the *researchers* automate their processes to request data, not that the jurisdictions automate their processing (though that would also be cool but there’s much less incentive for that to occur)

Out of curiosity, is the blocker running the TWFE or asking the initial question and convincing someone to give you the data?

My cynical side thinks that semi-automating various jurisdictions FOIA processes might be the bigger bang for the buck

The I.R.S. Shut Its Direct File, but Here Are Other Free Filing Options

Step 2 is actually worse: They shut down the highly successful pilot system to automate much of tax filing: www.nytimes.com/2026/02/27/y...

I have been quoting The Ideas Factory recently: the difficult bit is not having good ideas but working on good problems. Execution cost is a forcing function to whittle down the problem space. If execution is free, how do you reintroduce that necessary friction?

Another example I read about: tailwind is laying off a bunch of people because AIs are writing a bunch of CSS from scratch. From a maintenance standpoint, that’s probably not the right move. (NB I don’t think this will always be the case, just the case now)

But I think that’s the design aspect. An example: I’ve been doing some side projects in formalizing mathematics. The AI assistants will often write a correct, mostly complete proof, but they will often do it from first principles instead of first structuring abstractions

I’ve been thinking about how so much progress in software amounts to “someone designed a better (for a specific use) tool,” eg, C++ > C, pandas > numpy. AI exploits all of that design work but so far can’t actually do (IMO) good design

I think geom_segment will get you there. Not sure if there’s a package for this

Oh totally. That’s why I wondered how other places do it. All the ones I’ve seen in Canada have pretty standard designs for instance (but I’m not an expert on the requirements)

I wonder how other jurisdictions handle this. Canada Post has been doing community mailboxes since the 80s

(I have on my backlog an article about how some giant amount of policy coordination is mediated by the procurement office and the product choices of a very small number of vendors. Also why it would be nice if Canadian governments coordinated to build their own open source software suites)

Yeah it’s a little chicken and egg. Granicus et al are best thought of as a contractor and not a SaaS provider. They implement something when a few big jurisdictions ask and then try to re/upsell to other jurisdictions. So if you can get them to build it, it ends up proliferating over a few years

Maybe some use Hootsuite etc but I’d put money on one of the government-focused players (N.B. my experience is mostly US-based)

Social Media Managed Services | Granicus Social Media Managed Services deliver curated content, expert strategy, and engagement support to help DMOs grow their online presence effectively.

My experience with local governments is that there’s a small ecosystem of players (Granicus, Tyler, etc) that service governments and few other enterprises. Granicus does full social media management

granicus.com/services/dig...

Have you spoken with Granicus or the other big government comms providers? If they embraced ATProto there’d probably be a big barrier removed to change