Brian Slesinsky

It seems like the trouble is that the results of some other specialist’s problems might not generalize to *your* problems, and the only way to be sure is to do your own evals. Or maybe people will start publishing increasingly specialized industry benchmarks?

Or accordions.

The post Jim was replying to is gone, but this seems like good general advice. I read a lot of good stuff on Substack too.

Social networks seem quite useful for finding (and posting) *links* to more substantial articles and discussions, though?

Yeah, but they say you can feed it whatever you want, so I wonder what it does if you feed it a character's dialog? Is there a Sherlock Holmes persona vector?

Persona vectors: Monitoring and controlling character traits in language models A paper from Anthropic describing persona vectors and their applications to monitoring and controlling model behavior

What happens when you can give it any personality you want? Will some personalities seem more real? Or does being able to manipulate them make them seem less real?

If the AI seems upset then you could cheer it up by giving it an optimistic, bubbly personality.

www.anthropic.com/research/per...

Could also compare with the other anti-immigrant campaigns you wrote about too.

But maybe none of the above? Historical analogies suggest possibilities but they aren't a crystal ball.

The mysteries of Roman inscriptions are being solved with a new AI tool Aeneas, named after a hero from Greek and Roman mythology, can calculate when inscriptions were carved and predict lost text

I assume it's just a tool for a small part of the job, but this sounds useful?

This is great work!

I'm wondering about the persistence API. Suppose you asked Spark to write code that checks the current user against a list of authorized users before writing to the KV store. Is there another way to write to the KV store that bypasses that?

What does "communal" mean to you? The relationship between a musician and their audience can certainly be commercial. There are transactions. But it seems like it's a community too?

For me it was employers paying me too much money for some reason, not that I minded. And collectively, this sort of thing drove up the cost of housing for the entire region.

The people I know who are well off didn't do it that way, though?

They carelessly used AI because the people recruited to do the work believed in it themselves and didn't mind being sloppy. "Move fast and break things." Being willing to break things is part of Trump's brand, as is being unaccountable.

It undermines trust in government by making it untrustworthy.

Sure, it's a minor perk that's not worth it for a lot of people.

TSA precheck costs $78 for 5 years, about $16 a year. For someone who travels at least once a year, this is insignificant compared to the cost of the flight and related expenses. You will pay more for checked luggage, or for a sandwich.

Hardly an “elite status” thing. More like “planned ahead.”

I'm wondering there's a gap in public communication about flash flood risks? Do we need something like a "red flag warning" for fires, but for flooding?

If it were an airplane crash investigation, there would be discussion about what to fix.

I'm curious: which study is that?

I tried it. It doesn’t support constraints and the AI stuff didn’t work very well. I went back to OnShape.

I wonder if people would like it better if they sold tickets for a random drawing where the prize is a burrito? (Losers get a bag of chips instead.)

It seems like this calculation assumes people don't cut back on other spending and buy food instead, since food is more important. (That is, in those cases where they have something else to cut back on.) There would be some reduction but not the full amount.

*Well, that was ingenious #printbooks #devoured

simonwillison.net/2025/Jun/24/...

> It would take Iran just a few days to convert the 60% material into weapons-grade material.

www.theguardian.com/world/2024/n...

I think the biggest roadblock is widespread skepticism about the federal government, particularly now. I'm more optimistic about California.

I don't have a need for dictation myself, but the Recorder app on Pixel phones appears to have some new AI features?

android-developers.googleblog.com/2024/08/reco...

Maybe exporting the audio from Recorder would help? When looking at an individual recording, press the share icon, then select File, Audio, and “Files by Go… Download” as a destination, which will copy the audio file to your downloads folder.

Ok, I wrote a reply disagreeing with a bad take. That's helpful right?

Hmm. Maybe putting "be brief" in my custom instructions isn't doing me any favors here.

I tried asking about Iran and it was mediocre, but perhaps I'm holding it wrong. How specifically do you prompt for this?

They might if you also have the source code in the right place. Do npms commonly bundle both sourcemaps and source? (Haven't used npm in a while.)

This seems something like what LLM benchmarks attempt to do, by defining some "correct" responses in certain contexts. But there are many possible benchmarks.

Maybe for now, but the approach Deno took with jsr.io seems pretty promising? I dislike ending up in type-stripped, minimized JS in the debugger.