Brian Slesinsky

I believe that's true in practice, but an interesting exercise to prove it might be to temporarily delete one source file, have a coding agent regenerate it so it compiles and passes the tests, and then do a diff. How much knowledge is lost? Could that be captured by the tests, too?

Just as a data point, I never heard of him before and don’t know these things. I browsed his Bluesky account and read his Wikipedia article and I’m not entirely sure what he stands for.

Yeah but that means the search doesn't help me at all. The point of searching within the page is to scroll down to the post I want to look at.

The thing I was trying to figure out is why "For You" keeps recommending very ordinary posts by Kelsey Hightower to me, and I think it might just be because he has a lot of followers? Maybe posts by people with lots of followers need to be downranked more to make up for wide exposure? Example:

Also, everywhere there's a DID, put the username instead (or as well). Example: 'top curators who liked it'.

It's good to be able to get a DID, but it shouldn't be the default.

Also, searching within the page doesn't really work for some reason.

A simple change would be to make it clearer that you can just paste in your account name and don't need to know your 'did' (which is the first example).

My project started out that way, with the bot one-shotting a simple, generic link-sharing website. But since then, we've made over 900 commits as we built it out. These are all pretty routine changes that add up to something larger - a more elaborate link-sharing website.

It makes sense looking at the code it changed. Also, changing the code fixed it. It's pretty much the same thing as reviewing a bugfix made by a human.

I have standing instructions to write a failing test before changing the code. If it didn't write a new test then I tell it to write one.

My coding agent can investigate a bug and fix it, and explain what it did so *I* understand what the problem was.

I don’t see why I should care whether it “really” understood it or not. If it’s true that it didn’t “really” understand, then apparently it didn’t need to.

That’s certainly a thing that happened, particularly in the early days. Generating AI images is like that. But I haven’t seen it since I started using a coding agent in December. There are bugs but I have’t had to undo anything. It’s great at investigating and fixing bugs.

That study was done almost a year ago and the first model I tried that I thought was good didn’t come out until December. Meanwhile the tools have improved. You can’t expect studies like that to settle questions for all time in a rapidly moving field.

I've spent a fair amount of time working with a coding agent recently, and this doesn't seem to be a problem in practice. Sometimes it makes mistakes, but then it gets feedback and recovers well.

Being a bit out of date is not that bad when it can read up-to-date documentation and do experiments.

I guess it depends how you do it. Substack is apparently free, but maybe they wouldn't scale high enough for some newspapers?

I think the historical analogy works by coincidence, because mailing costs are no longer an issue. Distribution is nearly free for a website. There are different reasons why journalism needs subsidies. (I'd guess mostly labor?)

We have numbered design docs with a status field at top: "draft", "in progress", "completed". Completed docs go to a completed subdirectory. We don't look at completed docs much.

We can easily count conversations or Bluesky accounts with a database query. When we count agents, what are we counting?

Far-UVC Light Can Virtually Eliminate Airborne Virus in an Occupied Room Far-UVC light dramatically reduced airborne virus levels in a room where people were working, in the first study of the new air disinfection technology outside of an experimental setting.

It seems like far UVC machines (just coming on the market) would help a lot here?

LLMs are kind of like sails in that left free flowing they're completely useless but tightly bound and directed they can dramatically accelerate your progress

Maybe someone will take this as a challenge?

Depends what you mean by ”limited.” It’s will pick strings it wasn’t trained on. Randomly if temperature>0. The set of possible (though unlikely) strings is much larger than its training set.

The same is true of a random string generator. Very low probabilities aren’t practically possible, but…

The actual Turing test is a party game like Werewolf. As with all games it depends on the skills of the players; testing a chess bot against random undergrads wouldn't be the same as testing it against grandmasters. Who is testing LLM's using the actual Turing test against skilled players?

Wrote blog.exe.dev/expensively-... to dig into how cache reads costs dominate LLM agent conversations. Several visualizations and one terrible pun included!

There's a sense in which every possible string already exists in Borge's Library of Babel. When you write, you're picking a string that already exists.

I can't think of any practical implications, though. It seems like the normal way we use "exists," for things we actually wrote, is better?

I see it more like a dog personality. Claude Opus 4.5 makes sure you know that it’s into whatever coding task you’re doing today. I’ve tried other models that give the impression that they’re silently judging me.

I don’t really expect a smart ghost dog to bite its owner if they’re being bad.

I prefer "ghosts" because they're immaterial beings that resemble people. Maybe they'll become golems if they can get the robots to work well?

exe.dev - Persistent VMs via SSH Start VMs with persistent disks in seconds. The disk persists. You have sudo.

Alternatively check out Shelley from exe.dev, which has a nice web UI for chatting with a coding agent running in a VM. No routing through Telegram needed; you can go to the website directly. Works great for me in Chrome on Android, iPad, and laptop.

"Refuse small speed improvements especially if they make the code much slower" - oops, that doesn't make logical sense.

I find that asking a coding agent to read over my AGENTS file and offer suggestions will sometimes find stuff like this.

Archaeologists find a supersized medieval shipwreck in Denmark The sunken ship reveals that the medieval European economy was growing fast.

In Medieval Europe ships called cogs revolutionised shipping simply by their size. A cog named Svælget 2 was recently found off the coast of Copenhagen: it’s 28 metres from bow to stern and, preserved under sand, its rigging is still intact. buff.ly/tqiJIJQ
#ShareGoodNewsToo

Person: say, i am alive. Computer: I am alive. Person: oh my god.