Aaron Meurer

If you want to know how good an AI is, ask it a question about something you actually care about, not proxy questions that you would never actually want to know the answer to. Ask it how it would fix a problem you had yesterday, not how many r's are in "strawberry".

So while it's easy to "own" an AI by getting it to give a stupid answer to a simple question, you shouldn't let this fool you about its capabilities for things that you'd actually use them for.

If you want to know how good an AI is, ask it a question about something you actually care about, not proxy questions that you would never actually want to know the answer to. Ask it how it would fix a problem you had yesterday, not how many r's are in "strawberry".

That's why we ask stupid questions in interviews like whiteboard coding puzzles or "what's your biggest weakness?". Those things don't actually directly matter for the job, but they correlate enough that we can infer things from them.

If you want to know how good an AI is, ask it a question about something you actually care about, not proxy questions that you would never actually want to know the answer to. Ask it how it would fix a problem you had yesterday, not how many r's are in "strawberry".

They finally added the ability to get push notifications only for mentions on this app. So while I'll still not use it much I'll at least now see it if you mention me.

PSA: if you ever want to "fork" a project in the OSS sense, don't use the GitHub "fork" feature. Create a new repository and copy the history over instead.

One interesting thing about t-strings: in most monospace fonts, t looks very similar to an upside down f. I expect t"string" and f"string" to be very commonly confused in quick scanning, even by non-dyslexics. Although a simple solution would be to use T"string" instead.

Those of use who were in school between 2005–2010 (roughly) can attest that there were teachers who told us to not use Wikipedia at all. Wikipedia was a website that could be edited by *literally anyone*, and was therefore dangerous to use as a source of information.

Kids today are of course just ignoring their teachers and using AI anyways. Because they just have to use it once to recognize that their teachers are just wrong. AIs are helping students in ways that teachers never would.

Those of use who were in school between 2005–2010 (roughly) can attest that there were teachers who told us to not use Wikipedia at all. Wikipedia was a website that could be edited by *literally anyone*, and was therefore dangerous to use as a source of information.

Today, everyone recognizes that Wikipedia is a good thing, and "anyone can edit it" is actually a powerful feature that makes it *more* reliable. But what's funny is that in 2005 kids already knew this, because we were the ones actually using it.

Those of use who were in school between 2005–2010 (roughly) can attest that there were teachers who told us to not use Wikipedia at all. Wikipedia was a website that could be edited by *literally anyone*, and was therefore dangerous to use as a source of information.

These teachers could see directly from the very definition, a website that could be edited by anyone, that it could in no way be useful. It certainly didn't help the worst students would directly copy text from Wikipedia into their assignments.

I think I'm going to give up on Bluesky, at least for now. I will probably still check it from time to time and might post stuff here if it think it would interest the people who are on here, but I'm not going to go to the effort to cross-post everything here from X anymore.

Bluesky is so sad. Such a squandered opportunity. A great protocol but nothing there. I'd be willing to live with the thousand cuts of the terrible app if there was content, but there isn't. Or if there is, there's zero discoverability. No algorithm, useless "who to follow" recommendations.

Poor man's Claude Code: - Install the Claude Mac App - Install the filesystem MCP github.com/modelcontext... - Create a Claude project and add custom instructions telling where the project is on the filesystem - Push the project up to GitHub and add this as context to the project

Note the GitHub context indirection is required if you want to use thinking, since MCP commands cannot be run to get context before thinking. Maybe there's a prompting trick to do this in two stages you could use to avoid this.

Poor man's Claude Code: - Install the Claude Mac App - Install the filesystem MCP github.com/modelcontext... - Create a Claude project and add custom instructions telling where the project is on the filesystem - Push the project up to GitHub and add this as context to the project

Can't decide if t-strings are a good idea or not.