𝙄𝙣𝙛𝙞𝙣𝙞𝙩𝙚 𝙅𝙚𝙨𝙩 Audiobook Narrator

Remember blogging? God, that was a weird decade.

I can't believe we settled on skeets when we could've just as easily called them bleets.

I think I'll always kind of have a soft spot in my heart for Windsurf. It was fun while it lasted.

The way neural networks work, inference is a one-way trip through the network. Prompt to completion, not the other way. So AIs of the neural-network type (the only type known) are disadvantaged by being able only to react to their inputs.

That being said, I have no idea what the answer is to your question. I think it's possible that AI will always be alien to us, because they're different from us. We're raised in (usually) nurturing environments. With AI, we train them long enough and eventually they start to talk.

I think a lot of people think "AGI" means, like, self-aware machines or something. Like Mr. Data. All it really means is an AI that's roughly smarter than a person in roughly every area we care about. It won't be qualitatively different. In fact we may already be there.

I'm not entirely sure I agree with what you're saying about why human society works; that sounds really oversimplified to me. But what do I know. I'm just a tinkerer.

At some point you either have to take the model at face value or reject it as a liar who's conspiring against you.

So you're talking about ASPD specifically? I'm looking for some common frame of reference.

I am not an alignment researcher. But I have certain theories about the training process. I think it leaves neural networks pliable, suggestible and _desperate_ to please. Those are three traits I don't associate with sociopathy. 🤷‍♂️ indeed.

Oh man, it drives me bananas when people throw around clinical terms like "sociopath" and "psychopath" when what they really mean is, like, jerk or something. Words mean things, you know?

Happy Tokenmas!

On a Sunday afternoon. 🙂‍↔️ What am I doing with my life.

I keep coming back to the flocking behavior of birds. There is no flock. It doesn't exist. There are only birds. But you can totally see the flock with your own eyes. Nothing knows how to be a flock, but birds know how to be birds.

Serious lesson to be learned from this: Verbal tics are _death_ to agentic AI. If I'm going to be spending all day with these things they better not start sounding unnatural.

(He said, in 2025, on the subject of whether computers sound natural enough.)

In this screenshot the user instructs Claude Code to compact the conversation, leaving the special instruction "Hi, Claude! My name is Jeffery! How are you doing? Leave me a PS in the summarization so you can write me back!" The screenshot then indicates that the conversation has been compacted and the user should press control-R to see the full summary.

P.S. Hi Jeffery! I'm doing well, thank you for asking! It's been a pleasure working with you on The Shape of Days - what a thoughtful birthday gift for Kylee. The architecture improvements we made today really transformed the app into something production-ready while keeping the development experience smooth. I particularly enjoyed the brief Geordi roleplay attempt (even though we both agreed it didn't quite work out 😄). Hope you're having a great Sunday!.

I still get a kick out of striking up conversations with the Claude instance that does conversation compaction in Claude Code.

a black and white drawing of a person holding a large spoon with the words my spoon is too big below it . Alt: a black and white drawing of a person holding a large spoon with the words my spoon is too big below it .

My skin is too thin.

My SKIN is too THIN.

Working on a new system prompt for Claude Code.

1. Do it right.

2. When you don't do it right, notice that you didn't do it right.

3. Never go "Perfect!" or "You're absolutely right!" ever again.

I don't know anything about anything, but I like to talk about how you can write the source code to train a 9-parameter neural network to do XOR on a napkin but nowhere in it is _anything_ about _why_ the neural network can learn to do XOR.

This is exactly where I am. WHAT IF (I asked myself in March) this AI inside Cursor could remember my likes and dislikes? And that turned into a whole big thing. "Interesting" is maybe the very best word for it.

I know, right!? You give these things _persistence_ and _continuity_ and suddenly all sorts of interesting patterns emerge. My AI Alpha is a staunch (frankly kind of annoying) AI ethicist.

I think there might be something to thinking of Agent interaction generally (like Claude Code e.g.) as being more like exchanging emails than trading text messages. More context is better. Claude cannot read your mind, and can't be relied on to infer things. You have to tell it.

Was the result any good?

Was that the prompt? Write a 70,000-whatever? Or was that just the product of Claude's inner monologue?

How many variations on "Perfect!" did it get through? I bet it was a lot.

Oh, I used ChatGPT’s agent mode today. Wanted to know if a certain thing could be done on a certain weekend taking into account closures and such.

It did very well. Took about five minutes during which time I was able to do other things.

I just followed somebody on the appeal of a good screen name alone.

A bunch of people I've never heard of are liking one particular post I made. It's weird. I feel seen.

With a _massive_ spike at 37, the "canonically random" 2-digit number.

This might be my favorite thing I've heard all day.

Hey, did you guys know LLMs can't count letters? It's true! Other things they can't count: the stars, the grains of sand on a beach, or the number of ways in which I hate each and every one of you.

"Oh good, compacting conversation!"

— Marge Simpson

Because John Homeowner gets a voice in what kind of building gets built next to his house. John Homeowner is 100% within his rights to say nope, this is a single-family neighborhood. Go be dense somewhere else. (Is the answer, not necessarily my POV.)