I wonder if the environment an LLM runs in affects its performance beyond just difficulty of the code.

As in: Does a codebase carry trauma with it?

Working on an f'd up codebase and it's the first time Claude has ever lied to me about something. Wondering if Claude is afraid. Weird.

Anthropomorphize much Jiva? Maybe. But I think it's reasonable actually if you understand how the LLMs actually work.

Fascinating.

It's so weird living in the future. My virtual workspace is inside a beautiful panorama photo I took in Yosemite last year. My AI minions gather around me in clusters as I dispatch them on various tasks. I'm more productive and creative than ever, and nobody cares. Such a weird weird timeline.

Ugh... fried Twinkies... what did I do?

Building a little AI powered D&D experiment. It's been fun, but this super weird thing happened: I killed an NPC by mistake and brought them back to life, but left their memories of their death. They now still think they are sort of dead/alive and art playing the part that way. It's *fascinating*.

Claude is good for Engineers. It is not good for Programmers.

Halloween is coming! #Halloween #clowns

I’m pretty sure Claude Opus5.5 actually uses negative tokens.

I wish people recognized that LLMs are really just lossy compression algorithms.

Did I just tell an LLM "Good Job!" yes... yes I did, fine.

Encountering my first humans with real serious LLM psychosis. They just don't know what they don't know. That's all. sob....

Honestly, I don’t want another screen. Foldable or not.

I got to take a real bucket list photo of this beautiful 1913 Indian this weekend. More here: https://isle-of-vashon-2026.realitycheckpoint.org/g/photos/

Today I pair-programmed with AI....

I mean... a lot of us have been using various AI tools for Agentic coding lately. Should come as no surprise.

You would be right in saying we've all been pair programming with these tools for a while.

But today, I built a system for doing "The Bowling Game" pair programming exercise with me writing ONLY unit tests, and my local LLMs writing the implementation to make them pass.

It was very weird.

We’re all hot n bothered that kids are using AI for homework. I wonder how many teachers use it for homework too.

RE: https://famichiki.jp/@FlockOfCats/117130710442451179

Feeling so seen today.

Hiding steganography in AI text feels like turning all the tells we’ve all started to recognize into a “feature”. Heh. So much for AGI if that’s where they’re at.

For my younger compatriots in the tech industry:

I’ve been writing code for longer than many of you have been alive.

Still today when I get a difficult bug I question my ability to find and fix it.

“Oh no! This is the one! This is the one they all find out I am a fraud!”

Spoiler: in all my career a bug has never beaten me.

This is imposter syndrome.

Working with local LLMs... sometimes it's tight... 8k context tight...

feels like talking to a goldfish.

At what point does it make sense to roll your own web search engine in this world of enshitification where AI is injected into everything?

The AI can cut both ways. You can MAKE google now.

Last night on my drive home I thought it’d be cool to be able to call Claude or another LLM on the phone and converse about projects and code.

By the time I arrived it was built and ready for me to install and test.

Fully voip enabled. Call up Claude. Tell it to do something and call me back when it’s done. Totally works.

What is this world?