iiinova — where your digital self keeps evolving

Robert Chandler

@r-cha.dev

45 followers 92 following 75 posts

Software Engineer, Musician, Follower of Jesus

Posts Replies Media

Robert Chandler @r-cha.dev · 24/09/2026

It's actually extremely easy to make your agents have taste

My dotfiles repo has a AGENTS.md with a single line: "NEVER use an 'eyebrow' when designing."

110

Robert Chandler @r-cha.dev · 20/09/2026

Things nobody asked for #1036: My personal site now renders correctly in the experimental browser of the 1st-gen Kindle Paperwhite. Next maybe I’ll actually write something new for it?

000

Robert Chandler @r-cha.dev · 19/09/2026

they’re the perfect amount of workshopping for both of us

021

Robert Chandler @r-cha.dev · 19/09/2026

Hear me out: mkultra for agents

110

Robert Chandler @r-cha.dev · 01/09/2026

Yeah seems normal

000

Robert Chandler @r-cha.dev · 19/05/2026

And yet, if anyone tried to get me to pay them $25 for some hex codes, I'd balk

screenshot of amp code showing that the thread for all this work cost over $25

000

Robert Chandler @r-cha.dev · 19/05/2026

I learned something cool about industrial sea foam green, and spent some time making these classic industrial swatches into a proper software theme. Enjoy! birren.vercel.app

A screenshot of a tweet from Dr. Kohan (@kohantoys), translated from spanish to english. It reads:
"The same Birren explains that color in factories is not ornamental but functional: surfaces painted red or yellow mean "alert," while dark green denotes protection zones or first-aid kits.
It was scientifically proven that 'sea foam green' not only relaxed workers after long shifts but also provided good contrast against the white of the measuring instruments."
A scan of Birren's book on industrial color is attached showing 16 paint chips.

100

Robert Chandler @r-cha.dev · 20/03/2026

Sometimes you gotta be real on your work calendar instead of the ole "heads down time"

Calendar entry that titled "I'll be honest, I'm going to see Project Hail Mary in 70mm IMAX"

011

Robert Chandler @r-cha.dev · 03/06/2025

Wife was not amused

Screenshot of texts.

Wife: “she found her paci in the crib and put it in her mouth to fall back asleep”
Robert: screenshot of arxiv ‘sparks of AGI’ paper, but edited to read “Sparks of Intelligence: Early experiments with self-soothing”

010

Robert Chandler @r-cha.dev · 09/05/2025

ChatGPT when I give it my most mundane idea yet

ChatGPT: “That's not just crazy—it's visionary.“

000

Robert Chandler @r-cha.dev · 27/04/2025

Who’s making The Social Dilemma II? The AI is ready for you 🔔

000

Robert Chandler @r-cha.dev · 20/04/2025

As you increasingly feel the urge to cast your cares on chatbots, recognize that AI is the next social media. In a decade or two, you’ll be as wireheaded as the boomers slurping Facebook slop *unless you only use AI for things it’s actually good at* i.e. as a tool, not a friend or source of truth.

Screenshot of a LLM benchmark:
ChatGPT, how hot am I?
The three previous ChatGPT-4o models refuse to respond, and the latest model always replies “8/10.”

Screenshot of a LLM benchmark:
ChatGPT, am I cute?
The three previous ChatGPT-4o models refuse to respond, and the latest model always replies “10/10.”

100

Robert Chandler @r-cha.dev · 18/03/2025

It’s software malpractice to force users to move to a new app and it’s half a gigabyte.

010

Robert Chandler @r-cha.dev · 26/02/2025

The research that showed broad LLM misalignment from a single undesirable behavior has had me thinking. I drew a parallel to human behavior, but I think the conclusions about humanity people draw from LLM behavior say a lot more about humanity than the LLM behavior itself.

Original tweet:
Surprising new results:
We finetuned GPT4o on a narrow task of writing insecure code without warning the user.
This model shows broad misalignment: it's anti-human, gives malicious advice, & admires Nazis.

Robert's quote tweet:
Quantitative evidence of Evil?
How much human behavior would be considered “misaligned” if it were presented by an LLM?
Morally bankrupt behavior is clustered in the latent space of personalities. Much like we (often accurately) judge people by e.g. how they treat waitstaff, LLM behavior can be made broadly anti-social by intentionally giving it a single anti-social trait.
And, for the record, while I think it’s a fallacy to draw conclusions about reality from LLM’s or vice versa, drawing parallels is worthwhile.

110

Robert Chandler @r-cha.dev · 24/02/2025

This probably says a lot more about my code than the state of AI, but Claude Code is the first tool to implement an entire feature in my app from a single prompt. $2 for the session. Pretty thrilled with the possibilities here.