@simon have you ever tried this test with a feedback loop? (LLM gets to see its output and revise its svg)

@judell @nolan
It helps so much to have something you want to do that isn’t itself programming, but could be done via programming

(4/4)

Or we can refuse. And they can tear us down anyway, because we choose not to be morally empty and they hate that. They want everyone as hollow as they are so they can feel clean.
Funding DHS means funding ICE. No one gets to pretend they didn’t know what that money buys.
If you’re reading […]

(3/3)

We are watching people we love put themselves in danger to bear witness, to document what is being done in our neighborhoods. Sometimes it feels hopeless because the violations come with such boldness, and such volume, that each one barely registers. It becomes a blur until they cross the […]

(2/3)

Minnesota spent years choosing to be a safer place for people. We welcome immigrants and refugees. LGBTQ+ families moved here in droves because we built protections while they are being torn away in other states. We codified abortion rights and other freedoms that other states have been […]

They are driving around in minivans and roaming the streets to harass and abduct people. Conducting traffic stops randomly against people of color, including cars with children of color. There are reports they have begun putting up checkpoints at highway exits.

2/2
Specifically the Voice Mode LLM will give some response, and I'll say something like "no, you're missing this idea, which is why X". Then it replies "Exactly, spot on. Sometimes X is true because..."

And in retrospect yes, the LLM is very capable of mapping my problem space to X, even if it […]

First draft (drafted by an LLM) called CTX_MARK: https://gist.github.com/ianb/0aa30f94a12bb67b920f0993a42d0dc5

I'm not really happy with how collapsed content works.

It's heavy on highly structured content, but I would expect to use the same patterns for broad prompt structure and immutable […]

Looking around at the (fairly sparse) state of the art, some thoughts:

1. A fully normalized format isn't good. For instance a standard JSON load with timestamps, children referenced by ids, etc. Way too much distracting boilerplate (LLMs are very distractable!)
2. Though it looks crude, I see […]

I've been playing with LLM-assisted language learning and produced something along these lines for that purpose (see the attached image).

I find this is quite supportive when I don't know the language at all; lots of LLM learning tools only work when you […]

[Original post on hachyderm.io]