jtmoulia’s intermittent journal, guides, and recipes. See recent posts below, and links to archives in the header.

Atlanta to Home


Took off early this morning for home, basically a west-coast red-eye where we took off around 4a PT and made it in round 9a. Super tired, but had a great time out in Atlanta. Jeju was a highlight, and I’ll be searching for the San Jose version of Korean Baths.

Read more ⟶

Continuing to teach the LLM Starforged


Saturday RL’ll be coming round for a bike ride, so tomorrow I need to make sure everything is pumped and good to go. My poor surly has been flat for too long. Weather looks a bit drizzly that morning, but

Also looking forward to the Miami F1 race this weekend. Would love to see Sargent get in the mix for his home state race, despite the Williams.

GPT-4-GM

I’ve been seeing how much of the Starforged TTRPG ruleset GPT-4 can pick up if it’s fed in. Unsurprisingly, the answer is “a lot”, though the rules themselves are not super concise.

Read more ⟶

LLMs as GMs


Keeping with the erosion of humanity, I watched Villeneuve’s Blade Runner 2049 (2017). My bias is at play, but it did a good job of exploring LLM-like intelligence. The way Joi would follow commands, inferring what K Joe wanted and instantly pivoting: “I don’t want to read either”.

Another deep theme in the context of the LLMs, one well dissected by 2049, is reproduction. Neither the replicants nor LLMs are able to procreate, and I believe has profound impacts on their “values”. We’re shaped by natural selection and it’s harsh demands. The movie specifies birth, not intelligence, as a point of transcendence.

Read more ⟶

Drizzllms


Some useful LLM access

OpenAssistant has whipped up the most impressive LLaMa derivative I’ve interacted with so far. Here’s to them pushing the envelope with an open approach, and here’s their interface.

Also, Azure OpenAI is available if you apply, and offers GPT-4 via an API with excellent documentation. I see what the Microsoft $10B is paying for.

Exercise

I’m going to stop tracking each exercise here. It’s totally redundant with the strong android app I use, and is still too manual to write. It’s working for me as a tracking app and while I’d like to build my own, this is one I can save for the AI’s

Read more ⟶

Friday to Home


Watched the Azerbaijan F1 quali this morning. Ooolala Leclerc snatching pole from the Red Bulls – fellow loves his Baku.

I got access to OpenAI’s GPT-4 API and an alpha of the plugin interface today. The plugins tools seem to 1) use a custom model, and 2) still be incredibly [alpha] slow. I’ll be looking to develop a plugin of my own to get smart with the technology, though still figuring out what’d be the best integration 🤔

Read more ⟶