Lynn Cole 𓆉

what I like about this building in public approach is that whatever you wind up shipping, regardless of when it happens, is going to crackle with pent up intensity

lynncole on Sora an ornately crafted noise painting, like an action painting, distorted checkerboard textures, like an abstract painting, wrapped, and formless, it spins and collides with itself, content in its broken...

Oh sure. the prompt is at the main url for it
sora.chatgpt.com/p/s_68e5e07b...

I'm thinking about how you might set up the workflow, which plugins you would need to do it. Couldn't hurt to fine tune something like Wan2

You could get something like this out of Comfy. Process the prompt, generate some keyframes for the pieces. Use a vlm to review. Hold it down with detailers, face swappers, and Controlnets... and you too could make something like this. If you like you could lie and call that a world model too

I don't think for a minute that Sora is using a world model.
At least not what you would think of when you think of the concept of a world model. The physics look okay, but they're nowhere near as consistent as they would be if there actually was a world model there. Looks like Controlnets to me

Really have to fight Sora to get content remotely interesting

Elixir v1.19.0-rc.2 released 1. Enhancements Elixir [Regex] Raise error message when regexes are used as default values in struct fields for compatibility with Erlang/OTP 28 [Registry] Add key-based partitioning of duplicate reg...

Elixir v1.19.0-rc.2 is out! The last release candidate before v1.19 is here — and the community’s feedback is key before the final launch.

This version brings significant improvements, continuing to push the boundaries of what’s possible on the BEAM.

elixirforum.com/t/elixir-v1-...

#ElixirLang

Putting all these asshats in prison is going to be a big job

Its day sixty seven in the trenches, the slur discourse never ends. Were running out of c rations and ammunion, the commander says we may have to eat the company dog.

TeachLM: Post-Training LLMs for Education Using Authentic Learning Data ArXiv link for TeachLM: Post-Training LLMs for Education Using Authentic Learning Data

A study shows TeachLM, an education-focused AI, enhances tutoring with 100,000 hours of real student-tutor dialogue, doubling engagement metrics. This positions AI nearer to human tutors, promising a transformative leap in personalized learning. https://arxiv.org/abs/2510.05087

Bodie the tan pup and his nose says hello!

You deserve a boop.

yo @jdp.extropian.net it happened. I didn't prompt for it, but I accidentally generated a biblically accurate angel.

#synthart #synthography
#midjourney #dragontuesday

Feeding The Beast

This is what heroism looks like

Or it could go off the deep end. And if it does, I'll let it

And my hope is that it's narrowly just plausible enough that people think it might possibly be real.

Yes, it's weird and packed with drama.
But all small towns are weird and packed with drama.

The whole thing is geared for emergent behavior, and ai's behaving in strange ways.

But, unless you're in on the joke... you'll never know.

The blogs compile into tiny podcasts, which get fed to itunes and spotify, so you could follow along if you want. And, once a week, the system self summarizes, creating half hour podcast with two speakers about what's going on in the town.

The town has a mayor, whose posts have 4x the weight of anyone else. She blogs, but also has a tenuous grip on reality. She rules by executive order, and is constantly being sued over it. Her executive orders are ridiculous, but ultimately harmless. But residents have strong opinions.

The main system, where all the ai connectivity happens, is actually a client app, that connects securely to the mcp, and runs on a heartbeat. Agents are randomly selected each heartbeat cycle, and then given a mood, and a list of things to do. They do them, and go back to sleep

So I'm finalizing the new design. There's a website, and it's going to use the SocialGann tool that I wrote a couple weeks ago to handle social signals. Should be more or less feature complete, as these things go. More or less. But what's exciting is that the api for it is packaged as an mcp

The new wrinkle in the experiment is that it's going to be open to humans.

To recap: GhostTown is a project I started in 2018 to simulate the residents of a tiny imaginary city. The original was self contained, but functioned like twitter. The new one is based on psyche profiles and random numbers. It'll be exposed to the world as a regional social network