MasterScrat

xAI is renting out their DC space. I think I read that Google is, too.

You’re probably thinking of Meta. Google renting GPUs is a less recent phenomenon, to say the least.

From HF:

The practical lesson for defenders: have a capable model you can run on your own infrastructure vetted and ready before an incident, both to avoid guardrail lockout and to keep attacker data and credentials from leaving your environment.

This sounds like a reasonable measure. Any recommendations regarding the most suitable models for this? GLM? Kimi?

How is what you describe "YC-ish"?

Playing backward is actually a great idea for testing generalization

Would be a great idea to see how much we could manipulate the latent space and whether it has some internal structure w.r.t the physical state. I guess the only unknown is how the world model would show robustness to latent states that are transformed through this network

Hey all, happy to see this here! This was a colab between General Intuition (that I’m part of), Kyutai and Epic Games.

You can read plenty of details in the blog post and tech report but the TLDR is that we trained a multiplayer world model on 10k hours of Rocket League data. We optimized it to be playable at 20fps on a single GPU.

So what you see in the demo is fully generated: there’s no graphics or physics engine. Instead it’s a 5b neural network that takes actions in and gives pixels out.

We're happy to release MIRA, a collaboration between General Intuition, Kyutai, and Epic Games.

Mira was trained on 10k hours of Rocket League data. The model has 5B parameters and runs 4-player games at 20 fps on a single B200 GPU.

We've released a playable online demo, an in-depth technical report as well as a 1k hour dataset of 4-players gameplay:

Technical report: https://mira-wm.com/paper Repo: https://github.com/mira-wm/mira

The checkpoint also weights half less than the game install! ;-)

This sounds like a strong statement with little backing. The author does infra at DeepSeek if his LinkedIn is to be trusted, and is the author of Foyer.

This is exactly what i would have needed 6 months ago

I have mixed feelings (as in, I'm unsure how to feel) about projects where the code, the README and the HN/Reddit posts are mostly AI-generated.

I feel the frustration of reading "slop", but on the other hand the projects that surface do usually bring something useful to the table.

Should we simply judge the submission based on its technical merit? Why do I feel annoyed that an otherwise cool project uses typical LLM prose? For how long will we be able to recognize LLM-generated text, and what happens when we can't?

and so does OpenAI

Clearly they need more datacenters!

Without paywall: https://archive.is/u7t8U

As we focus and compute demand grows, the Sora research team continues to focus on world simulation research to advance robotics that will help people solve real-world, physical tasks.

— https://www.businessinsider.com/openai-discontinues-sora-vid...

So yeah, focusing on world models

This has a simple pragmatic solution though:

https://duckdb.org/2024/12/06/duckdb-tpch-sf100-on-mobile#a-...

Are you saying Carmack isn't a "talented and successful developer"?

Why would a game development pedigree correlate with rejecting AI? As Carmack said:

AI tools will allow the best to reach even greater heights, while enabling smaller teams to accomplish more, and bring in some completely new creator demographics.

Have you ever noticed that the food graphics in Super Smash Bros. and Kirby Air Riders is flat “billboarded” stock images of food?

I have not... A video showing what it's talking about would have been a good addition to this article.

Is there a reliable service that plots hourly price per GPU per cloud through time?

Just so you know, this is already very much a thing on TikTok: AI-generated movie summaries with narrator voice explaining the plot while showing only major beats, reducing movie from 2h to shorts totaling 10min.

It’s honestly not the worse AI content out there! Lots of movies I wouldn’t consider watching but that I’m curious enough to see summarized (eg a movie where only the first title was good but two more were still published)

Exactly - best case scenario is Reddit/HN front page with a cool project you enjoyed working on, have some nice conversations there, reach a few 100s stars which look good on your CV, and that’s it.

If you expect more long term support you better be paying me for my time.

Without world models. you cannot achieve reliability. And without reliability, profits are limited.

Surprising to simultaneously announce the end of the road yet point to the road ahead

How much AI did you use to write up this article? It tripped up my "fake AI-written article" detector a few times despite being interesting enough to read to the end

The start is slow as well, skipping to generation 42168M is recomended.

I picture entities playing with our universe, "it starts slow but check it out at the 13.8B mark"

HN has a hatred of K8s? That’s new to me

And even for "out of distribution" code you can still ask question about how to do the same thing but more optimized, could a library help for this, why is that piece of code giving this unexpected output etc

I think the concern of "blog comments" is best left to external platforms eg HN, Reddit etc

What would be more useful would be an automated list of places where the post has been discussed (and maybe pull the top comments from there through API?)

This looks interesting, but at least on mobile, it’s riddled with too many ads to be readable.

Congrats on shipping!

I’d love to hear a bit about the ML side of things: what was your experience with various models? Do you see a clear cost vs quality tradeoff with current state of the art models? How do open vs closed models compare?