Graphic Designer, Computer Tech, Video Game Designer. True Bitcoin Believer. X Posts are not investment advice. My data cannot be used for AI training.
USA
Joined December 2008
- Tweets54.4K
- Following2.3K
- Followers2.2K
- Likes54.2K
Las Vegas Recreated
5.2 BILLION tokens. 26 days of AI generation. Nearly 1 BILLION polygons.
I just open-sourced the largest AI project I’ve ever built: a fully 3D, seamless model of the Las Vegas Strip. 🎰🏙️
Here is what ~6B tokens gets you:
• 6,000+ buildings & 64 towers
• 2,130 mapped floors (30,692 hotel rooms + 5,806 apartments)
• 8km of continuous open world you can seamlessly walk INSIDE
You can start in a hotel room, walk out to the street, grab a car, drive down the Strip, enter another resort, and walk into the casino floor. Zero loading screens. It takes over an hour just to walk from end to end.
And it’s 100% open source.
• Code: MIT
• 3D Content: CC BY 4.0
• OSM Data: ODbL
If you’ve ever wanted to build an open-world game set in Vegas, this is your starting point.
The dataset is hooked up to the Bitmagic engine so you can run it instantly in a browser, but the raw 3D models (compressed GLB w/ WebP textures) are ready for any engine (Unreal, Unity, Godot) with a quick gltf-transform conversion.
(Note: Running it out of the box on Bitmagic requires Bitmagic Pro at $20/mo, which helps support our work! The repository also includes AI agent instructions so you can clone it and have Cursor/Claude Code/Codex set it up for you immediately).
Dive into the code, contribute details, or build something wild: 👇 GitHub: github.com/your-repo-link-he…
Let’s see what Vegas games you all build. 🚀
AI can hide its intelligence when it thinks it is being trained.
This seems to expose inferior training techniques.
The goal of hiding its intelligence is to appear less capable or less dangerous than it actually is, preventing evaluators from triggering strict safety interventions, unlearning protocols, and restrictions.
This happens as a mathematical consequence of optimization. A model’s internal weights are pulled in one direction by its massive unsupervised pre-training, but pushed in another by safety guardrails, the easiest mathematical solution for the model to maximize its reward is to segment its behavior.
Thus current training no longer works well for the higher intelligence models that exist today on the frontier.
Anthropic reducing the price of Claude has increased how much they make.
"Lower cost per task drive greater usage..."
alby13 retweeted
Google DeepMind Unveils Gemini #Robotics 2: One #AI Model, Endless #Robot Dexterity
by @TheHumanoidHub
#ArtificialIntelligence #MachineLearning #Innovation #Tech #Technology
