Small Lab Cracked Computer Use Agents! They’re ACTUALLY Generalizing!

A small lab just shipped something that might be the most important AI agent demo I’ve seen this year. Standard Intelligence built FDM-1 — a computer use model trained on 11 million hours of video data that can actually generalize to any software interface. It used Blender to model a gear. It drove a real car through San Francisco using arrow keys. With less than one hour of fine-tuning data. This is what true computer use looks like. Also in today’s video: Mercury 2’s 1,000 tokens/sec diffusion AI, handcrafted 343-parameter models, Nano Banana 2 impressions, and Nvidia’s Sonic robot model.
▼ Link(s) From Today’s Video:
SI’s Computer Use FDM-1:
Inception Labs Mercury 2:
Codex Hand Crafts Model weights:
Nano banana 2:
Nvidia Robotics Model:
► MattVidPro Discord:
► Follow Me on Twitter:
► Buy me a Coffee!
▼ Extra Links of Interest:
General AI Playlist:
AI I use to edit videos:
Instagram: instagram.com/mattvidpro
Tiktok: tiktok.com/@mattvidpro
Gaming & Extras Channel:
Let’s work together!
– For brand & sponsorship inquiries:
– For all other business inquiries: [email protected]
Thanks for watching Matt Video Productions! I make all sorts of videos here on Youtube! Technology, Tutorials, and Reviews! Enjoy Your stay here, and subscribe!
All Suggestions, Thoughts And Comments Are Greatly Appreciated… Because I Actually Read Them.
0:00 — What’s In Today’s Video (Overview)
0:45 — Standard Intelligence FDM-1: Computer Use AI
4:15 — Why This Is True Generalization
5:10 — Mercury 2: Diffusion AI With Reasoning
5:45 — How Diffusion LLMs Actually Work
7:15 — Testing It Live
9:40 — Handcrafted 343-Parameter AI Models
10:20 — Nano Banana 2 Early Access Impressions
11:00 — Nvidia Sonic: AI Native to a Robot Body
13:00 — Music, VR, Text Prompts — It Does All of It
13:45 — Open Source + Isaac Lab Training Explained
14:20 — Outro

source