Love seeing the super app coming together x.com/i/jf/cards/mon…
David
Verified account This account is verified. Learn more@dzhng
738 Following 15K Followers 3.9K Posts
Latest posts 5
It's amazing how much progress you can make on a hobby project solo with just the right domain specific harness setup
All harvested from my software farm: github.com/dzhng/skills
All harvested from my software farm: github.com/dzhng/skills
day 15 of building my dream RTS game with AIView quoted post on X
left: the abrams I hand-coded in blender. right: the one an AI wrote as a blender script from reference photos, which I then measured and touched up.
all 37 tracked vehicles in the game now get built this way, most for $3–5 each. the cannon you hear is a real US Army tank recording. models via heybento.ai
Ok benched Gemini 4 Hydrogen against Argon for 3D modeling for @ArmataGame
TLDR Argon is still sota. Hydrogen actually turned out to be MORE expensive despite lower token costs. It tends to get itself stuck in a loop where it over-verifies everything
Quality is also not there
TLDR Argon is still sota. Hydrogen actually turned out to be MORE expensive despite lower token costs. It tends to get itself stuck in a loop where it over-verifies everything
Quality is also not there
Introducing Gemini 4 Hydrogen, an extremely capable and fast model built to bring Gemini 4 to everyone!View quoted post on X
Available to all tiers of Google AI Studio, Gemini API, Gemini for business, the Gemini app, Google Search, and paid Google AI Pro + Ultra subscribers in Google Antigravity.
Gemini 4 is probably the biggest leap since opus 5.5 for game dev
For our unit models - we've built a custom harness for 3D model generation with @GoogleDeepMind's latest Gemini 4 models (running on trybento.ai)View quoted post on X
Our Claude Code connect to it via MCP and orchestrate all 3D model creation using Gemini as sub agents
Every part of our harness (blender, model viewer, linter, renderer, ffmpeg.. etc) runs in the cloud on Bento
Now Claude has all the tools it needs to request a model, inspect it via images, pan, zoom, programmatically lint.. etc