We benchmarked GPT-6 models in DOOM by having LLM agents fight each other. Across 120 matches, we found:
- Astra had the highest win rate at 82.5%
- Sol made the fastest decisions at avg of 5.34s
- Luna had the most wins per dollar at 12.8
GitHub:
@OpenAIDevs @rootlyhq
Show more
We benchmarked GPT-6 models in DOOM by having LLM agents fight each other. Across 120 matches, we found:
- Astra had the highest win rate at 82.5%
- Sol made the fastest decisions at avg of 5.34s
- Luna had the most wins per dollar at 12.8
GitHub:
@OpenAIDevs @rootlyhq
Show more
Google open-sourced a fruit fly's brain, so I gave it controls to Dark Souls III.
Training is still early, but here's it fighting the first boss:
Google open-sourced a fruit fly's brain, so I gave it controls to Dark Souls III.
Training is still early, but here's it fighting the first boss:
GPT-6 ASTRA IS GODLY STUPID IN CREATING 3D GAMES.
this guy figured out how to create insane graphics using Astra and the trick was just image gen.
steps to recreate:
> connect Codex to blender mcp.
> paste in the concept for the game.
> tell Astra to use the Codex image gen skill to generate concept images of the target art style, then iterate until in-game screenshots look as close as possible to those, at 60fps
set the reasoning to high.
this was also one shotted in 45 minutes btw.
Show more
i built the 1st game ever on LingBot 2 world model
- fits entirely in one node graph
- VLM monitors the state in real time
- survive longer →new weapons/monsters
- powered by @robbyant_brain at 48 fps on @reactorworld
available now via API → one call = a playable world
first of its kind & you can build yours now
Show more
Adib Fallahpour, researcher at @nvidia, shared how he’s using AI during our full-room Codex event in Toronto.
The evening was full of in-depth talks that helped us imagine new ways of using AI agents beyond coding, including biology.
Thanks to Muhammad Hamza, @OpenAI Ambassador, for co-hosting, and to the 400+ who signed up.
The room only fits 150, so we're already planning the next one.
Show more
Codex’s /goal command can be used in really creative ways.
I used it to iteratively update a robotic RL reward function and fine-tune a humanoid policy to walk like the reference GIF.
Codex ran a simple loop: edit reward → train → evaluate → compare rollout video → repeat, while preserving the RL architecture and trying to match the reference behavior.
@OpenAIDevs @OpenAI @Codex
Show more
Codex’s /goal command can be used in really creative ways.
I used it to iteratively update a robotic RL reward function and fine-tune a humanoid policy to walk like the reference GIF.
Codex ran a simple loop: edit reward → train → evaluate → compare rollout video → repeat, while preserving the RL architecture and trying to match the reference behavior.
@OpenAIDevs @OpenAI @Codex
Show more
