oh boy!

https://openai.com/open-models/

8/5/2025 1:02:19 PM

it’s gpt-5 release day

https://www.theverge.com/news/752091/openai-gpt-5-model-announcement-github-leak

8/7/2025 12:31:27 PM

ComfyUI has a fully offline model. But there are some gigantic downloads needed.

8/8/2025 9:24:43 PM

Been building demos at work using Geminicli to take screenshots and use pyautogui to click around and do things. It’s slow but mostly works, the bugs come from being slow to verify what it did worked.

Built an automation for Sage (a hellspawn accounting software) with Claude code in about a week for my father in law. Been using Claude more than anything else for ai things.

8/14/2025 5:13:44 PM

Meta totally pausing new AI hires

8/22/2025 4:25:07 AM

Found this image model that is pretty sick: https://dinov2.metademolab.com/demos?category=segmentation

Saw it in reference to these papers: https://huggingface.co/papers/2506.01844

Basically they are taking the output of the vision model and creating actions that can be processed by a standard LLM then translating it into actions for a robot.

The SO-101 arm https://github.com/TheRobotStudio/SO-ARM100/tree/main uses this technique.

Video of someone teleoperating 2 SO101 robots: https://youtube.com/shorts/dyqhE2--H4s?si=wG-EhnNSuRbpeEv_

The data collection can be used to train a the SMOL VLA and perform somewhat novel actions (like sorting M&Ms).

[Edited on September 3, 2025 at 10:39 AM. Reason : a]

9/3/2025 10:39:30 AM