Gen AI Artist | Applied Scientist | Creative Technologist | Agents & Substrates | Midjourney Guide | ComfyUI | UE | Film R&D | Balanced Neutral Good | INTJ 💕
los angeles
Joined May 2017
- Tweets7.7K
- Following2.2K
- Followers1.5K
- Likes62K
Pinned Tweet
I keep coming back to one question:
What if the fundamental creative asset of the future isn't a model, a scene, or even an image, but an executable visual behavior?
Something that knows how to look, move, react, and render, at whatever level of dimensionality the result actually requires.
and just in case you were thinking all you'd get is content here and no fluff... i got this new detangler spray that is SO GOOD! but wow is it girlie strong smelling! phew! but sooo soffft omg. like silky candy. like if there was a giant candy dragon and it blew bubble gum kisses at fat baby unicorns rolling in confectioners sugar.
but so smooooth and soft.
there you go... fluff.
1
1
81
Techno beats while Hermes'ing = /Yolo
You can now approve Hermes actions by NODDING YOUR HEAD. With AirPods. 🤯
And somehow, that's only one highlight from a ridiculous day of Hermes development.
198 PRs merged on November 8, including 140+ plugin additions and updates.
5 changes worth knowing:
▸ AirPods Approvals lets Mac users respond to Hermes approval prompts with a nod or head shake, using a new community plugin
▸ Phone Autopilot lets Hermes learn an Android phone workflow once, then replay it without making another model call
▸ Managed oMLX brings local Apple Silicon model downloads, setup, and controls into Hermes Desktop
▸ A new on_session_open hook lets plugins take action the moment you open or resume a conversation, before the model starts working
▸ Messages and /steer commands no longer get stuck behind pending clarification questions. The default question timeout also drops from an hour to 90 seconds
Luke has the full 198 below 👇
We're not seeing the ceiling of what's possible...
We're watching people discover how much of the floor hasn't even been built yet. ✨
HyperFlow H3 Flash shipped INT8 weights with no way to run them in ComfyUI. So I built one. 🔥
The fastest H3 setup I've ever run at this resolution, and a real speedup: their own INT8 weights, copied bit-exact, running on ComfyUI's native kernels with one loader node.
One RTX 5090, 1344x768, 5 s with audio:
46.1 s, vs 66.0 s for the official GGUF
About 38 s with sparse attention
3x faster than stock H3
All clips AI-generated locally on a single 5090.
Links, nodes and weights below. 👇
7
296
