Hermes Canvas, a plugin for @NousResearch Hermes Agent.
Your agent plans on a live canvas beside the chat. Boxes light up as work happens, and you edit the plan mid-build.
Box status comes from hooks, not the model: real file writes and test exit codes.
https://t.co/o8CcMpBgpL
I noticed that Anthropic intentionally didn't train Fable 5.2 for GPU optimization.I wonder if it's for open source? Even so, it's still very powerful . Furthermore, Fable 5.2 can seamlessly write non grammar LLM weights without training.
A tiny black hole that lives on my Mac desktop.
Drag it over files and it pulls them into the Trash. It gets a little bigger with each one.
Double-click it and it spits the last one back out.
Hermes Canvas, a plugin for @NousResearch Hermes Agent.
Your agent plans on a live canvas beside the chat. Boxes light up as work happens, and you edit the plan mid-build.
Box status comes from hooks, not the model: real file writes and test exit codes.
https://t.co/o8CcMpBgpL
Lesson: readouts fit on one prompt learn the template, not the numbers (other phrasings fell 268 -> 31). The operand digits become prompt-independent only after block 20 attention, so the circuit now sits in blocks 19-24.
Code and notes: https://t.co/wAvA2cbCZ5 5/5
Experiment: can you add a missing skill to a small LLM by writing neurons by hand, with no training.
Qwen3-0.6B gets many 2-digit x 2-digit products wrong. I wrote ~4,200 MLP neurons into its own layers that compute the missing step. Results and failures below. 1/5
Results (one build, greedy, paired):
- Prompt form used in the fit: 642 -> 907 of 1000 new products (p~2e-56)
- Phrasings never seen in the fit: 269 -> 294 of 400 (p~0.04)
- GSM8K: no change
- Word problems: worse, 284 -> 210. It fires where it cannot read the operands. 4/5
Built a tiny menu bar monitor for my M5 Pro MacBook Pro.
CPU per core cluster, GPU load, memory pressure, CPU/GPU temps, fan rpm, system watts, disk and network. No root, no Electron.
Native Swift + IOKit/SMC. About 0.5% of one core and 26 MB RAM.
https://t.co/GdLMgoeGwz
Hermes Canvas, a plugin for @NousResearch Hermes Agent.
Your agent plans on a live canvas beside the chat. Boxes light up as work happens, and you edit the plan mid-build.
Box status comes from hooks, not the model: real file writes and test exit codes.
https://t.co/o8CcMpBgpL
All local on MLX. In this video:
- Qwen3.8 27B: chat agent
- MOSS-Image: photo
- LTX-2.5: image to video, extend
- ACE-Step 1.5: music
Also: FLUX.3 Klein + Edit, Fish Drama 3 voice, Qwen3-ASR + Nemotron 3.5 Transcribe, LTX-2.5 text/voice to video
MIT: https://t.co/nG0mE9ti3O
Everything in this video was made on one Mac. No cloud, no account, nothing uploaded.
MLX Media Workstation: a local app where a chat agent runs open image, video and music models for you on Apple Silicon.
Real run on an M5 Pro (48 GB), sped-up parts are marked.
Links below
Real run with Claude Opus 5.5 in Hermes Desktop. Sped-up parts are marked in the video.
Install:
hermes plugins install 0xbyt4/hermes-canvas --enable
Then restart the gateway and open the Canvas pane. Steps are in the README.
Hermes Canvas, a plugin for @NousResearch Hermes Agent.
Your agent plans on a live canvas beside the chat. Boxes light up as work happens, and you edit the plan mid-build.
Box status comes from hooks, not the model: real file writes and test exit codes.
https://t.co/o8CcMpBgpL
While the agent builds, you can:
- link boxes to change the order
- leave a note; it reaches the subagent on that box
- remove a box; its next write is blocked
- see writes outside the plan as drift boxes
- replay every revision
A failing test holds the agent's finish.