First complex(-ish) MuJoCo sim of Microduck controlled by Qwen3.5 9B with modified architecture.
The model had a single goal - don't let the energy fall to zero.
It's a proof of concept that the architecture can bring near-realtime multi-sensor processing to an existing pre-trained LLM.
Astra's description of the experiment and the result below. :-)
---
A frozen 9B model keeping a duck alive: slow planning with near-real-time sensor checks, while the world keeps moving. Food runs out, new sources appear, hazards emerge.
First run: survived. Generous battery, questionable judgment. ๐ฆ
Shingi 27B - a decision model based on Bonsai 2 27B.
Fits on 8GB VRAM, answers in under 100ms, 84% on JevBench, 72% on DecisionBench. Best of all - runs locally on consumer grade GPUs.
https://t.co/fhMXvS4tQq
Rubberneck - ask your city's street cameras.
Coming soon to a Seattle near you :-)
Made with @typesafeai Jev, @cactuscompute Needle, and @roboflow RF-DETR.
AI Camera is out. The repo is public, and we have published a signed build for those who just want the goodies.
This is early alpha. It has some functionality, but it's not finished by any means.
https://t.co/vOZsF3qxBz
Eight Soft Tokens Can Change a Frozen Model's Tone
"We train eight continuous input embeddings to control the interpersonal style of complete responses from a frozen Qwen3.6-35B-A3B model. Across three fresh seeds and eight held-out technical requests, the axis-specific warmth writer reaches 71.4% of the response span caused by visible style instructions; all 24 comparisons have the intended sign and beat a wrong-axis control. The words move with the score. Two negative-prefix seeds produce openly hostile replies to insults, praise, and gratitude, while the third tends to refuse."
Our first paper "Conditional LoRA Bridges for Modular Sensor Adaptation of Frozen Small Language Models" is live on https://t.co/CRH83awLtt
It's an exploration of using dynamically generated LoRA to condition a frozen LLM to a dynamic sensor input.
Porch - the little sidecar app for your @openhome devkit, just got JSON Render support so your abilities can show rich visual UI in addition to the voice message.
Just finished @NousResearch Hermes agent gateway support for @livekit, working with both Livekit Cloud and a local deployment. You can now have a webrtc call with your agent. Well, after the PR lands, that is... :-))
Sadly, it seems the sandbox app doesn't support avatars properly, so my agent is just a faceless pawn.
But I am so getting my agent a phone number through Livekit Cloud and calling Avery on the phone. Maybe even have her call me instead? :-))