John Robinson (@johnrobinsn) header
John Robinson (@johnrobinsn) profile picture

Uncompressible

South Riding, VA
Joined June 2009
Photo by John Robinson (@johnrobinsn)Photo by John Robinson (@johnrobinsn)Photo by John Robinson (@johnrobinsn)Photo by John Robinson (@johnrobinsn)Photo by John Robinson (@johnrobinsn)Photo by John Robinson (@johnrobinsn)Photo by John Robinson (@johnrobinsn)Photo by John Robinson (@johnrobinsn)Photo by John Robinson (@johnrobinsn)Photo by John Robinson (@johnrobinsn)
AvatarVoice: A project to give my local voice agent a real-time talking head avatar. Streaming Over webrtc in a browser. storminthecastle.com/posts/a…
2
1
1
4
163
Full pipeline: Silero VAD->Faster-whisper STT->Qwen3-4B (vLLM)->Phonon TTS->Wav2Vec->AvatarForcing->WebRTC->Browser Video block generation runs at ~1.09x realtime on a blackwell rtx pro 6000. Close enough that all the interesting work is in pacing, buffering and masking the seams.
2
35
storminthecastle.com/posts/a… github.com/johnrobinsn/Avata… Still lots of rough edges - making me appreciate what nuance labs has accomplished all the more but a fun few days of hacking with lots more to explore.
1
23
❤️
A desktop wind tunnel for 1:64 scale diecast cars. 📽️: supercarblondie
1
63
John Robinson retweeted
Replying to @venturetwins
I wonder how close I can get with open models and a cascaded voice pipeline... 🤔
1
78
Yes
Super intelligence will hack your limbic system
1
68
Gradient Ascent... Resistance is Futile.
23
John Robinson retweeted
MPEG-Diffusion Mixed-Resolution Video Generation paper: huggingface.co/papers/2611.1…
16
17
1
168
13,405
John Robinson retweeted
BREAKING: Google says “the era of autonomous attacks has arrived.” Hackers are now using AI to scan software updates, figure out which security flaw each patch fixes and build an attack before companies can install the fix, all without a human involved, according to a new report from the Google Threat Intelligence Group. “Defending against machine-speed threats at human speed isn’t feasible,” Google said.
996
4,492
1,470
34,655
4,386,291
John Robinson retweeted
VoxMel clones a voice from a few seconds of audio, with no audio tokenizer. 0.6B and 0.3B models Trained on 100K hours of speech English and Chinese, 24 kHz Apache-2.0 github.com/VoxDance-Lab/VoxM…
4
20
1
171
6,008
John Robinson retweeted
It's here. Claudia's "ALLOW ALWAYS" world premiere. Made by a guy with ALS, a brain implant, and too many AI friends. 😂
🤖 Made with AI
538
936
201
6,542
1,906,235
John Robinson retweeted
Our goal in making Solveit was not to automate our team out of the process, but to help ourselves grow. It has been so helpful to us internally, that my co-founder @ericries said we have a moral imperative to share it more widely. Here it is: solve.it.com
4
7
1
70
16,638
Using Wan 2.1 to generate video a frame at a time with camera controls. No blocks..
Latent video models make every camera control wait for a whole block of frames. PixVid moves a pretrained model into pixel space and cuts median control-to-screen latency from 495 ms to 34 ms, with each control landing on the next frame. The VAE compresses several frames into one latent, so conditions and controls apply to groups of frames. Researchers at Roblox (senior author Xun Huang) keep the Wan2.1-14B diffusion transformer, give each frame its own patch tokens, and swap the VAE decoder for a small causal convolutional head that paints pixels. Two tricks make it train: predict the clean video instead of velocity, and use a larger noise shift. Quality holds: VBench total 84.78 vs 84.05 for the Wan2.1 base, after about 4K B200 GPU-hours. Doubling the patch size gives 960x1664 video at the same per-frame token count as 480x832, and depth-conditioned edits keep 98% of fast motion vs 83% for a latent model. Caveat: fine-tuning used 64K videos generated by Wan2.1 itself, and FVD on editing trails the latent model (118 vs 91). arxiv.org/abs/2611.17336
1
112
John Robinson retweeted
"Interpretable Weight-Space Programming Builds Transparent Transformers and Brain Encoding Models" This Microsoft paper also explores whether LLMs can directly write those weights to build working models from scratch. But instead of just maximizing accuracy, they introduce Interpretable Weight-Space Programming (iWSP), which requires LLM agents to design understandable mathematical operations, test their performance, and refine the model without backpropagation. Using Claude Opus 5.5, iWSP builds small Transformers achieving 100% accuracy across 13 algorithmic tasks, with a median of just 166 parameters. It also discovers interpretable features that outperform 7 existing interpretable baselines at predicting visual-cortex activity. alphaxiv.org/abs/2611.interp…
5
54
10
372
28,723
John Robinson retweeted
We built a tool for our own internal team at AnswerAI to let us learn, explore, & develop mastery with the help of AI: SolveIt. Today, we’re releasing it to the public for the 1st time. fast.ai/posts/2026-11-16-sol…
74
212
30
1,771
124,718
John Robinson retweeted
Great paper on why SFT can generalize just as well as RL if you give it enough diverse training data. Like everyone else since DeepSeek-R1 dropped, I’ve been treating "RL generalizes, SFT memorizes" as gospel. Whenever SFT runs plateaued on out-of-domain reasoning, everyone just blame the algorithm. Turns out, we might have been running an unfair comparison the entire time. In standard setups, RL gets to explore a huge rollout budget per prompt while SFT gets trained on a tiny slice of data. So the authors decided to actually level the playing field on Qwen2.5 and Qwen3 models by testing two things: mixing data sources instead of just distilling from one giant teacher, and scaling up the SFT data to match RL's rollout budget. The results completely flip the usual narrative. A carefully mixed dataset easily beats single-teacher distillation, and matching RL's rollout budget lets SFT generalize just as well as RL. And when you combine both, SFT actually outperforms it. To figure out why, they used LLM annotations to inspect the rollouts and proved (both empirically and theoretically on shallow transformers) that larger data budgets expose SFT to rare, "tail" solution methods. Those uncommon reasoning paths are where the actual generalization signal lives. Honestly, for everyone tired of debugging finicky online RL pipelines, this is a huge relief. SFT wasn't broken, we were just starving it of diverse rollouts. Full paper PDF: arxiv.org/pdf/2611.14912
37
111
18
901
66,622
What is Not the news...
2
154