Huck Yang

@huckiyang

computer lover | agent word error rate pypi.org/project/agwer/ | past: @NvidiaAI @GoogleAI @GeorgiaTech 🔬 ex-TSMC

Earth

huckiyang.github.io

Replies Reposts Media Articles

Huck Yang
@huckiyang
Sep 3
really interesting to see array runtime inference with mlx and mlx-lm for oss local search. turning your own battery power into intelligence. 🙂
Perplexity
@perplexity_ai
Sep 2
Replying to @perplexity_ai
Hybrid compute splits work between cloud models and a local model on the Mac. Local inference must keep pace with the rest of the task. MLX-LM is a general-purpose framework, while Lily is purpose-built for this inference workload.
Huck Yang
@huckiyang
Aug 22
@yuntiandeng’s translator for native Claudish speakers
Yuntian Deng
@yuntiandeng
Aug 22
Claude has become a language, so I built a translator. English <-> Claudish
1
Huck Yang
@huckiyang
Aug 12
@SemiAnalysis_ release date ≠ pre-training cutoff. A model (i.e. Nemotron) out this year was likely pre-trained ~6 months ago - that committee data simply isn’t fully in the weights yet. imprecise takes like this circulate widely, and nobody ends up accountable for them.
SemiAnalysis
@SemiAnalysis_
Aug 12
Replying to @SemiAnalysis_ and @deepseek_ai
Furthermore, Mistral fumbled Nemotron's pretraining. NVIDIA's single-committee structure means that one member's failure drags everyone down with it, and the rest of the committee is upset. 2/3🧵
Huck Yang
@huckiyang
Jul 30
🖱️computer agents now get a "voice memory," a sub-KB inference-time only learnable acoustics 🎶. voice memory rewrites at inference, no backward pass. WER 8.36 to 7.52% over 10 domains and a new harmful edits 64% to 35%. See arxiv: arxiv.org/pdf/2607.26410 and demo:
00:00
Huck Yang
@huckiyang
Jul 27
purely a for-fun hack: running Inkling-975B on a laptop with ~35 GB of RAM to spare. since ~95% of its weights are MoE experts, a moe-offload keeps the always-needed parts resident (attention, shared experts, embeddings, norms, router, vision/audio towers, ~35 GB) and pages the