Machine learning @ ; audio & language; Freigeisterei und Vielgeisterei; "at once a man of business and a man of rhyme"
Cambridge, MA, USA
Joined September 2017
- Tweets1.2K
- Following993
- Followers2K
- Likes31.8K
Loren Lugosch retweeted
1/4
Meet Qwenputer: Qwen3.5-4B with a full RISC-V CPU inside its weights, using the model's own Gated DeltaNet layers as RAM.
It runs C and MicroPython in its forward passes, no external runtime.
On DeepMind math: base 55–90%, base + Python tool 95–100%, Qwenputer 97.5–100%.
Loren Lugosch retweeted
We generated 458,000 RL environments to train an agent on stateful, multi-step on-device tasks. We share in our preprint how we decided along which axes to enlargen our RL env corpus, via data composition scaling laws.
📄 arxiv.org/abs/2610.37648
🧵1/9
Loren Lugosch retweeted
Introducing a routing control framework to train agentic MoEs: aligning turn-level experts with agent actions while encouraging token-level consistency leads to major performance gains.
arxiv.org/pdf/2610.07332
@lblaoke Ting-Yao Hu @cydhsieh @schowdhury671 @OncelTuzel
Loren Lugosch retweeted
Introducing RSI-Inference Bench from 🍎 Apple ML Research
🌐 rsi-inference.com
💻 github.com/apple-aiml-resear…
Can AI agents improve the systems that run AI?
We spent $100K API tokens, 50K GPU hours to find out. We built environments and benchmark from real-world #vLLM and #SGLang tasks, covering the full lifecycle of model inference engineering:
🚀 Day-0 Model Support
⚡ Inference Optimization
🔧 Bug Fix
→ Unsaturated: best agent at 26.7% pass@1
→ Comprehensive coverage of the inference stack
→ Verifiable rewards: binary and continuous
Watch the video to find out more!
Loren Lugosch retweeted
So proud of the MLX team!
Muse Spark Open, on your desk: 816B params at 36+ tokens/sec on 4 Mac Studio computers.
We've been hard at work on better M5 GPU Neural Accelerator support, a smoother Thunderbolt experience (including multi-cable meshes), and kernel optimizations for MoE architectures.
Oh — and MLX LM (mlx_lm.generate, mlx_lm.server) now supports the vision modality for Muse Spark Open!
More on using MLX for distributed inference of frontier-scale models on Mac Studio:
machinelearning.apple.com/re…
Loren Lugosch retweeted
Continuous Diffusion LMs went from niche to contender this year, but the recipe is shaky: why self-conditioning? How to embed tokens? 🤨
Our new work takes your one-hot flow from 4% to 50% on TinyGSM without self-conditioning, and explores the design space of CDLMs. 🔥🚀 (1/6)
Loren Lugosch retweeted
📣 We’re hiring PhD research interns at Apple ML Research!
Focus: Efficient #multimodal models, video & memory 📹
Goal: Publish high-impact research 📝
Duration: 4–10 months (Nov '26 – Sept '27)
Building on our recent work in fast vision-language models: machinelearning.apple.com/re…
Apply at the link below and be sure to include "#efficient-ml" in your resume/cover letter so we can spot your application! 👇
jobs.apple.com/en-us/details…
Loren Lugosch retweeted
Apple presents LoopCD:
- Enables halving the number of recurrent loops while still matching or exceeding full-depth baselines
- AIME 2024 pass@1 from 61.88% to 73.33%
arxiv.org/abs/2610.02185
Loren Lugosch retweeted
We let an agent self-improve by exploring and internalizing ”on this sort of task, keep this sort of thing in mind“ bits of self-feedback. It learned to solve tasks where the original policy failed 128 attempts in a row and GRPO flatlined.
Paper: arxiv.org/abs/2609.37633
🧵1/8
Opus 5 kept making up nonsense about how something I wanted to do with the Mimi codec wouldn't work, and I had to stubbornly fight with it until an experiment finally convinced it I was right.
Possibly one of the last times in my life in which I know more than my coding agent.
3
243
Loren Lugosch retweeted
Tired of building complex environments to train AI agents? 🛑
Check out our environment-free way to generate high-quality agentic trajectories
#DigitalWorldModeling
arxiv.org/abs/2607.16900
@seanie_12 @schowdhury671 @chaojiang06 @cydhsieh Ting-Yao Hu @alexttoshev @OncelTuzel
Loren Lugosch retweeted
10 years today since we unveiled WaveNet!
Autoregression with long context before it was cool😅
Maybe it looks a bit silly now that we didn't use a Transformer, but we had a good reason: it would take another year for that to be invented🙃 Thanks @heiga_zen for the reminder!
Excited to share our latest research on speech and music generation! Hear our #WaveNets speak for themselves: goo.gl/WBKO7g
