Chat if people come up to you the day after to talk about your karaoke performance of Thrift Shop is that good or bad
Anyways, had a blast at CoRL/Humanoids. Next year in Austin!! 🤠
to be clear, not dunking on it. the joke was that in a room full of humanoids and dexterous hands, the manipulation researchers are all talking about an autonomous vehicle
cybercab is awesome
Hosting a robotics event at Antler VC in downtown Austin from 6:30-9:00PM
We’ll have poster presentations and a robotics learning panel with CoRL researchers
We are at capacity with 600 registrations right now, but if you are from out of town and would like to attend DM me and I can try to get you in
https://t.co/MzewdMptVA
Fable 5.2 shows substantially improved skills in controlling physical robots, even on tasks it has never explicitly been trained to do.
On RoboDojo, a benchmark with a simulated two-arm robot, it outperformed other models and most AI systems built to control robots.
"Roboticists are too pain tolerant for their own good"
Earlier this year we held a Robot Learning panel with three postdoctoral researchers from UT Austin: Dr. Nathan Tsoi who works on HRI, Dr. Shuijing Liu (@dafeijing) who works on human-centered robot learning, and Dr. Claas Voelcker (@c_voelcker) who works on RL algorithms.
We covered whether better policies or better evaluation moves the field more, data-driven vs. hybrid methods, whether vision is all you need or if robots need tactile/force sensing, where robot learning sits relative to a "GPT-3 moment," and each panelist's own recent work (co-adaptive value alignment, MimicDroid, and REPPO).
Thank you to Nathan, Shuijing, and Claas for joining this conversation!
@JasonJZLiu Nice hardware and demo. The extra dof from the 7-dof Nero arms really shows in the box folding task for wrist / arm orientations. We are prototyping a similar platform. Nice job, looking forward to seeing more on your project and research.
It’s a wonderful thing when you see two great interests of your life unexpectedly collide. This is a clip from ‘His Girl Friday’ a 1940s Howard Hawks film starring Cary Grant and Rosalind Russell. The movie centers around the relationship between a newspaper editor (Grant) and an estranged reporter (Russell) as they get pulled together pursuing the biggest story of the decade. The film features my favorite example of the ‘Hawksian Woman’ character type made famous by Hawks, where the female lead is witty, competent, powerful, and full of cinematic energy. I could go on about the merits of this film at length. The writing and dialogue is wonderful, the direction and editing is nearly perfect, the performances by Grant and Russell are amongst the best in their careers…it is simply one of the great films, and a personal favorite. It is an especially good watch if you’ve seen other Hawks films and some of the other American comedies of this era.
High marks to the Nemotron team for demoing their transcription model on this clip. When Hawks was working on ‘His Girl Friday’, he envisioned a comedy with ultra-fast and overlapping dialogue exchanges (a natural feature of the fast-paced fast-talking world of the newspaper business). This style influenced later directors like Robert Altman who used overlapping dialogue to great effect.
From a model perspective, Nemotron 3.5-transcribe looks awesome. It builds on the Nemotron family that uses a mamba-2 / attention / moe architecture (great to see SSM / attn and linear attn / attn hybrid architectures continue to advance). The model is 34b with 4.3b active. I ran initial tests on a RTX Pro 6000 96gb (vLLM, bf16, 67.5gb model weights), got ~9hrs audio transcribed per minute of wall clock time with good accuracy (LibriSpeech test-clean 1.15% WER).
Meet NVIDIA Nemotron 3.5 Transcribe 🎙️
An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint.
Watch it keep track of who said what when everyone’s talking at once.
1/ 🧵
@saba_khalilnaji Just put in an order for one XT30 and three XT30(2+2). Easy to use website, and good to see there is also an API, I’ll try that at some point. Thanks for offering this
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking: