VP, AI @ Atoms | Ex Meta, Arm | CMU Ph.D.
Los Altos, CA
Joined April 2009
- Tweets347
- Following218
- Followers1.3K
- Likes819
I have joined Atoms as VP of AI, where I will be building foundation models for the physical world.
For 8 years my job was making models small enough and fast enough to live on AI glasses. The physical world is a harder version of the same problem. Almost everything in the world around us is grown or mined, manufactured and moved. Industrial automation was built one task at a time, a separate stack for every machine, unable to carry a lesson from one site to the next. Foundation models break that pattern. A model that perceives a scene and predicts what happens next sees far less difference between a forklift, an excavator, and a kitchen than we do.
That means multimodal perception that fuses cameras with lidar and radar where the physics calls for it. It also means world models that predict how a scene evolves when you act on it, reasoning that holds a long task together, and action heads that turn all of it into control at the rate a machine needs. Not one of those is settled, and unlike language there is no consensus architecture to inherit yet.
Not many places let you do all of that under one roof. That is a large part of why I joined Atoms. @travisk has been building Atoms in stealth for 8 years. The machines are already in production across food, mining, and transport, with one team owning the whole problem from what the model learns to how it performs in the real world. If this works, the cost of making and moving physical things starts to approach the cost of materials and energy and very little else, which is not a better product but a different economy.
We are hiring across foundation models, robotics, and systems, so if you want your models running on something with scale and real world impact, come talk to me. Back to work.
The video of my keynote talk at this year's @EmbVisionSummit is now published.
Scaling Down is the New Scaling Up
youtu.be/amdqNKnNL-s
Vision Language Models are native 3D learners!
We just released MobileMoE, first sub-B-active-parameter MoE language model family.
MoE isn't just for 100B+ parameter models on servers. At sub-B scale, sparse expert routing lets you match dense models at 2-4x fewer FLOPs while fitting in mobile DRAM.
arxiv.org/pdf/2605.27358
Vikas Chandra retweeted
2/ the work by @vikasc and others at Meta AI shows that Trajectory Reduction Policy Optimization (dTPRO) significantly reduces the training cost of diffusion LLMs (dLLMs), a key step forward as scaled training has been a hurdle for dLLM to date. watch out for what comes from Meta here
arxiv.org/abs/2603.18806
Grateful to @sallywf and @EETimes for the thoughtful writeup of my Embedded Vision Summit keynote. The thesis in one line: the next decade of AI won't be won by the biggest model, but by the smartest, most efficient one that lives on the devices you wear!
eetimes.com/scaling-down-is-…
Audio is the most ignored perception modality in on-device AI.
Every smart glass, robot, and drone has a mic. Almost none fuse audio + vision at perception time.
Vision-only is the vibe-coded version of multimodal perception.
Vikas Chandra retweeted
the market loves to talk hardware because stock prices go brrr but
the most interesting thing happening right now is in alternative model architectures
Diffusion models couldn't reason because RL was too expensive, not because the architecture was wrong. dTRPO collapses trajectory computation to one forward pass. On a 7B model: +9.6% GPQA, +4.3% HumanEval+. The architecture question is open again.
Paper: arxiv.org/abs/2603.18806
1/ New @ieeeICASSP 2026 (Oral): "Exploring Audio Hallucination in Egocentric Video Understanding."
Audio-visual LLMs often "hear" things they didn't, inferring sounds from visual cues alone. We built a benchmark to quantify it.
2/ Setup: 300 egocentric videos, 1,000 sound-focused Q/As, with a taxonomy that separates foreground action sounds (from the user's activity) and background ambient sounds.
3/ Even Qwen2.5 Omni hits just 27.3% on foreground and 39.5% on background sound accuracy. Audio hallucination is widespread in today's AV-LLMs, and robust evaluation has to be a first-class metric for AR/wearable use cases.
📄 arxiv.org/pdf/2604.23860
Efficient Video Intelligence in 2026 🧵
Five years ago video understanding meant action recognition on Kinetics-400. Now VLMs reason over hour-long footage, foundation-grade tracking runs at 16 FPS on a phone, and one sub-100M backbone replaces four specialized encoders.
What's still hard is mostly deployment: streaming at hour-plus durations, sub-watt AR glasses, open-set anomaly detection, cross-camera reasoning, spatial grounding through cuts, closed-loop eval.
The bottleneck moved from models to the stack around them.
Long post pulling the field together, leaning on my group's work (EUPE, EfficientSAM, Efficient Track Anything, EdgeTAM, LongVU, EgoAVU, VideoAuto-R1, DepthLM, ParetoQ) placed against the parallel work in each section.
v-chandra.github.io/efficien…
Diffusion models couldn't reason because RL was too expensive, not because the architecture was wrong. dTRPO collapses trajectory computation to one forward pass. On a 7B model: +9.6% GPQA, +4.3% HumanEval+. The architecture question is open again.
Paper: arxiv.org/abs/2603.18806
Standard approach to long video: more frames, bigger context. Tempo flips it. A small VLM reads the question first, then compresses the video around it. 6B params. 8K visual tokens. Outperforms GPT-4o and Gemini 1.5 Pro on hour-long videos. arxiv.org/pdf/2604.08120
Vikas Chandra retweeted
Been thinking about what this paper really means.
"Video diffusion" and "World Models" are becoming synonymous. Neural Computers are basically video diffusion world models for terminal envs and GUI.
Lots of talk last week about automating Manim videos. In theory, we should be able to train these world models on a 10000 hours of diverse manim videos and "see where it goes"
If a NC can generate outputs to terminal commands, it should be able to generate videos like this directly from prompt too.
Without writing code.
Classical computers run programs. Agents wrap models around programs. Neural Computers ask: what if the model is the program, the memory, and the machine? New paper exploring fully learned runtimes where computation emerges from weights alone.
Paper: arxiv.org/abs/2604.06425
Vikas Chandra retweeted
the market is reacting to the memory shortage by buying Sandisk assuming that labs continue to run less optimized versions of their own models and overspend on hardware but the longer term signal to take from the memory wall is to focus on the continued scaling down of models
deployed compute has scaled 3x every 2 years while memory bandwidth has only scaled 1.6x over 20 years - every new GPU generation widens this gap - but we've also been grossly overusing memory via the common training mode. APOLLO showed that AdamW, the standard LLM optimizer, stores redundant state for every single parameter, and that coarser gradient approximations achieve the same training quality with a fraction of the memory, enabling model pre-training on a 1/8 the GPU capacity. many more proof points are coming to light re: optimizing algorithms and efficiency of smaller models. in fact small models can outperform much larger models by spending more compute at inference time
good read on this from @vikasc
🤖 Made with AI
