Vineeth Yeevani
@vyeevani.bsky.social
Working on robotics. Prev @ Apple on Vision Pro.
vyeevani.github.io
vyeevani.github.io
Is anyone else baffled by this? Why build this if you have the Optimus robot? Based on the teleoperation + current behavior cloning paradigm, it should be exceedingly trivial to learn the repetitive sequence of motions to handle cleaning this car. What am i missing?
m.youtube.com/watch?v=vVFX...
m.youtube.com/watch?v=vVFX...
This Robot Sucks
YouTube video by Tesla
m.youtube.com
February 12, 2025 at 4:33 PM
Is anyone else baffled by this? Why build this if you have the Optimus robot? Based on the teleoperation + current behavior cloning paradigm, it should be exceedingly trivial to learn the repetitive sequence of motions to handle cleaning this car. What am i missing?
m.youtube.com/watch?v=vVFX...
m.youtube.com/watch?v=vVFX...
FSD not great in SF.
1. Doesn’t follow taxi + bus lanes
2. Doesn’t understand streets have a lane that’s sometimes parking sometimes driving
3. Always messes up left turns onto van ness
4. Always have to intervene for Valencia/market intersection
1. Doesn’t follow taxi + bus lanes
2. Doesn’t understand streets have a lane that’s sometimes parking sometimes driving
3. Always messes up left turns onto van ness
4. Always have to intervene for Valencia/market intersection
February 10, 2025 at 10:03 PM
FSD not great in SF.
1. Doesn’t follow taxi + bus lanes
2. Doesn’t understand streets have a lane that’s sometimes parking sometimes driving
3. Always messes up left turns onto van ness
4. Always have to intervene for Valencia/market intersection
1. Doesn’t follow taxi + bus lanes
2. Doesn’t understand streets have a lane that’s sometimes parking sometimes driving
3. Always messes up left turns onto van ness
4. Always have to intervene for Valencia/market intersection
Deepseek r1 proves Yann LeCun’s thesis that models need physical grounding
Deepseek: good prior + easy verification = 🔥.
Good prior = video
easy verification = robotics in physical world
Deepseek: good prior + easy verification = 🔥.
Good prior = video
easy verification = robotics in physical world
February 9, 2025 at 11:33 PM
Deepseek r1 proves Yann LeCun’s thesis that models need physical grounding
Deepseek: good prior + easy verification = 🔥.
Good prior = video
easy verification = robotics in physical world
Deepseek: good prior + easy verification = 🔥.
Good prior = video
easy verification = robotics in physical world
If time travel existed, wouldn’t we have met future us by now? Maybe it’s a beacon—you can only travel back to when it was first turned on. No paradox, just a switch we haven’t flipped yet.
February 9, 2025 at 6:43 PM
If time travel existed, wouldn’t we have met future us by now? Maybe it’s a beacon—you can only travel back to when it was first turned on. No paradox, just a switch we haven’t flipped yet.
Let’s revisit code as policy in the deepseek-r1 era
February 1, 2025 at 9:55 PM
Let’s revisit code as policy in the deepseek-r1 era
Why is instant policy better than using point cloud matching/alignment approaches? Can we not just sprinkle genAI on everything without making a clear cut case for why scaling a method with DL is better than classic approach?
January 30, 2025 at 8:36 PM
Why is instant policy better than using point cloud matching/alignment approaches? Can we not just sprinkle genAI on everything without making a clear cut case for why scaling a method with DL is better than classic approach?
Been seeing “competence threshold” take hold as an idea. High level: for a model to benefit from RL, it has to have some minimal competence. My view is, competence threshold is a function of the number of rollouts per question. Larger the number of rollouts, less competent the base model has to be.
January 30, 2025 at 6:18 PM
Been seeing “competence threshold” take hold as an idea. High level: for a model to benefit from RL, it has to have some minimal competence. My view is, competence threshold is a function of the number of rollouts per question. Larger the number of rollouts, less competent the base model has to be.
Rerun is the best ml logger I’ve used. Miles better than wandb and tensorboard.
December 29, 2024 at 7:05 PM
Rerun is the best ml logger I’ve used. Miles better than wandb and tensorboard.
Anyone know how to improve sample diversity with flow matching when examples are really similar?
December 25, 2024 at 8:37 PM
Anyone know how to improve sample diversity with flow matching when examples are really similar?
Why write clean code for ml? Took me ten min to convert a video diffusion training library to flow matching.
December 23, 2024 at 6:39 PM
Why write clean code for ml? Took me ten min to convert a video diffusion training library to flow matching.
Pre-LN transformer layers don’t work for perceivers with fixed encodings
December 22, 2024 at 4:57 PM
Pre-LN transformer layers don’t work for perceivers with fixed encodings
Reposted by Vineeth Yeevani
New AI Snake Oil essay: Last month the AI industry's narrative suddenly flipped — model scaling is dead, but "inference scaling" is taking over. This has left people outside AI confused. What changed? Is AI capability progress slowing? We look at the evidence. 🧵 www.aisnakeoil.com/p/is-ai-prog...
December 19, 2024 at 12:16 PM
New AI Snake Oil essay: Last month the AI industry's narrative suddenly flipped — model scaling is dead, but "inference scaling" is taking over. This has left people outside AI confused. What changed? Is AI capability progress slowing? We look at the evidence. 🧵 www.aisnakeoil.com/p/is-ai-prog...
Next scale prediction demonstrates how to explicitly simulate CNNs using transformers.
December 16, 2024 at 3:03 AM
Next scale prediction demonstrates how to explicitly simulate CNNs using transformers.
A self driving car dataset has examples going left around a tree and right around a tree. The algorithm averages the two and goes straight into the tree. Hallucination.
December 5, 2024 at 1:01 AM
A self driving car dataset has examples going left around a tree and right around a tree. The algorithm averages the two and goes straight into the tree. Hallucination.
It’s hard to make transformer masks for nested sequences - essential when working with perceivers or robotics. Wrote a library to help github.com/vyeevani/eas...
GitHub - vyeevani/easy-transformer-masks: Really flexible masks for transformers
Really flexible masks for transformers. Contribute to vyeevani/easy-transformer-masks development by creating an account on GitHub.
github.com
November 29, 2024 at 6:06 PM
It’s hard to make transformer masks for nested sequences - essential when working with perceivers or robotics. Wrote a library to help github.com/vyeevani/eas...
Utility library for diffusion forcing in Jax. I’ve tested this on two projects and it works. github.com/vyeevani/dif...
GitHub - vyeevani/diffusion_forcing_utils: Set of utility functions that make it easy to do diffusion forcing with Jax
Set of utility functions that make it easy to do diffusion forcing with Jax - vyeevani/diffusion_forcing_utils
github.com
November 29, 2024 at 5:16 PM
Utility library for diffusion forcing in Jax. I’ve tested this on two projects and it works. github.com/vyeevani/dif...
Reposted by Vineeth Yeevani
Inspired from this, I also created one for robotics and vision :)
go.bsky.app/HcQYMj
go.bsky.app/HcQYMj
November 21, 2024 at 11:05 PM
Inspired from this, I also created one for robotics and vision :)
go.bsky.app/HcQYMj
go.bsky.app/HcQYMj
Reposted by Vineeth Yeevani
November 19, 2024 at 11:00 PM
When you build for a new product category, do you build for luxury or for the masses? Cultured meat can’t compete on cost so they are turning to luxuries with foie gras. What do people think will happen in robotics?
November 29, 2024 at 3:53 PM
When you build for a new product category, do you build for luxury or for the masses? Cultured meat can’t compete on cost so they are turning to luxuries with foie gras. What do people think will happen in robotics?
Wasted a morning stopping a TPU instance stuck in creating. If there was other good GPU alternatives I’d switch from Google cloud and never come back.
November 28, 2024 at 9:58 PM
Wasted a morning stopping a TPU instance stuck in creating. If there was other good GPU alternatives I’d switch from Google cloud and never come back.
Rust library to control feetech. Currently tested on the sts3125. gist.github.com/vyeevani/6a3...
basic code to drive a Feetech sts3125 from rust
basic code to drive a Feetech sts3125 from rust. GitHub Gist: instantly share code, notes, and snippets.
gist.github.com
November 28, 2024 at 9:51 PM
Rust library to control feetech. Currently tested on the sts3125. gist.github.com/vyeevani/6a3...
Rust library to control the Cybergear from xiaomi: gist.github.com/vyeevani/baa...
https://gist.github.com/vyeevani/baafcd0980831285fae289e00c469eb6.js"></script>
November 28, 2024 at 9:50 PM
Rust library to control the Cybergear from xiaomi: gist.github.com/vyeevani/baa...
DaxBench is awesome. Easy to use cross platform clothes folding simulations.
November 19, 2024 at 8:55 PM
DaxBench is awesome. Easy to use cross platform clothes folding simulations.