Kyle Vedder
@KyleVedder
Followers
1K
Following
3K
Statuses
908
Researcher @dynarobots. CS PhD @Penn
Redwood City, CA
Joined August 2014
For the record: attention, scale, and a sufficiently hard problem is all you need
0
0
10
@chris_j_paxton consider the possibility that brett wasn't the principal agent but is spinning it as such for pr reasons
0
0
1
I wonder how much faster JAX is than Pytorch to train this my money is on ~2x
Many of you asked for code & weights for π₀, we are happy to announce that we are releasing π₀ and pre-trained checkpoints in our new openpi repository! We tested the model on a few public robots, and we include code for you to fine-tune it yourself.
0
0
5
@geogristle @snoopsonar this isn't surprising, the whole point of GRPO is it's more efficient than InstructGPTs PPO. if you have enough resources you can just do InstructGPT PPO and also train the whole value network as part of your advantage estimation that GRPO avoided
0
0
2
@geogristle @snoopsonar Replicated as wrote code? Or replicated as in actually trained it at scale? I don't think there's been enough wall clock time passed since the paper dropped to even train such full size models, let alone implement + train. whole point of GRPO is it's more efficient at scale
1
0
2
excited to announce Neural Eulerian Scene Flow Fields was accepted at ICLR 2025! see you all in Singapore!
Neural Eulerian Scene Flow Fields EulerFlow is - unsupervised - state of the art for point cloud scene flow - flexible (works without tuning across domains) - canonical frame-free - also a point tracker a thread 🧵👇
3
0
22
this is true of the field, the project, and what you're doing that day and remember that good infra and vis tools are chronically undervalued. unless there's a more obvious value play, better tooling to find new insights is often the best value for your time
0
0
4
I will say that I have increased my conviction in my criticisms of the "Just Ask For Generalization" agenda for a variety of reasons including limits of LLM pre-training, PIs results, and 1x's own issues with scaling Original book review can be found here
1
0
1
squared error for values <<1 and its consequences have been a disaster for the human race
2
0
11
1
0
1
@pariljain @chris_j_paxton JP Morgan believes Waymo is unit profitable in San Francisco
Main Street: Thank God Google’s search business is so profitable, it can forever subsidize my Waymo rides. Wall Street: Waymo has positive unit economics in SF. -- JMP research report: Waymo has 63% cash contribution margins. The vehicle & sensor capex is large (~$150k/vehicle), so accounting for their depreciation, contribution margins are just slightly positive. Today Waymo operates at just 35% utilization, vs Uber at 55% in NY. Waymo should achieve higher utilization than Uber over time. If Waymo operates at 55% utilization, that’s an incremental 58% revenue at high cash contribution margin. The IRRs on the upfront vehicle & sensor cost go from ~0% to 30%+. There’s also an expectation the vehicles and sensors get more affordable. All of this ignores operating expenses (R&D, G&A, centralized mgmt), but suggests Waymo can profitably scale its fleet.
1
0
2
@chris_j_paxton yeah they're rolling pieces into main company L2/L3 effort, similar to Latitude
1
0
1