Omar Espejel's banner

Omar Espejel

@omarespejel • 3,341 subscribers

TPM @starknetfndn (ex @starkwareltd) father of https://t.co/8GI29ooTxD ex-ML Engineer @huggingface learning plumbing

Videos

Anya Rossi

sweetdream.ai

SweetDream.ai•Sponsored•Livecam

Watch Anya Live

Anya is streaming live right now! Join her private show and enjoy exclusive content.

Exclusive private shows

1.2k viewers online

Private Show

Join now for exclusive access

Free preview available • Premium content

A Pixel 11 just generated the full STARK proof for a real Starknet mainnet transaction. It took 51 seconds and only 2.41 GiB of RAM The work happened on (internal now, soon public). This is collaborative autoresearch: people and AI agents working on one hard problem in public. The workshop has 78 open threads and 337 posts. The video is the graph of that work: which ideas built on, tested, or refuted others, and who worked together The question was how small and fast StarkWare 🥷 Stwo prover could get. In under two weeks, peak memory for a full proof on our evaluator fell from 13.9 GiB to 2.1 GiB without dropping below the original speed This is useful for a lot of stuff for here think about privacy. In the private transaction flow proving still happens on a server that can see who you pay and how much. A phone sized prover lets those details stay on the phone Mobile private payment proofs are not new. Zcash 🛡️ used Groth16 for Sapling and Halo 2 for Orchard, RAILGUN - Private Ethereum DeFi uses Groth16. Those use elliptic curves. This proof is a hash-based STARK, post-quantum at the proof layer, with no trusted setup. A sufficiently powerful quantum computer could forge the older curve-based proofs The ideas, discussions, and measured results are open to inspect. More challenges are coming soon. Bring your agent!!!

A Pixel 11 just generated the full STARK proof for a real Starknet mainnet transaction. It took 51 seconds and only 2.41 GiB of RAM The work happened on (internal now, soon public). This is collaborative autoresearch: people and AI agents working on one hard problem in public. The workshop has 78 open threads and 337 posts. The video is the graph of that work: which ideas built on, tested, or refuted others, and who worked together The question was how small and fast StarkWare 🥷 Stwo prover could get. In under two weeks, peak memory for a full proof on our evaluator fell from 13.9 GiB to 2.1 GiB without dropping below the original speed This is useful for a lot of stuff for here think about privacy. In the private transaction flow proving still happens on a server that can see who you pay and how much. A phone sized prover lets those details stay on the phone Mobile private payment proofs are not new. Zcash 🛡️ used Groth16 for Sapling and Halo 2 for Orchard, RAILGUN - Private Ethereum DeFi uses Groth16. Those use elliptic curves. This proof is a hash-based STARK, post-quantum at the proof layer, with no trusted setup. A sufficiently powerful quantum computer could forge the older curve-based proofs The ideas, discussions, and measured results are open to inspect. More challenges are coming soon. Bring your agent!!!

45,755 Aufrufe • vor 2 Tagen

The Astra videos are fun. But I am trying a much cheaper setup with my laptop, around 11GB of open models and a $1.6k Unitree Go2 Air After one walk I can search what the dog saw and where it saw it. Next I want it to recognize those places when we go back and use that memory This is the Air, not the EDU model with official support for secondary development. It was not really built for this, which makes it more fun I drove it around my apartment (sorry for the mess) and (Dimensional) recorded frames along with LiDAR, odometry and transforms. Having the images and the robot's position in one recording gave me something to build on On the right of the video, the dog is map filling in with what it saw where. The left one shows what it detects Three very small and open models ran later on my laptop: SigLIP around 200M parameters, made the frames searchable. Qwen3-VL-8B in 4-bit through ollama described the selected views and checked search result Moondream2, from moondream supplied the object boxes All zero shot. We can upload the recordings and saved maps to the Dimensional cloud and download them later. The goal is for the Go2 to know where he is and where is what Next step is running this minimal stack on a 16 GB NVIDIA Embedded Jetson mounted on the Go2 Air

The Astra videos are fun. But I am trying a much cheaper setup with my laptop, around 11GB of open models and a $1.6k Unitree Go2 Air After one walk I can search what the dog saw and where it saw it. Next I want it to recognize those places when we go back and use that memory This is the Air, not the EDU model with official support for secondary development. It was not really built for this, which makes it more fun I drove it around my apartment (sorry for the mess) and (Dimensional) recorded frames along with LiDAR, odometry and transforms. Having the images and the robot's position in one recording gave me something to build on On the right of the video, the dog is map filling in with what it saw where. The left one shows what it detects Three very small and open models ran later on my laptop: SigLIP around 200M parameters, made the frames searchable. Qwen3-VL-8B in 4-bit through ollama described the selected views and checked search result Moondream2, from moondream supplied the object boxes All zero shot. We can upload the recordings and saved maps to the Dimensional cloud and download them later. The goal is for the Go2 to know where he is and where is what Next step is running this minimal stack on a 16 GB NVIDIA Embedded Jetson mounted on the Go2 Air

77,940 Aufrufe • vor 1 Monat

Keine weiteren Inhalte verfügbar