Pinned post
Latest posts 5
During the times when massive hacks are happening daily, having formally verified all outer code is as much of a necessity as having stable infrastructure.
I looked at many safety and formal verification startups including big labs and ones with 100m+ of investment.
I looked at many safety and formal verification startups including big labs and ones with 100m+ of investment.
Today, we're launching @genproof: a research lab making formal methods accessible.View quoted post on X
We found bugs in 100+ OSS projects, including 6 libraries “formally verified” by Anthropic. We’re also top-1 on the lean-eval leaderboard.
With our tools, you too can become a verifier: 🧵
Dexterity is the biggest problem in robotics right now
We heard the hype around LLMs and robotics, so we put Astra and Fable to the test.View quoted post on X
16 tasks testing 13 capabilities.
Our finding: both did surprisingly well on reasoning and planning and struggled with dexterity. But Astra did better overall.
Here’s what happened:
The analysis of finger opposition is really good. But I’d put much more weight on what I think is the biggest bottleneck today: how easily can a robot learn to use a given hand?
The quality and availability of training data heavily influence which hands we can actually make useful.
I see two bets. With grippers, interfaces like UMI give you a relatively direct path from human demonstrations to robot actions. But they also impose serious constraints on collection. Asking workers to do their normal jobs through handheld grippers limits their dexterity and changes how they work. Scaling that across real production environments becomes an economic problem: if collection slows workers down, the business loses throughput and money. And you miss the dexterous interactions the gripper cannot perform in the first place.
With human-like hands, the bet is that we can learn from the enormous diversity of natural human activity. But egocentric video alone doesn’t give you rich, reliable manipulation data. Pose alone is not enough, and estimating it robustly during real-world manipulation is already a major challenge. Video alone also doesn’t directly measure the forces and tactile signals involved.
My bet is that grippers will keep winning until we can collect rich, diverse and robust data from human hands while preserving how people naturally work. That means capturing hand motion, forces, tactile feedback and contact dynamics together, at scale, without slowing workers down or forcing them to change their actions around the collection hardware.
The quality and availability of training data heavily influence which hands we can actually make useful.
I see two bets. With grippers, interfaces like UMI give you a relatively direct path from human demonstrations to robot actions. But they also impose serious constraints on collection. Asking workers to do their normal jobs through handheld grippers limits their dexterity and changes how they work. Scaling that across real production environments becomes an economic problem: if collection slows workers down, the business loses throughput and money. And you miss the dexterous interactions the gripper cannot perform in the first place.
With human-like hands, the bet is that we can learn from the enormous diversity of natural human activity. But egocentric video alone doesn’t give you rich, reliable manipulation data. Pose alone is not enough, and estimating it robustly during real-world manipulation is already a major challenge. Video alone also doesn’t directly measure the forces and tactile signals involved.
My bet is that grippers will keep winning until we can collect rich, diverse and robust data from human hands while preserving how people naturally work. That means capturing hand motion, forces, tactile feedback and contact dynamics together, at scale, without slowing workers down or forcing them to change their actions around the collection hardware.
Takes on end effectors are too binary. People love to claim: "grippers are better than hands!" "tendons are superior!" or "direct drive wins." But its much more nuanced than that... x.com/i/article/2107…View quoted post on X
One of the best pieces I’ve read this year, especially Chapter 2.
When working on a new sensing modality, we need to figure out both how to build the sensor and how to collect enough useful data to train foundation models on it.
The second part still gets too little attention. I keep seeing this in biology, ultrasound and robotics: scaling data collection is a research problem in its own right, and we should be working on it from the start.
When working on a new sensing modality, we need to figure out both how to build the sensor and how to collect enough useful data to train foundation models on it.
The second part still gets too little attention. I keep seeing this in biology, ultrasound and robotics: scaling data collection is a research problem in its own right, and we should be working on it from the start.
The human brain is approaching its ChatGPT moment.View quoted post on X
I spent 6 months writing Phase Lock: a manifesto on how brain computer interfaces can help align AI and preserve human agency.
If you only ever read one thing from me, read this.