My name is Ben. I am an AI researcher and robotics engineer.

About Me

I am the founder and CEO of a company called K-Scale Labs. We are building general-purpose robots (GPRs) to bring embodied intelligence into the real world. We are making this technology open-source and free for anyone to audit, build on, and optimize for their own use cases, because we believe this is both the best way to increase adoption of embodied intelligence and to ensure that it is maximally beneficial to humanity.

Previously, I have worked as an AI researcher and engineer at Tesla, Meta, Google and Amazon. At Tesla, I trained and deployed the first autoregressive transformer for outputting car waypoints, which required rewriting the neural network compiler for Tesla's HW3 ASIC. I also wrote the CUDA kernels to generate the ground truth data for training our voxel occupancy network, which was later adapted to be used by the Optimus robot. At Meta, I trained and deployed the first transformer model for content moderation, before transitioning into AI research. I co-developed one of the first large-scale speech foundation models (a billion parameters was large by 2021 standards), which has since been adopted by the open-source community for applications like offline voice cloning and speech generation. I also worked briefly on semantic mapping for robotics.

I obtained my degree in Mathematics and Computer Science from Emory University, where I also did extensive research in computational neuroscience as part of an NIH-funded training grant, the highlight of which was meeting the woman who is now my wife. Most of this grant was spent working in the Hasler lab at Georgia Tech on neuromorphic computing with analog circuits, which is where I first developed an interest in deep learning. After graduating, I lived in China for three months before starting work to practice my language skills and enjoy learning about a beautiful country.

Goals

My goal for the next two decades is to build the infrastructure for a Type-I Kardashev-scale civilization, and make this technology maximally beneficial to humanity. Specifically, this will be achieved in three parts:

• Build the best general-purpose robot, and make it available to everyone. Bringing embodied intelligence into the world in a safe, maximally-beneficial way will require a hardware embodiment that can be audited by everyone. Unlike smartphones or computers, closed-source embodied intelligence presents an existential risk to humanity, and there is no future in which such technology is allowed to achieve its fullest potential.
• Build the best medium for embodied intelligence. Making GPRs useful and scalable in the real world will require different silicon substrates. This shift will be significantly larger than the shift from desktop computers to smartphones, because neural networks are a fundamentally different software paradigm. Achieving the best performance will require thinking outside of the Von Neumann architecture.
• Understand human consciousness, and make it reproducible and transmittable. My wife is one of a very small number of people in the world who performs cyborg surgery, which is shedding light on the most likely path to a Feynman-type understanding of human consciousness. The end state of GPRs will be as its conduits.