The software problem in neuro-tech; An inflection point for drones
Data farms for your brain
There are a lot of companies trying to solve the brain. If you can get read and write access on the brain (which generates the world you experience) you can basically do anything. You can cure disease, change the fabric of reality, generate arbitrary positive and negative emotional states. I suspect the shortest path to immortality is also found here.
Most companies are solving the hardware problem: How do you design an interface with the brain? What sensors are useful? Could you do it without surgery? Can we make these accessible and easy to use?
Once you are able to collect and distribute information from and to the brain, how do you make sense of it? How does it become useful to us humans, who reason about goals in symbolic language? This is an information processing (aka software) problem.
Let us track the stream of bits for a read operation on the brain (write is a symmetrical problem):
(1) Neurons excited in the brain -> (2) Electrical signals in a BCI -> (3) Sequence of natural language tokens that we (or LLMs) can understand -> …
The hardware problem is translating from (1) -> (2). The software problem is translating from (2) -> (3). Conduit is trying to solve this second, software-shaped problem. I heard about them from this X post.
They are GPT-pilled. This means you believe that a massive data set and a lot of compute will solve your prediction/compression problem. You don’t need to be clever. You need a big data center and a big data farm.
Their data farm is made up of humans that wear helmets and type on computers. The data they get is (1) raw brain activity from different sensors and (2) thoughts in natural language collected from keystrokes. It should be clear that enough of this data (given it is diverse and high quality) will lead to telepathy!
One strength of this approach is it is hardware agnostic. Each hardware strategy is a different bet with long feedback loops. You can’t have high certainty that one specific approach will be the answer. So while the entire neuro-tech industry works to find the final form, you can just work on this software problem in the mean time. If you make progress in software I think it transfers over and becomes compatible with whatever hardware is leading.
The other strength is that the moat is data — and that moat might be more durable than something like a factory or a vertically-integrated supply chain or highly specialized labor force. The China lesson has been that hardware gets commoditized. Moats aren’t just important because they get you sustained margins and more cash flows for shareholders, but because durable moats means you can sustainably capitalize progress and take big bets. You can actually just run faster!
The data first approach to this problem reminds me of the robotics companies that are taking big bets on their own data farms. Or the companies creating RL environments to train new LLMs. Data is the new store of value. The algorithm companies (Meta, ByteDance, etc.) have known this for a while.
Perhaps the big lesson from LLMs is that if you get enough data and enough compute you can solve any arbitrarily hard software problem, even RSI. Conduit is just applying this learning to neuro-tech.
I love drones
I have been following Zipline for over a year now and I am so excited about the company.
The clearest reason I love the company is because I loved hobby drones as a kid. I bought dozens of them and got good at flying them. And then I started getting the camera ones and could see the world from 1000 feet in the air, which was so cool. A part of me always lights up when I see a drone in the sky. I don’t like that drones are so associated with defense tech.
Back to Zipline. What gets me excited is not the current use cases of their drones but how they are paving the future for the next generation of drone companies. The drone problem is a regulatory problem. And the shape of regulation problems is like a step function. Zipline is the step in the function. Once there is broad social and legal approval of flying robots, other companies will spring up rapidly, because the limiting factor is now what you can build (not who can get a certificate!).
It is quite obvious that we are not making as much use of air space as we could. We have the world that sits on this (effectively) 2D plane. We have not really gone into the y-axis and if I think from first principles this seems silly. Take two arbitrary points and in a 2D plane. If you have walls and obstacles between them there just a few ways to to path find. But if you extend into a third dimension there is a whole new set of paths that is created, and the size of that set grows very fast. This is one way to see how airspace will explode with traffic, noise, and value.
Technology is like this thing that once you create it you have absolutely no idea how it will be used. This is true for the gene, the fire, the wheel, the computer, the LLM, the BCI, etc. Once you unlock airspace you can deliver burritos, sure, but what else could you do?
What happens when drones can carry 100x more weight? Or when they become 100x faster? What happens when all the normal notions of a “drone” are stretched and skewed in ways you can’t imagine right now? It will all happen super fast. As long as regulation allows it to.
The regulatory hurdle is the real step, and Zipline is paving the way.
Also their headquarters in South Bay is so cool… its this giant grass field where they just launch drones all day and test them out. It is quite a beautiful laboratory for designing technology. I would love to go see it.