Labs

What we’re thinking about right now

• Digital and physical AI Operating Systems

  We’ve built dozens of AI Operating Systems but they all stop at the computer screen. With a majority of the world’s economic value existing in the physical world, it seems to us like the next natural progression is how to integrate these digital AI agents with their real world physical counterparts to capture a larger share of the economic pie. This doesn’t necessarily mean we’re building robots, but rather seeing how our existing mapping and terraforming engine can touch the real world through hardware that already exists where the economic value is created. How do we program a Raspberry Pi with a cheap local model to turn a CCTV camera into a safety asset in a factory? How do we integrate biometric data from a wearable with emergency dispatch to launch a drone carrying antivenom to a spider bite victim who is hiking in a remote canyon? How do we integrate sensors, drones, cameras, and other machines to sense and actuate within a physical bounded environment the same way Claude Code does on a computer? How do we make the system of digital and physical agents share the same underlying foundation and language to keep everything running smoothly and safely?

• What use cases do the new AI fueled advances in mathematics, physics, computer science, and more open up?

  Foundational AI models are solving mathematical problems that have been outstanding for hundreds of years, which rewrites the possible solution space for countless industries overnight. This is like rewriting the laws of physics overnight and it’s happening all the time. To us, that means things that were once not possible might now be worth exploring. Whether that’s new technologies, business models, markets, and more.

• How does a team of AI agents that act like employees change the economics of what’s reasonable?

  When you can run a business that has the output of 1000 with a team of 5 plus a ton of AI agents, what things become feasible that were not before? This is the same concept as the one person billion dollar company, but what if an individual’s profit was not the entire end goal? What if the end goal was creating a system to fight fires effectively that actually fits within a tight local county budget? What about building armies of AI agents to run millions of parallel experiments that would have previously taken researchers a decade? Or how do we create something amazing and use agents to bring to market quickly and cheaply enough that it allows the owner to avoid predatory outside capital?

• Making cheap models and hardware punch way above their weight

  The smartest models are expensive and it’s not clear what happens to your data when you submit a query. Also in many cases they are way overpowered for the particular use case, like driving a leased Ferrari through the McDonald’s drive-thru. As hardware improves, quantization of small models improves so you can do way more with less, hardware improves so you can actually run a smart local model on a Raspberry Pi, and fine tuning techniques improve we think you can get small models to do one or two things really well. If you are intentional about those things, you can spend a few thousand dollars one time and have a fully local AI system that perfectly fits your use case.

• Local AI models

  To play off the leased Ferrari to McDonald’s example, how do we build the fully paid off Toyota Camry of AI systems? It’s the daily driver that is reliable, paid off, and perfectly suited for your everyday use case. And most importantly you own it and as you drive it more often, you start to get a sense for what it can and can’t do. Besides just cost, data sovereignty is an important thing to consider. For law firms, businesses working on top secret work, and medical companies working with sensitive data. You can’t just not use AI in today’s day and age. Just like we’re building physical and digital AI operating systems, how do you build a fully local AI operating system that actually provides value and doesn’t need millions of dollars of computing hardware to be effective. Between setting up the hardware correctly, fine tuning smaller and dumber models to execute tasks specifically, orchestrating agents, and keeping everything on a set of servers so that your data never leaves the building, there is a lot to get right. That is exactly the kind of problem we like.

• How do you trust an AI that is sending commands to a machine that will do things in the real world?

  All signs are pointing to AI being able to actuate machines without a human in the loop, but with LLMs being probabilistic in nature, it’s important that there is a layer of trust between the LLM and the machine, so both sides verify that the command is correct and the actuation has been executed by the machine. And just as importantly, how do we create a log so humans can verify everything that happened and have detailed trace logs whenever they need it. And how does that layer sit in coordination with the rest of a digital and physical AI operating system?

• Agent to agent commerce

  What does the economy look like when agents are executing financial transactions, negotiating on behalf of their human with other agents representing their human? We’re already seeing this with search and shortlisting things on behalf of a human. But what happens when AI agents are handling the entire commerce stack? How does the go to market motion change when agents are making the full decision just based on the constraints of a human? What are the downstream effects on business models that rely on advertising as their sole source of revenue when humans are no longer the ones that need to be influenced to make a purchase? What happens to negotiations and complex sales cycles when AI agents are the ones doing all the back and forth? How do we trust the AI agents and create systems where an AI agent can access the right amount of information to effectively negotiate and not hallucinate?

• AI safety and alignment

  How do we make sure that within AI operating systems of all shapes and sizes, that these agents are strictly bounded and tight governance is enforced?