Requests for Startups

In the 1980s, a small group of funds started using computers to analyze markets. At the time it seemed silly, but quantitative trading is now obvious. We're at a similar inflection point now, and the next Renaissance, Bridgewater, and D.E. Shaw's are going to be built on AI. The biggest funds in the world have been slow to adapt. I worked as a quant researcher at one of these funds, and when I asked compliance to let us use ChatGPT, I didn't even get a response. It made it clear to me that the hedge funds of the future won't just bolt AI onto their existing strategies. They'll use it to come up with entirely new ones. That's where the alpha is. We've already got swarms of Claude agents writing our codebases. Imagine swarms of agents doing what hedge fund traders do now - combing through 10-Ks, earnings calls, and SEC filings, synthesizing analyst ideas and making trades. An AI-native hedge fund will be the first to do this well.

Agencies have always been crazy hard to scale. Low margins, slow manual work, and the only way to grow is to add more people. But AI changes this. Now instead of selling software to customers to help them do the work, you can charge way more by using the software yourself and selling them the finished product at 100x the price. Think of a design firm that uses AI to produce custom design work for clients upfront, to win the business before the contract is even signed. Or an ad agency that uses AI to create stunning video ads without the time and expense of setting up a physical shoot. Or a law firm that uses AI to write legal docs in minutes, rather than weeks. That's why agencies of the future will look more like software companies, with software margins. And they'll scale far bigger than any agencies that exist in these fragmented markets today. If you're rethinking how agencies and service businesses of the future will be built, we'd love to hear from you.

Stablecoins are rapidly becoming critical infrastructure for global finance, yet much of the financial services layer remains unbuilt. The GENIUS and CLARITY Acts are placing stablecoins in a unique position between DeFi and TradFi, compliant but crypto-native. This creates room for financial services that offer DeFi benefits like better yield or access to tokenized real-world assets while operating under traditional compliance frameworks. Today, businesses and individuals must choose between regulated financial products with limited upside and unregulated crypto with real risk. Stablecoins sitting in the regulatory middle ground can bridge this gap, whether that's yield-bearing accounts, new investment access, or infrastructure that makes money move faster and cheaper across borders. The regulatory window is open. The rails are being laid. It's the perfect time to build something that blurs the line between the two worlds.

The first wave of AI companies has helped businesses and normal people fill in forms and complete online applications with unprecedented speed and accuracy. On the flip side, many of these forms will be received by local, state, and federal government, where they're currently printing them out and processing them by hand. Government desperately needs AI tools to deal with the huge increase that's coming down the line. And the benefit is that it will also make government much more cost-effective and responsive. We've seen hints of this digital government in places like Estonia, but we need to spread it to the rest of the world. This kind of startup is not for the faint of heart. Selling to government is extremely hard, but once you've figured out how to land your first customer, they tend to be very sticky and can expand to huge contracts.

When people talk about reindustrializing America, they usually focus on labor costs or geopolitics. But a bigger problem is hiding in plain sight: American metal mills are slow by design. If you buy rolled aluminum or steel tube in the U.S., lead times of 8 to 30 weeks are normal. Most buyers can't even purchase directly from mills. And despite high prices, mills still operate on thin margins. That's not because demand is weak or workers are unskilled—it's because the systems running these mills were designed decades ago. Production planning, scheduling, quoting, and execution are fragmented. Mills optimize for tonnage and utilization, not speed, flexibility, or margin. Short runs and spec changes are treated as disruptions instead of opportunities. Automation has lagged at the exact moment the workforce is shrinking. Material handling, changeovers, inspection, and quality control still rely on tribal knowledge held by a few experienced operators. Automation is mostly used to push more tons through a slow system, not to eliminate setup time or variability. Energy is the other half of the problem. Aluminum and steel are extremely energy-intensive, yet most mills rely on legacy power contracts and inflexible grids. New energy models—on-site generation, smarter power management, even next-generation nuclear—could dramatically reduce costs, but they're rarely designed into mills from the start. What's changed is that software and energy technology are finally good enough to rethink the entire system. AI-driven planning, real-time MES, and modern automation can compress lead times and raise margins at the same time. We think this creates an opportunity to build modern, software-defined American mills—especially in aluminum rolling and steel tube—where long lead times and energy costs are most entrenched. Modernizing mills isn't just about going faster. It's about making domestic metal cheaper, more flexible, and more profitable—and rebuilding the industrial foundation of the U.S.

You know that scene in The Matrix, where Neo plugs a cable into the back of his head and wakes up a while later and says "I know Kung Fu"? Physical work is about to get something similar – not through brain implants, but through real-time AI guidance. Much of the AI conversation focuses on which desk jobs will get replaced. But for physical work—stuff like field services, manufacturing, healthcare—AI can't yet act in the world. What it can do is see, reason, and guide the human who does. Imagine wearing a small camera while an AI sees what you see and talks you through the job: "turn off that valve", "use the ⅜ inch wrench", "that part looks worn, replace it". Instead of needing months or years of training, workers can become effective immediately, with AI coaching them and accessing new skills when needed. Why now? Three things have converged. First, multimodal models can now see and reason about real-world situations reliably. Second, the hardware is already everywhere – phones, AirPods, Smart Glasses. And third, skilled labor shortages make this economically urgent and a high wage job for millions of people. There are a few approaches you could take. The most obvious is to build this system and sell it to companies with existing workforces. Or, you could pick a vertical, like HVAC repair or nursing, and build a full-stack superpowered workforce. Or, you could build a platform that lets anyone sign up and become a skilled worker or start their own business. If you're interested in giving physical workers the same type of AI superpowers that Claude Code gives you, we'd love to see you apply.

Large language models have driven most of the recent breakthroughs in AI, but their impact has been constrained to domains that can be expressed primarily through language. Unlocking the next wave of AI capability, and enabling artificial general intelligence, will require models that are capable of spatial reasoning. Today's systems can handle limited spatial tasks, such as basic relationships or depth estimation, but they cannot robustly reason about spatial manipulation, 2D and 3D features, their relationships, or operations like mental rotation. This limits AI's ability to understand and interact with the physical world. There is an opportunity to build large-scale spatial reasoning models that treat geometry and physical structure as first-class primitives, not approximations layered on top of language. Such models would enable AI systems to reason about and design real-world objects and environments. A company that succeeds in building this capability could define the next AI foundation model, on the scale of OpenAI or Anthropic.

We want to fund startups that bring government fraud investigation into the modern era. Government is the biggest customer on earth—it spends trillions annually at the federal, state and local levels, and it hemorrhages a commensurate amount in fraud. Medicare alone loses tens of billions a year to improper payments. One of the most effective ways to claw this money back at scale is the qui tam provision under the False Claims Act. This lets private citizens file lawsuits on behalf of the government against companies defrauding it. If the case succeeds, these citizens get to keep a percentage of whatever's recovered. At the moment, this process is painfully slow: An insider tips off a law firm, and then the firm spends months or years manually pulling documents and building the case. This should be accelerated with software. Not dashboards, but intelligent systems that can take an insider tip and organize the evidence around it—parsing messy PDFs, tracing opaque corporate structures, and packaging the findings into complaint-ready files. Some startups are already filing FCA claims themselves, but we think there's a big opportunity to build tools that dramatically speed up whistleblower law firms, state AGs, and inspectors general. Founder profile matters here. We're looking for teams where at least one founder has actually done work like this, whether that's a former FCA counsel, compliance lead or auditor. Now is the time to build this: the AI capabilities are finally here, and there's bipartisan tailwind to act. If you can make fraud recovery 10x faster, you'll build a huge business — and return billions to taxpayers.

Training large language models is still surprisingly difficult. My co-founder Eric and I have spent the last three years training diffusion and language models at Can of Soup, and despite all the attention AI has received, the tooling has barely improved. On any given day we may spend significant time dealing with broken SDKs, SSHing into busted GPU instances (that you only find out are busted after spinning them up for half an hour), or discovering major bugs in open-source tooling. Not to mention the work of managing, sourcing, processing, and visualizing terabytes of data. I'd love to use products that make LLM training easy. • APIs that abstract training. • Databases to easily manage very large datasets. • Dev environments built with ML research in mind. As post-training and model specialization become more important, I could see these products becoming the foundation of how software is built in the future.