跳到正文
Air Street Press· Air Street Press·· 1 小时前AI 评分38

Air Street Capital 领投 LATO 470 万美元种子轮,用 AI 智能体做专家访谈调研

Our investment in LATO

AI 导读

Air Street Capital 领投 LATO 470 万美元种子轮,YC 及 ElevenLabs、Vercel、Snowflake、Rogo 联合创始人等天使参投。

正文 · 原文

On a recent investment opportunity, I needed to understand what made a company’s product special in a market with well-funded competitors. And I needed to do it fast. In two days, I had 14 hours of conversations with 12 category experts, buyers, and technical leaders across the US, Asia, and Europe. Except, I hadn’t personally taken any of these calls.

As I shared with Matt Turck on the MAD Podcast, my Codex agents and LATO designed the study together. They used the work I’d already done to refine the hypotheses I needed to test. LATO recruited the experts and its voice agents interviewed them. My agents then worked with LATO to analyze the transcripts and revise the questions for the next interviews. I could go straight from a finding that piqued my interest to the conversation behind it and hear what the person actually said.

If you’re AI-pilled, you’ll see where this is going. A recursive research system that learns from your taste and feedback can take on more of the work each time you use it.

That’s why I’m excited to share that Air Street Capital led LATO’s $4.7M seed round alongside YC and angels including co-founders of ElevenLabs, Vercel, Snowflake, and Rogo.

Who can you get on the phone?

The investment world runs on primary research. You want to understand why a customer picked one product over another, how they weighed the trade-offs, and what would make them change their minds. The people with the best answers might sit three introductions away, work in another time zone, speak a language you don’t, or just be inaccessible to you. Being too slow can mean you miss the deal. Being ill-informed is, well, even worse.

With LATO, I can start with a hypothesis before I know whom to call. Its agents help work out which questions would offer the most alpha, then find and interview experts while I get on with everything else. If two customers disagree, I can listen to both and decide whom to ask next. As voice agents and frontier models improve, these interviews can get better at unpacking nuance, detecting uncertainty, and following up on a surface-level answer.

Probing the AI infrastructure buildout

For its first public study, Ear to the Ground, LATO asked how much of the promised AI infrastructure will be ready when it’s needed. Anthropic alone has disclosed roughly $901B in future payment commitments, almost all for compute and infrastructure. LATO interviewed 29 people who build, power, supply, and finance data centers and reviewed 261 publications.

A sales director at a switchgear manufacturer told LATO that newer manufacturers were winning about a quarter of the bids he saw by promising delivery in under a year. Established makers were quoting thirteen months. Some newcomers had venture or private equity backing. So far, you can see the investment thesis emerging.

But then you ask where they get their parts. According to the director, they all buy medium-voltage breakers from the same few suppliers, including Square D, Siemens, and ABB, with delivery taking 42 to 50 weeks. Those suppliers also compete for the finished orders and prioritize their own projects when parts are scarce, he said. He expects the newcomers to deliver three to six months later than promised.

If I were considering an investment here, I’d want that conversation before underwriting a company’s ability to beat the incumbents on delivery. Now I have a specific assumption to test with the founder and their suppliers.

Read "Ear to the Ground"

Power creates another wait. A construction manager in Northern Virginia described two 40 MW buildings whose equipment sat installed for almost four months while crews heated and dehumidified the rooms. The finished buildings you see in satellite pictures on social media can still be months away from serving a customer.

These delays come on top of schedules that already run into years. For a site of about 100 MW, six of nine experts whose answers could be compared estimated at least two years from first request to power. For 500 MW or more, three of seven estimated more than three years. One said five to seven. Think about how many generations of AI models and chips could come and go while you’re waiting to switch the site on…

Eleven of the 12 people who offered a view on capacity under construction expected most of it to get built. For projects due in 2027 and 2028 that had yet to start, four of the five forecasts below put cancellations at 20%. The fifth, from someone working on Texas grid connections, put them at 50%.

LATO’s Ear to the Ground. Nine forecasts from eight interviewees, grouped by whether construction had started.

One equipment lender described the squeeze on newer AI cloud providers. A year earlier, they could borrow the full purchase price of equipment and repay over four to five years. By November 2026, they could finance only 60% and had two to three years to repay. On a $100M order, that’s $40M to find elsewhere before a single GPU earns any revenue.

LATO’s Ear to the Ground. The credit chief at one equipment lender compared the terms his firm offered a year apart.

Why wouldn’t you interview everybody?

Each conversation an investor has with an expert offers new assumptions to test. But once agents can do the work, why stop at 10 interviews if 50 would provide even more signal? You can seek out people who disagree, ask follow-up questions, and revisit customers six months later. The conversations cost money, but they can run around the world while you sleep. You get to decide how much more you want to learn. The opportunity expands much further than just investors too.

If I ran a technology company, I’d want to talk to everyone experiencing the pain we’re trying to solve. How do they deal with it today? What makes their workaround good enough? I’d interview our customers, competitors’ customers, and buyers we’d lost before committing the next quarter’s product roadmap.

If I were a limited partner, I’d use LATO to understand the companies my managers keep telling me will drive their returns. Interview their customers and current or former employees. Find out why people buy, whether they’re spending more, and how the business has changed.

Subscribe now

A museum for the mind

I think AI can become a museum for the mind. Consider how museums safeguard the artifacts of civilization for future generations. By analogy, interviewing people lets us preserve their expertise and how they think. What did they notice that others missed? Why did they make a decision that looked wrong at the time? What changed their mind?

LATO can do this at a scale no professional interviewer could manage. Imagine interviewing the best people in hundreds of fields, asking them to work through difficult problems, and following up where they disagree. With their permission, we could use those conversations to teach an AI how these experts reason.

I think there’s a path from there to AGI. Learn from the best human expertise, then test those approaches in reinforcement learning environments where the model can act and learn from the consequences. Add open-ended exploration, where it discovers new problems to solve, and recursive self-improvement, where it improves how it learns. That’s the longer-term opportunity I see in what LATO is building.

The right team to rethink research

Tymek Staniszewski spent six years in growth and private equity. His co-founder Tien Chu, whom he befriended in high school in Warsaw, spent eight years in AI and software engineering. As a user and now investor, I find them sharp, kind, insanely ambitious, and a lot of fun to work with.

Since their first study at the end of July, LATO has run more than 1,000 expert interviews in over 40 countries for funds managing nearly $500B, and passed $1M in annualized run-rate revenue in under 100 days. And we’re just getting started!

If you make decisions that would benefit from this, or want to help build it, reach out at hi@latolabs.io.

Read "Ear to the Ground"

来源:Air Street Press · press.airstreet.com