Moat

Foody claims data quality is power-law, not additive: on a 100-person project most of the model improvement comes from the top 10 to 20% of contributors, mirroring how value concentrates inside any company. The defensibility therefore lives in two compounding assets, the referral network that sources rare 10x experts and the matching layer that routes them to tasks where they shine, which he argues is very hard to copy. It is his answer to the charge that every labor marketplace now says the same things in podcasts.

> "The outcomes of data and the people that contribute to it are extremely power law. Similar to a company, if you have 100 people on a project, oftentimes the majority of the model improvement is coming from the top 10 to 20% of people." > — [[12:30]](https://www.youtube.com/watch?v=FzftvxA84z8&t=750s) > "When we're able to build proprietary advantages in the way that we have, not only our supply base and the referral network to access them, but also the way that we match those experts with the opportunities where they're going to do phenomenal work, it creates so much value for customers that it's extremely difficult to compete against." > — [[13:00]](https://www.youtube.com/watch?v=FzftvxA84z8&t=780s)

For judging which AI revenue is real, Foody points to retention and customer love rather than pilot counts: a company where 95% of pilots fail is suspect, while unparalleled retention plus customers who rave is the signal of true fit. He pairs this with Sam Altman's durability test, whether models being far better in one to two years would help or hurt your business, as the single most important question a founder can ask. Together they form a filter for separating durable demand from subsidized or pilot-driven growth.

> "The most important thing is looking at the numbers and anecdotes around retention to see the revenue health and whether there's real value. If you meet an application-layer company where 95% of their pilots are failing, it's probably not going to be a good investment. But if you meet a business that has extraordinary, unparalleled retention numbers, and you talk to those customers and hear how much they love the product, then of course it's a really exciting opportunity." > — [[36:00]](https://www.youtube.com/watch?v=FzftvxA84z8&t=2160s) > "I really like the thing Sam Altman says of, will models being dramatically better in one to two years improve your business or worsen it. That is in so many ways the most important question to see if you're building a business that's durable." > — [[56:00]](https://www.youtube.com/watch?v=FzftvxA84z8&t=3360s)

Foody's closing thesis is that RL environments, where humans build the frameworks and verifiers and models learn to execute, will subsume the entire economy because it makes little sense for humans to keep doing monotonous, redundant knowledge work. He frames the human role as constructing the scaffolding rather than doing the task, and pegs Mercor at roughly 50 to 60% of the RL-environments market, the fastest-growing new data type. This connects the human-versus-model TAM argument to a concrete market-share claim worth verifying later.

> "There's the new data types that everyone's moving towards, called RL environments, where we're, call it a rough estimate, like 50 to 60% of the market." > — [[59:00]](https://www.youtube.com/watch?v=FzftvxA84z8&t=3540s) > "I've talked to multiple executives, CEOs at leading labs, that believe RL environments will subsume the entire economy, because it doesn't make sense that humans would be doing monotonous, redundant work. It makes way more sense for humans to build the framework of how to do that so that models can then learn how to do it and do it for us." > — [[59:30]](https://www.youtube.com/watch?v=FzftvxA84z8&t=3570s)

Foody reframes Mercor not as a body shop but as the company that caught a market transition: early LLMs were trained on cheap, low-skill crowdsourced data from Scale and Surge, but frontier post-training now needs Goldman and McKinsey caliber experts who can build data that researchers cannot even interpret on their own. He locates Mercor's takeoff in that engagement-model shift, sourcing and vetting elite talent that works directly with researchers, rather than in headcount or facilitation. It is the founding thesis the rest of the interview builds on.

> "There was this enormous transition underway, moving away from the crowdsourcing paradigm that Scale and Surge pioneered, of how do you get low- and medium-skilled people that write barely grammatically correct sentences for early LLMs, and very quickly moving towards this sourcing-and-vetting paradigm of how do you find the Goldman, the McKinsey analysts, the FANG software engineers, the top doctors and lawyers that can work directly with researchers to help them build the highest-complexity data on Earth." > — [[08:00]](https://www.youtube.com/watch?v=FzftvxA84z8&t=480s) > "It's really that trend around a different engagement model and higher-caliber work that caused us to take off and really catalyze this meteoric growth." > — [[09:00]](https://www.youtube.com/watch?v=FzftvxA84z8&t=540s)

Yao observes that as of early 2026 no scenario has yet formed a genuine data flywheel, and outside of agentic coding there is no purely AI-native application that has become highly successful. Durable advantage therefore still sits mostly on the model side rather than in the application layer.

> "目前没有哪一个场景真正形成了数据飞轮。甚至AI纯粹原生的应用场景，目前除了Agentic coding，就是写代码之外，没有哪个场景是AI真正原生的场景，变得非常成功。" > — Whisper transcript [00:16:40] ("没有哪一个场景真正的形成了数据飞轮，甚至AI纯粹原生……"), chapter 竞争与逃逸; research "飞轮" via coding revisited at [02:11:43]. Quote verified against 虎嗅 digest #20. > Gloss: No real data flywheel exists yet; apart from agentic coding, no AI-native app has truly succeeded, so moats stay on the model side.

A thin app built on someone else's model (a 壳) can survive in only two ways Yao can imagine: outrun the labs by growing fast enough to capture user mindshare and evolve its own model before they copy the product form, or operate in a market so small the labs cannot be bothered to enter it. Anything in between gets absorbed.

> "壳（AI产品）在目前这个情况下活下来，有两种我大概能想象的方式。一种是逃得足够快……以至于在模型公司反应过来的时候，我其实已经占领了大量的用户心智……另一种方式就是这市场足够小，小到模型公司根本懒得去管。" > — Whisper transcript [00:18:00]–[00:18:29] ("壳……活下来有两种……占领了大量的用户心智……我又自己研发出了自己的模型"), chapter 竞争与逃逸; he names Cursor as an example of the "outrun" path. Quote verified against 虎嗅 digest #21. > Gloss: A wrapper survives only by outrunning the labs to mindshare or by being too small to be worth their attention.