Feb 20, 2025 · 46m · news
Adarsh Hiremath @ Mercor: The Fastest Growing Startup in Silicon Valley | E1261 · 20VC with Harry Stebbings
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
This episode of the 20VC podcast features an in-depth interview with Mercor co-founder Adarsh Hiremath, exploring the startup's rapid ascent to a $2 billion valuation and its high-intensity culture. Hiremath discusses the evolution of global labor, the critical role of elite human data in AI training, and Mercor's mission to fully automate talent sourcing.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 23.4% of the talking time here. How this is scored →
Harry as informed peer 2.6 Guest teaching 2.5 Guest disagreement 0.8 Harry pushing back 2.1
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 19:39 Defending human expert data over synthetic data
Adarsh explicitly rejects the premise that synthetic data replaces human post-training, arguing that model evaluations and domain expertise inherently require human intelligence.
Hardest push from Harry ▶ 19:39 Challenging data bottlenecks using Grok insider perspectivesHarry forcefully counters Adarsh's stance on human data by bringing up direct counter-arguments from Grok leadership about synthetic data quality.
Biggest teaching moment ▶ 21:47 Explaining the 70/30 human-AI collaboration splitAdarsh educates the host on how AI models handle initial bulk work while creating higher demand for specialized humans to complete the complex remaining percentage.
Harry holds his own ▶ 19:39 Demonstrating deep AI domain knowledge via industry quotesHarry displays his technical industry reach by citing specific technical leaders from Grok to challenge the guest's thesis on synthetic versus human data.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| 20VC Title Sequence | 1 | 2 | 0 | 0 | Harry introduces Adarsh and opens with background questions regarding his high school debate history with co-founder Surya. Adarsh politely explains how policy debate prepared them for startup partnership equity and decision making. | |
| Dropping Out of Harvard | 1 | 1 | 0 | 0 | Harry asks Adarsh about the decision-making process behind dropping out of Harvard. Adarsh shares a personal story about his nocturnal roommate and an emotional conversation in Palo Alto with his co-founder. | |
| Seed Round Success and $500 Salaries | 2 | 2 | 1 | 2 | Harry cites specific metrics like 50M ARR and asks about rumored 996 work hours. Adarsh mildly reframes 996 as an organic byproduct of passion rather than an explicit corporate requirement. | |
| Hyper-Growth Stress Tests | 2 | 3 | 1 | 1 | Harry inquires about culture stress tests under 50 percent monthly growth. Adarsh educates the host on how human data labeling has transformed into a high-stakes talent assessment challenge. | |
| AI Lab Partnerships and Business Metrics | 2 | 2 | 0 | 1 | Harry asks detailed operational questions regarding lab partnerships and success metrics. Adarsh outlines how Mercor's AI interviewer generalizes across different professions. | |
| Model Infrastructure and OpenAI Integration | 3 | 2 | 0 | 1 | Harry asks about underlying model provider choices and model ecosystem architecture. Adarsh details the market shift toward reinforcement learning paradigms. | |
| Human Data as the AI Bottleneck | 5 | 4 | 3 | 5 | Harry challenges Adarsh by citing Grok's stance that synthetic data surpasses low-quality human data. Adarsh defends his view, reframing high-quality human data as an essential talent matching problem. | |
| The Long Tail of Tasks and Human-AI Collaboration | 4 | 4 | 3 | 5 | Harry pushes back directly on why human labor will remain necessary as model capabilities reach 80 to 90 percent. Adarsh reframes the relationship into human-AI collaboration across the long tail of tasks. | |
| Operating with Zero Sales Representatives | 3 | 3 | 2 | 4 | Harry presses repeatedly to get Adarsh to reveal Mercor's take rate percentages. Adarsh initially focuses on talent quality value before clarifying the variable pricing structure. | |
| Recruiting Roots in India and Sourcing US Talent | 2 | 3 | 0 | 1 | Harry asks about geographic candidate distribution and whether young people should study computer science. Adarsh explains that programming is moving to higher levels of abstraction. | |
| AI Coding Tools and Software Commoditization | 3 | 3 | 0 | 2 | Harry asks about coding tools like Cursor and the impact of software commoditization. Adarsh explains why network effects become the primary moat when software costs approach zero. | |
| Mercor's Two Moats: Market Liquidity and Data Flywheels | 2 | 2 | 0 | 1 | Harry prompts Adarsh to identify Mercor's core competitive moats. Adarsh outlines two distinct flywheels: marketplace liquidity and outcome prediction data. | |
| The Necessity of In-Person Energy | 3 | 2 | 0 | 1 | Harry discusses in-person work intensity and asks about past product mistakes. Adarsh candidly admits that over-indexing on chat-only user interfaces was a mistimed strategy. | |
| Felicis Partners and Scaling to Eight Figures | 3 | 2 | 0 | 3 | Harry asks detailed questions about investment timelines and ARR scale at the time of fundraising. Adarsh confirms scaling to eight-figure revenues before accepting Felicis capital. | |
| The $100 Million Fundraising Round | 4 | 3 | 2 | 5 | Harry directly challenges why a highly profitable startup needs to raise a 100M round. Adarsh explains that labor market aggregation requires a massive balance sheet long term. | |
| Quick-Fire: Sourcing Efficiency and Matching | 3 | 3 | 1 | 3 | In quick-fire mode, Harry asks if pushing for lean headcount contradicts Mercor's recruiting model. Adarsh clarifies that lean teams require even higher matching precision. | |
| Quick-Fire: Hard Lessons and Sam Altman | 2 | 2 | 0 | 1 | Harry wraps up with quick-fire questions on hard lessons, Sam Altman, and 10-year company scale. Adarsh lays out a grand vision of managing billions of jobs globally. |
Statements from this episode (41)
Hiremath: Mercor raised $100M at a $2B valuation
“The round is a hundred million. It was at two billion.”
Hiremath: Future AI ecosystem will feature many specialized models
“I think we'll live in a world with many, many models with different use cases.”
Hiremath: Recruiter is the most prestigious position in any company
“I think being a recruiter is the highest prestige position in any company. The recruiter is the one who controls the talent inflows and outflows of every company.”
Hiremath: Zero software cost businesses must rely on network effects
“The businesses that succeed in a world where software costs approach zero will be built on network effects.”
Hiremath: Mercor originally launched as a software development shop
“Brendon Suria and I started working together without any business ambition. Necessarily. We just started a dev shop together.”
Hiremath: Mercor pivoted to recruitment after hiring Indian developers
“And what we ended up doing is recruiting these really, really exceptional folks from India to help us out with our dev shop. And then very, very quickly we realized, you know, the software was one thing, but we had found some really, really exceptional people …”
Adarsh Hiremath: Major founder decisions like dropping out are made emotionally
“Like most of these decisions, you just make them completely emotionally, and I just knew I wanted to work with my best friends.”
Hiremath: Mercor founders dropped out before raising seed or getting Thiel Fellowships
“No seed round, a little bit of revenue, no series A, no Teal Fellowship, nothing. We were just three friends working in a small office in Palo Alto with our amazing team in India.”
Mercor Founders Paid Themselves $500 per Month Before Seed Funding
“We changed our like salaries and gusto to 500 dollars a month.”
Hiremath: Mercor Promoted 996 Schedule to Protect Sundays from Work
“The only reason we actually just floated those numbers out is because we didn't want our team working on Sundays.”
Hiremath: Historically All Successful Companies Had Intense Work Cultures
“It's like all of the successful companies have had pretty intense cultures historically, and it's just a function of a startup, right? You got to work harder than everyone else, obviously in a sustainable way to succeed.”
Hiremath: You Can Teach Technical Skills, But You Can't Teach Care
“One realization that we've come to is that you can teach people a lot of things, whether it be, you know, technically or going to market or whatever, but The one thing that you can't quite teach people is to care, and that's one thing that we index on pretty h…”
Hiremath: Human data labeling and talent assessment are now the same problem
“Actually, our insight about the market is that human data and talent assessment have actually become the same thing, right?”
Hiremath: Top AI labs use Mercor to hire talent for model post-training
“It looks exactly the same as placing someone to work at any company, right? So just like Mercore might work with startups making their first hires or companies hiring in a more traditional full-time capacity. It's the exact same thing for a lot of the large AI…”
Hiremath: Mercor's AI Interviewer Can Launch in Under 10 Seconds
“You can literally spin up this interview in under 10 seconds.”
Hiremath: Shift to reinforcement learning will create domain-specific reasoning AI
“The whole market is shifting to reinforcement learning, right? You're already seeing this with O-one, O-three, the deep seek models. And as a result, I think we're going to see really, really powerful models in specific domains that can reason extremely well.”
Hiremath: Mercor aims to significantly outperform expert human hiring managers
“For us, it's hiring and Beating the expert hiring manager by a large margin.”
Hiremath: Only a few companies will build foundational AI models
“There will only be a couple of companies that are able to build these foundation models that everyone builds off of. I think OpenAI is a great example of one of those companies. And I think, you know, that analogy roughly holds. I don't anticipate there being …”
Hiremath: AI model evaluation definitionally requires human-created datasets
“I think a lot of it will be human data going forward, and I think a great example of this is evals, right? Evals for models definitionally have to be outside of model capability, right? In order to see whether model is doing well at a particular task, You need…”
Hiremath: Expert human data will be the primary bottleneck for AI
“So synthetic data will certainly be a part of the equation, but in a lot of ways, the bottleneck to unlocking and unleashing the next level of intelligence will be expert humans, which brings me back to the question and phrase that you used, low quality human …”
Hiremath: AI will only finish 60-80% of tasks, increasing human value
“But I think the more realistic breakdown is the AI for a specific use case might be able to get us 6070, 80% of the way there. But for that remaining 40, 30, 20%, you're going to need a human To be able to take you all the way there. And the reality of the sit…”
Hiremath: Future labor markets will shift toward extreme specialty and sophistication
“I think the key thing is that the market will move towards specialty and sophistication. Meaning the types of work that we see 50 years from now will be more specialized and oftentimes will require people with kind of like a higher level of sophistication in t…”
Hiremath: Mercor has no sales team outside the co-founders
“One interesting thing about Mercore is we don't have a sales team. There is not a single person who works on sales at Mercore outside of, you know, the founders.”
Hiremath: Mercor's entire recruitment and payout pipeline is completely automated
“So on our end, the entire process is automated. So this is everything from a candidate hearing about Mercore and going onto the Mercore platform via job listing, us pulling in their resume, their salary expectations and whatnot. Administering a personalized in…”
Hiremath: Top 0.1% talent delivers exponential value, making customers price-insensitive
“There's a huge difference between the top .1% and then the 80th percentile person. So usually for customers, it's not a question of price. It's a question of quality.”
Hiremath: Mercor's client take rates exceed 30% for some accounts
“It's on a case by case basis. For some customers it can be, you know, over 30%. For some it can be less.”
Hiremath: The US is the top country of origin for Mercor candidates
“Actually the number one place that, you know, workers on the Mercore platform who, you know, have jobs through, through us are from is actually the United States.”
Hiremath: Learning Programming Is Still Essential, Shifting to Natural Language Abstraction
“My take is that programming is actually more important today And it's just going to happen at a different level of abstraction. One could argue that the leap from assembly to Python was actually maybe even a bigger leap than the leap from Python to natural lan…”
Hiremath: Software will be commoditized quickly as coding agents improve
“And I think like the implication for software is that software is going to get commoditized very, very quickly as these coding agents get really, really good.”
Hiremath: Next era of SaaS will replace entire manual services end-to-end
“I think what we consider SaaS will change, right? In the sense that the next era of SaaS will be replacing entire services, right? You know, whether it's the end-to-end process of a recruiting agency like Mercor or another, you know, different service or that,…”
Mercor Builds Predictive Candidate Flywheel from End-to-End Job Performance Data
“The second network effect is or like data flywheel is around this job prediction piece where we're able to see who is performing well on jobs and the specific reasons why they're performing well on jobs and use that, that end to end Data on people's outcomes t…”
Hiremath: The greatest technology companies of this generation are usage-based
“And I think a lot of the greatest products or companies of our generation have been usage based, right? I think Stripe is a great example of this.”
Hiremath: Mercor is moving toward fully automating its internal hiring process
“Maybe running our entire, you know, internal hiring process for Mercore in a completely automated way, meaning Brandon, Surya, and I don't even talk to someone when they come into the office, and then we walk in the conference room to meet them the first time,…”
Stebbings: Intensive 996 work cultures require in-person collaboration
“I don't think you can do nine nine six and do it effectively, unless you're in person. I think that motivation, that intensity, you feel in the same room.”
Mercor initially launched exclusively as a chatbot interface
“One of the initial iterations of the Mercore product was just built around a chat interface. Like there was pretty much no other way to hire people unless you use the Mercore chat bot, because we were so bullish on, on chat.”
Hiremath: Mercor Did Not Intend to Do Any Stated Fundraising Rounds
“An interesting dynamic about of all of our, you know, fundraising rounds is just that, like, we didn't intend on, on doing the fundraising at the time, and it sort of just came to us.”
Hiremath: Mercor reached eight-figure revenue before raising from Felicis
“It was doing, you know, eight figures in, in revenue, right?”
Hiremath: Mercor's board consists only of its co-founders and Benchmark
“It's, you know, Brendan, Suri, and I, and Benchmark on the board.”
Hiremath: Mercor does not plan to immediately deploy $100M raise
“Our goal isn't to deploy a hundred million dollars tomorrow.”
Adarsh Hiremath defines AGI as automating economically valuable work and research
“When we achieve AGI or sort of like think about AGI, it will certainly involve You know, doing more economically valuable work, right? So when more and more and more of economically valuable work has been automated to some extent, you know, research has been a…”