Vlad Tenev - Navigating Robinhood's Evolution - [Invest Like the Best, EP.384]
What this covers
The conversation moves between two subjects: the frontier of AI and the history of Robinhood. On AI, Vlad Tenev says the big labs assume reasoning will emerge from larger models trained on more English text. He thinks that is an incomplete path. He points to Lean, whose fully released Lean 4 version moves mathematics from chalkboards into VS Code and GitHub, so proofs become code that people can formalize and debug. A mathematical superintelligence could then write software and prove it bug-free to its specification, which matters most for mission-critical algorithms. He also argues that models have hit a data wall and that synthetic data from the previous model generation will add little, since otherwise models would be a perpetual motion machine.
The Robinhood half starts with the founders' failed high-frequency trading firm and the software they then sold to banks. Exchanges were competing on latency, with a 40-microsecond matching engine at NASDAQ, and that pushed per-trade costs toward zero. Tenev also covers the revenue-versus-satisfaction inverse he called a five-alarm fire, the shift to net interest income as rates rose, and the 3% cashback credit card, which he compares with Fidelity's.
Later sections cover the hardest stretches. GameStop had a clear playbook, while the 2022 rate reversal eroded the business slowly. Tenev recounts using a New CEO Thought Experiment built around Frank Slootman, and says trust took roughly three years to rebuild. The more contested turns come near the end: inverted advisory fees, why robo-advisors stalled, stablecoin use abroad, and a fringe forecast that token offerings could displace traditional IPOs.
Tenev argues that treating math as reasoning, and pairing formal verification tools like Lean with modern AI, will produce mathematical superintelligence that makes software provably correct, while Robinhood's success came from applying deep engineering and aesthetic craft to a mobile product that made trading free and most AI value lies in digitizing paid human financial advice.
- Mathematics is reasoning, so a system that solves math at superhuman levels is the path to general reasoning and provably correct software
- Robinhood's free trading came from copying institutional trading infrastructure for retail, not from a consumer-first insight
- AI will create the most value by automating expensive human advisory services, not robo-allocation
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Robinhood's founders began as a high-frequency trading firm after grad school, which failed commercially, then pivoted to selling trading software to other HFT firms and big banks, giving them deep technical knowledge of market plumbing.
“this high-frequency trading business that we tried to create, which wasn't very successful, led to us creating a software business where we served other high-frequency traders and big banks”
Exchanges competed on latency, with NASDAQ offering 40-microsecond matching engine latency, turning trade execution into a pure computer science problem and driving trading costs toward zero.
“NASDAQ came out and said, we can offer you matching engine latencies of 40 microseconds. This became purely a computer science problem”
Leaders resist reversing decisions they publicly announced, which is among the most dangerous organizational traps, whereas an outsider can reverse them easily by blaming the prior leadership.
“it's very, very difficult to go up in front of the company and say, Hey, this thing that I just told you that we were going to do three months ago, I was wrong”
The Lean proof language, fully released as Lean 4, moves mathematics from chalkboards into VS Code and GitHub, turning proofs into code that mathematicians formalize, collaborate on, and debug.
“What Lean has done is it's taken math out of the realm of chalkboard and chalk and put it into the realm of VS Code and GitHub”
Outside the U.S., crypto transaction volume is rising mainly because people in weak-currency markets use dollar stablecoins, such as USDT on TRON, to hold dollars cheaply, which has surpassed traditional financial infrastructure for that use.
“what's driving that increase is largely dollar, stablecoin volume on low-cost networks”
Advisory fees are inverted: small investors pay 2% to 3% of assets for mediocre service, while wealthy clients pay a fixed percentage that yields roughly ten times more fees for essentially the same service, so satisfaction falls as wealth rises.
“the more money you make, the more dissatisfied you are with your financial advisor relationship, and it really should be the opposite”
Robo-advisors failed to gain share because automating asset allocation is a commodity; the valuable part of advice is holistic planning such as estate, tax, budgeting, and private banking, which human advisors spend little time on portfolio work to deliver.
“automating the actual asset allocation is the easy part, and I think that's a commodity service”
The dominant AI labs assume reasoning will emerge automatically from bigger models trained on more English text; Tenev says that may partly work but he sees it as an incomplete path.
“reasoning capabilities, which includes mathematical reasoning is going to be an emergent phenomenon, an emergent property of bigger and bigger models”
Robinhood's experience running both a crypto and a traditional brokerage shows that operating costs are dramatically lower on blockchain-based infrastructure, and Tenev believes the migration off mainframes is inevitable.
“The cost of offering those services is dramatically lower on the crypto side”
Models have already trained on nearly all public internet data, creating a data wall; synthetic data generated by the previous model generation will probably be much less valuable than human data, because otherwise it would be a perpetual motion machine for capability gains.
“My bet is not significantly. It will probably be much less valuable than the human data because otherwise, you would have some kind of perpetual motion machine.”
Startups that issue liquid tokens to employees at formation may be preferred over traditional equity that vests over four years and is illiquid until IPO, which could shift companies away from traditional public listings toward token offerings.
“companies will do ICOs instead, and that could fundamentally change the equity markets”
Tenev bets that synthetic data generated by the previous model generation will not significantly increase capability, because otherwise models would be a perpetual motion machine that learns from its own output.
“My bet is not significantly. It will probably be much less valuable than the human data because otherwise, you would have some kind of perpetual motion machine”
Net interest margin has overtaken transaction revenue as Robinhood's largest revenue contributor, because in the zero-rate era payment for order flow dominated, but higher rates have shifted the mix toward cash and asset-building products.
“interest revenue actually flipped transactions. It used to be very, very heavy on the transaction side”
Robinhood's free trading came from Tenev and co-founder Baiju's background building trading software for high-frequency firms, where latency fell to microseconds and per-trade costs approached zero, so retail could replicate that stack.
“the cost of the trades basically was zero”
A useful exercise in a turnaround is to imagine a very good new CEO arriving and listing what they would change, because it surfaces decisions the incumbent team is clinging to for consistency.
“sit down and imagine what the company would look like if a new CEO came in to run it”
The AI applications most certain to work are tasks where companies or customers already pay humans a decent amount, such as advisory and support, because they can be automated and reduce cost, unlike speculative AI applications.
“the things that I know will work are things that you're paying a human a decent amount of money for already that could be done by AI”
Imagining how a strong new CEO like Frank Slootman would fix the company surfaced obvious problems the team had been avoiding, because leaders cling to past decisions they have publicly committed to, and an outsider can reverse them easily by blaming the predecessor.
“imagine what the company would look like if a new CEO came in to run it and a new CEO who's very, very good.”
GameStop was a difficult but acute crisis with a clear playbook of serving customers, raising capital, and strengthening the balance sheet, whereas the 2022 interest-rate reversal was harder because it slowly eroded a trading-dependent business built on 2021 assumptions.
“I think what was more difficult for me is navigating 2022”
Reputations can be lost very quickly, NPS fell to the bottom after GameStop and recovered gradually over about three years, so time is the biggest factor in rebuilding trust, with good products only modestly speeding it up.
“reputations can be lost very, very quickly. We see this in our data”
A mathematical superintelligence could write software and prove it correct and bug-free to its specifications, a capability that is most valuable for algorithms in mission-critical software.
“write software really well and ensure that, that software is provably correct so that there's no bugs and it actually works to its specifications”
Robinhood's 3% cashback on all categories was designed to disrupt credit card economics the way free trading disrupted brokerage, and it is priced to be profitable through interest on revolving balances and secondary account effects.
“that's where we got the 3% cashback on all categories, value prop”
At Robinhood, the more active a trader was, the more revenue they generated and the less satisfied they were, an inverse relationship between revenue and NPS that Tenev treated as a five-alarm fire.
“the more active a trader you were on Robinhood, the more revenue you're generating for us, the less satisfied you were with our product”
Pattern day trader restrictions blocked active options traders, so building an off-ramp into cash accounts let them keep trading on Robinhood, which improved satisfaction and retention among the most active users.
“And the solve for that was building cash accounts, which is a different account type, non-margin”
Robo-advisors were marketed as the original AI in investing but did little sophisticated AI, and the real opportunity is to deliver the experience of a private wealth manager digitally at lower cost to many more people.
“I think what people really want is the experience of a private wealth manager or a private banker delivered to them digitally”
Robinhood wants a small number of elite, well-paid people working in person rather than large numbers of low-cost staff, believing small elite groups push each other to do better than individuals and that offshoring creates mediocre output, and the speaker says the company is only about 25% of the way toward this culture.
“I would rather have a small number of just elite people that I pay more”
Fidelity offers a 3% rewards card to clients with about $2 million under advisory management, which signals that a primary card relationship is accretive through deposits and the rest of the financial relationship, not charity.
“I don't think they're doing this out of the goodness of their hearts”
The most valuable AI application in finance is delivering the experience of a private wealth manager or banker digitally at low cost, so that far more people can access holistic advice, which the speaker expects generative AI to transform within a few years.
“I think what people really want is the experience of a private wealth manager or a private banker delivered to them digitally”
Tenev prefers a small number of elite, highly compensated people over the offshoring model of large numbers of low-cost mediocre staff, because elite groups push each other to produce results greater than individuals could.
“my preference would be a small number of amazing people that are elite in their fields that we compensate very, very well”
Robinhood's founders found no fundamental reason retail traders could not get the stack of a sophisticated HFT firm, suggesting the exclusion was a product gap rather than a regulatory or technical necessity.
“there was no fundamental reason. There was nothing that said you can actually like replicate the stack of a sophisticated high-frequency trading firm and combine it with the retail platform.”
Tenev's team concluded that retail investors had no fundamental reason to be denied institutional-grade trading technology, only inertia and assumptions, which led them to build a consumer version.
“there was no fundamental reason. There was nothing that said you can actually like replicate the stack of a sophisticated high-frequency trading firm and combine it with the retail platform”
Mathematicians judge proofs by elegance and beauty the way art enthusiasts judge paintings, so the same aesthetic sensibility is valid for product design, and Robinhood is motivated by aesthetics and not only performance.
“there's a huge overlap between math and art. When you talk to mathematicians and they're talking about a proof to you”
Small product improvements were not prioritized because only a handful of engineers worked on options, a multi-hundred-million-dollar business, and the team was nomadic, starting projects that would not pay off for two to three years.
“little paper-cut improvements, even significant ones were just not getting prioritized when you have a handful of people working on options”
In financial services, incumbents benefit from decades of customer confidence, but a challenger's trust gap narrows over time, so tenure becomes a tailwind for Robinhood if it avoids missteps.
“time actually helps us in financial services”
Tenev wants promotions and hiring to reward doing big things with as few people as possible rather than leading large teams, because scope-based hiring incentivizes empire builders.
“what I instead want is someone that's done big things with as few people as possible”
Software is verified by hand today, which is expensive and artisanal, and firms do it only where a mistake would cause enormous financial damage; AI will eventually automate this.
“So it's kind of like an artisanal thing that's very expensive, and it's only basically done where there's already a huge financial impact for doing things wrong”
Harmonic combines formal mathematics in Lean, which digitizes math into code, with AI advances like AlphaZero and GPT-4 that take first steps into reasoning, to build a system that reasons with verifiable correctness.
“combining these two things, lean and formal mathematics that digitizes math and turns it into a programming language and all the advances in AI popularized by AlphaZero and GPT-4”
The GameStop episode was an acute crisis handled by a defined playbook of serving customers, raising capital, and strengthening the balance sheet, and the speaker admits the public messaging was not handled well, framing it as a meta problem rather than a product problem.
“I don't think we did a particularly great job of that. It's like a meta problem rather than an actual problem with the product or the business.”
Robinhood's NPS dropped to its lowest point immediately after GameStop and has recovered gradually over about three years, returning to roughly pre-GameStop levels, with the largest gains happening right after the crisis.
“It took a dive pretty much immediately. In our NPS survey that we ran right after the GameStop events, it was at the bottom”
Rapid rate hikes in 2022 collapsed trading volume, exposing that Robinhood had built heavy headcount and long-term projects in 2021 on the assumption that trading growth would continue indefinitely, forcing headcount cuts and a focus on the core.
“We were acting in 2021 like those trends would continue indefinitely”
Because mathematics is reasoning, building a mathematical superintelligence means building something that can take logical steps, make deductions, detect its own wrong deductions, and course-correct at a level beyond human capability.
“My belief is that mathematics is reasoning, and so when you talk about building a mathematical superintelligence, that's another way of saying something that can reason”
If provably correct software became cheap, regulators would require it for properties such as preventing double spending and enforcing risk checks; the only reason it is not required now is that it is impractically costly.
“I think that we'll get to a situation where the regulators will actually require that”
Robinhood's 3% card reward exceeds the 2% to 2.5% interchange rebate it receives from networks, so the company covers the difference with interest on carried balances and underwrites the card near zero profit to be conservative.
“So we're actually paying a greater reward than we get from the network, and we compensate for that in two ways”
The litmus test for mathematical superintelligence is a Millennium Prize problem; Tenev believes AI will solve the next one, and once it does, all future difficult math problems will likely be solved by AI.
“the litmus test for this will be Millennium Prize problems”
Robinhood chose mobile as its focus because the phone is always with active traders, and because the company had no fundamental advantage on the web where incumbents were established.
“this device is with you at all times. And if you're an active trader and you really care about what's going on in the markets, that's just the capability that you can't get on desktop”
By going mobile first, Robinhood redefined the market so that incumbents rushed to adapt their interfaces to look like Robinhood, because customers came to expect that design.
“the incumbents that had to rush in to make sure they didn't lose their existing market had to adapt their interfaces to look exactly like Robinhood too”
Capital One's card revenue divided by transaction volume is about 6%, and it pays out roughly 2% in rewards, which Tenev uses as the benchmark for the profit room available to Robinhood.
“you'll get to something like 6%, which basically means there's like a 6% revenue yield on transactions”
Preserving a path back to a previous role lowers the psychological cost of a risky venture and makes people more willing to take it, as the speaker experienced when a department head offered a welcome-back guarantee.
“it sort of empowered me to feel comfortable leaving, just knowing that I could come back”
Verified AI-written software will first impact safety-critical domains such as financial services, blockchain smart contracts, automotive, and aerospace, because the cost of bugs there is highest.
“particularly in safety-critical domains like financial services, blockchain smart contracts”
Once a customer has Robinhood as a primary credit card, they are expected to deposit more money into Gold, feeding the Gold flywheel that has been working effectively.
“once someone has a primary card relationship with Robinhood, they'll invest more money, they'll deposit more money in our gold offering”
The GameStop crisis drove user growth and a large capital raise, which left Robinhood with about $6 billion of cash that let it turn around the core business quickly and go public just before the IPO window closed.
“we had $6 billion of cash in the bank, which can also solve a lot of problems”
Options trading generates an annualized revenue range of about $500 million to $1 billion, making it the largest transaction revenue line, with crypto around $150 million quarterly and equities around $100 million quarterly.
“Options trading is annualized in the $500 million to $1 billion revenue range”
Entrepreneurs usually underestimate what they can accomplish in a year, and most problems can be fixed given a long enough horizon, even if they seem daunting now.
“it really is true that you tend to underestimate what you can do in a year”
Customer support is a market of roughly $500 billion annually in call center spending that AI will reallocate within about ten years, changing the jobs of those workers.
“something on the order of $500 billion spent annually on call center, and that's going to be reallocated in the next 10 years”
If advanced civilizations had billions of years of head start, they should have expanded through the galaxy, and the absence of visible colonization is the Fermi paradox.
“where are all of the advanced civilizations? Why can't we see any?”
The short-term crypto narrative is damaged by low-quality tokens created easily on blockchains, which distracts from the underlying potential of the technology.
“people see all of the mean coin activity and lots and lots of things that don't have any fundamental backing being created”
The speaker is driven by the question of why the laws of physics take their particular form and what preceded the Big Bang, and believes a simpler theory than string theory likely explains them.
“I don't think string theory provides a very compelling answer, I think there's probably something simpler.”
Regulation and entrenched special interests will push token-based corporate structures outside the U.S. first, but the U.S. will eventually catch up as crypto infrastructure embeds into traditional projects.
“there's regulation that doesn't want this to happen and a lot of embedded special interest, so it looks like that might happen outside of the U.S. first”
Leonardo de Moura joined AWS because high-scale operations make cryptographic or encryption bugs potentially catastrophic, so AWS invests in manual verification now, work Vlad expects AI to eventually automate.
“if there's something wrong with the encryption library or their cryptography, the results could be catastrophic and incinerate tens to hundreds of billions of value”
Beyond software, mathematical superintelligence could make discoveries in engineering and physics and run experiments, and it would be most valuable for work needing correctness across chains of logical thought, though it adds little over existing tools for writing essays or poems.
“It can make new discoveries in engineering and physics. It can run experiments.”
A Department Vice Chair's encouraging response to a leave-of-absence request helped Tenev leave his PhD to start a business, and he concludes that people should ask for what they need because they will be surprised by the response.
“Yes, I think sometimes you have to ask for stuff, and you'll be surprised what happens”
Close to $100 trillion in wealth will pass from baby boomers over the next one to two decades, and Robinhood intends to capture a large share by serving millennial heirs.
“close to $100 trillion in wealth passed down from baby boomers”
A faulty smart contract led to the Wormhole bridge exploit, which drained roughly $400 million, showing the financial cost of unverified blockchain code.
“the Wormhole exploit that bled out, I think, $400 million from this Wormhole bridge a couple of years ago”
Frontier models such as GPT-4 and Claude 3.5 were trained on nearly all public internet data, creating a data wall where further progress requires new data sources.
“we basically trained on nearly all the public data that's on the Internet with the latest generation of models being GPT-4 and Claude 3.5”
Robinhood is fundamentally an engineering company in financial services, whose end product is high-criticality software, and small bugs have cost the firm tens of millions of dollars.
“Robinhood is fundamentally an engineering company in financial services. We produce a lot of software”
Engaged Robinhood users almost always become Gold subscribers, and Gold creates an incentive to adopt other products that offer industry-leading economics.
“Once you get sufficiently engaged in Robinhood and you find value, you almost always end up being a Gold subscriber”
Younger, more technologically savvy mathematicians such as Terry Tao and Ravi Vakil are formalizing papers in Lean, using it to find errors, and running distributed collaboration, while older mathematicians resist.
“the older mathematicians are kind of resisting this to a certain extent”
The Robinhood Gold Card was launched with a 1980s luxury aesthetic inspired by print magazine ads, with a physical card, app experience, and acquired company X1's features surfaced to deliver craftsmanship.
“we looked at print magazine adverts as inspiration”
AWS invests heavily in manual verification of its encryption and cryptography libraries because a flaw could incinerate tens to hundreds of billions of dollars of value.
“if there's something wrong with the encryption library or their cryptography, the results could be catastrophic and incinerate tens to hundreds of billions of value”
Mathematical superintelligence is an artificial intelligence that can solve math problems at a superior capability to the sum total of all humans on earth, which most people misunderstand as a fancy calculator rather than a reasoning agent.
“Mathematical superintelligence is basically an artificial intelligence that can solve math problems at a superior capability to the sum total of all humans on earth”
Robinhood acquired Pluto, an AI-based investment research product that synthesizes public filings to answer questions, and plans to integrate it into self-directed and advisory offerings.
“we just recently, announced an acquisition of Pluto”
Current AI models make obvious reasoning mistakes and very confidently say wrong things, particularly in math and language-based logical deduction, which is why reasoning is the next frontier.
“they just make obvious mistakes and they very, very confidently say wrong things”
A Metaculus prediction market gave about a 43% chance that the next Millennium Prize problem will be solved by AI, and Tenev thinks the true probability is higher.
“last I checked, it was a 43% chance that the next Millennium Prize problem will be solved by AI”
Robinhood has more millennial customers than Schwab, Fidelity, Vanguard, and E*TRADE combined, which Tenev uses to justify capturing the wealth transfer from baby boomers.
“Robinhood has more millennial customers than all the major brokers combined, more than Schwab, Fidelity, Vanguard, E*TRADE put together”
Robinhood's business is organized around three priorities: being the number one venue for active trading, growing wallet share with millennials, and expanding internationally.
“the business is focused on three things. The first one is being #1 in active trading”
Tenev favors in-person work and does not believe the future is fully remote and distributed while humans are involved, though he stays remote tolerant.
“I don't think the future is fully remote and distributed companies, at least while humans are involved”
Early observers doubted the founders could build a consumer product because they were mathematicians and physicists without consumer DNA, a skepticism the founders countered by focusing on beautiful design alongside their technical strengths.
“They're not going to know the first thing about building consumer products.”
Crypto and AI are the two major technology trends for financial services, and Robinhood intends to lead both, already leading in crypto and investing heavily to lead in AI, having acquired Pluto for AI-based investment research.
“there's two big technology trends that could affect financial services. One is crypto, the other is AI. And Robinhood will be a leader in both.”
Robinhood Gold is a nine-figure revenue business with 1.7 million subscribers, mostly paying $5 a month, and it works as the product form of growing wallet share by unlocking better economics across products.
“Gold subscriptions are now a nine-figure revenue business as well”
Very few mathematics department members start businesses, and an encouraging departmental response shapes whether talented people pursue entrepreneurship rather than treating math as a dead-end career.
“Very, very few people actually from the mathematics department start businesses”
Robinhood still has too much bureaucracy and process, and Tenev estimates the company is only 25 to 30 percent of the way to the culture he wants.
“We're maybe 25% of the way to building the company that I want to build from a culture standpoint”