Unconventional building advice for the new AI era | Keith Rabois (Khosla Ventures)

Why does it matter? Because the most common startup advice — talk to customers, hire experienced people, protect psychological safety — may be actively destroying your company.

Keith Rabois has been on the inside of PayPal, Square, LinkedIn, and the boards of Stripe, Airbnb, DoorDash, and Ramp. His operating philosophy runs directly against the consensus. This conversation is a systematic dismantling of received startup wisdom, replaced with a sharper, more demanding framework built from 25 years of pattern recognition at the highest levels.

• Only 2–15 people in any company can independently drive an initiative from inception to success — hiring more people without finding more of them just adds drag
• For consumer and SMB products, customer research is not just unhelpful — it is directionally wrong
• The CEO's job is to push hardest when the company is winning, not when it's struggling
• The PM role as traditionally defined is already obsolete; the skill that replaces it looks more like being a CEO

Most companies hire more people when they need more output — that's exactly backwards

Hiring more headcount without expanding what Rabois calls "barrels" doesn't accelerate a company — it slows it down. A barrel, in his framework, is a person who can take an initiative from inception to success independently: they accumulate resources, motivate people, measure what matters, and deliver the outcome without hand-holding. Ammunition — everyone else — is necessary but only multiplies what a barrel can already do.

At PayPal, a company widely considered one of the greatest talent concentrations in tech history, Rabois counts 12 to 17 barrels among 254 people in Mountain View at acquisition. When he asked Sam Altman on a podcast how many barrels Gladstone had, the answer was two — and Rabois called that a more common answer for a very good company. The ratio of barrels to ammunition, he argues, is the only thing that determines how many important initiatives a company can pursue simultaneously.

Stacking more ammunition behind the same barrel doesn't create parallelism — it creates coordination tax and a drag coefficient that makes the company slower. The smoothie test at Square became a canonical illustration: an intern named Taylor Francis solved on his second day what an entire office team had failed to deliver for weeks. That's a barrel.

The diagnostic question before any hire: is this person a barrel, or am I just adding ammunition behind an existing one? If it's the latter, the hire will make things worse, not better.

Asking consumers what they want is not just useless — it actively points you in the wrong direction

"I hate talking to customers. I refuse to allow colleagues of mine to talk to customers." Rabois isn't being provocative for effect — he believes customer research for consumer and SMB products is directionally harmful, not merely unhelpful.

The mechanism: consumer purchases are subconscious decisions. When you ask someone to consciously explain a subconscious choice, they construct a plausible-sounding rationalization that bears little resemblance to what actually drove the decision. His go-to example: ask anyone who drives a Porsche or Lamborghini why they bought it, and 99% of the time they will give you every reason except the real one.

The implication is brutal. Every team that has sat through a meeting where someone says "I talked to eight customers" and then pontificates for an hour is consuming information that is not merely noisy — it is structurally misleading. Rabois sees this as so fundamental that he extends it to SMB products like Square: "anything sub-mid is directionally dangerous."

The exception is hardcore enterprise, where a named decision-maker is making a predominantly utilitarian choice. There, customer development works. But the moment you're targeting consumers at scale — a billion people, Tony Xu's delivery button, Brian Chesky's room-sharing marketplace — you need foundational insight instead. When Brian Chesky pitched Airbnb, what convinced Rabois wasn't customer interviews. It was a specific Craigslist data point: 30 listings in the Bay Area from people who had already typed in, unprompted, that they wanted to rent out a bedroom.

Push hardest when you're winning — be the coach, not the critic, when you're losing

Mike Moritz's answer to the question of what the best CEOs have in common: "the relentless application of force." Rabois adopted this, but with a crucial inversion that most leaders get backwards.

Complacency doesn't emerge from failure — it emerges from success. The more traction a company has, the more the organization naturally coasts. And the CEO's singular job, in Rabois's view, is to offset that complacency. So the direction of pressure should be precisely the opposite of what feels intuitive: maximum criticism when things are thriving, maximum support when things are struggling.

When a company is struggling, the founder already knows. Piling on criticism doesn't help them solve their problems — it just demoralizes. Rabois describes his own behavior as a board member in these moments: coach and supporter, not critic. But when the company is thriving and everyone is borderline complacent, that's exactly when to isolate the problems that will eventually become existential.

The reason this works isn't just strategic — it's psychographic. Truly talented people are like elite athletes. When things are going well and the team is coasting, morale among the best people actually drops. They have what Rabois calls an "internal claw tempo" — a compulsive drive to create and drive forward. Lenny confirmed this from his time at Airbnb: the brief periods of relative calm saw morale decline, not rise. The best people need the friction.

Criticizing people in public isn't aggression — it's how you optimize the system instead of just the individual

Private feedback is an optimization for the atomic unit — the individual receiving the note. Public feedback optimizes the system. Rabois absorbed this from one of the founders he works with closely, and the logic, once laid out, is hard to argue with.

When you pull someone aside to address a problem, their colleagues already have suspicions something is off. By siloing the feedback, you leave the rest of the team in a state of anxious uncertainty — they can see the problem persisting but have no visibility into whether leadership is aware or acting on it. Public criticism solves this: it signals that the issue is identified, it's being addressed, and it invites the team to help. Colleagues who might be able to contribute a solution can raise their hand. The feedback loop closes faster.

The cultural prerequisite is a team organized around winning, not around comfort. "High performance machines don't have psychological safety. They're about winning." Rabois recommends the Jordan Rules as the reading assignment for anyone who finds this framing uncomfortable — Michael Jordan's approach to teammates was not gentle, and the Bulls won six championships. There's an art to the mix: some things belong in front of the team, some to the individual, and the best coaches know which is which. But the default should run toward public, not private.

The PM role is already obsolete — what replaces it looks like CEO-level business judgment

Peter Fenton convinced Rabois, and Rabois is now convinced: the traditional PM function — customer inputs, sequential roadmaps organized by quarter, intermediary between engineering and business — doesn't survive the current rate of AI capability improvement.

"There are things that were impossible to do in November that are actually pretty easy to do right now in March." A year-long roadmap in this environment isn't just suboptimal — it's incoherent. The organization that wins will notice something newly possible this week and have it in front of customers next week.

What replaces the PM? Rabois frames it as CEO-level skill: understanding what to build and why, grounded in genuine business acumen — knowing the company's equation, where value is created, and what the inputs and outputs actually are. Alfred Lin's chef analogy resonates with him: a great chef at a famous restaurant isn't cooking the dishes, they're making commercial decisions about positioning, differentiation, segment, and pricing.

The engineering director at Ramp who personally ships as much code as he did as an individual contributor — while managing a team of 20, using AI as a second team — is Rabois's model for where this goes. Max Levchin had commercial instincts as an engineer. Jeremy Stoppelman had them before he started Yelp. The premium going forward goes to whoever combines technical execution with business judgment, regardless of what their title used to be.

Startups can't win a salary-cap war for known talent — the alpha lives in candidates that big-company recruiting machines will mis-evaluate

Peter Thiel told Rabois this on a jog around the Stanford campus during his first week at PayPal: the only way to build against large incumbents with infinite compensation budgets is to find undiscovered talent. Rabois has been on that crusade for 25 years.

A startup operates at roughly one-tenth the salary cap of its competitors. Competing for recognized stars is a structurally losing game. The better strategy is to develop a systematic lens for identifying why a specific candidate will get incorrectly filtered by large-company recruiting processes — and then go get that person.

Large organizations evaluate homogeneously. They run candidates through a black box. Rabois's heuristic: imagine this person is interviewing at Meta, Google, Block, or Coinbase. What are they going to miss, and why? That gap is the recruiting alpha. One reason this skews toward younger candidates isn't ageism — it's that younger people have less data in the employment equivalent of a FICO score, making them harder for the homogeneous machine to evaluate and easier for a contrarian recruiter to see clearly.

Most of the thriving companies Rabois works with have made this a strategy rather than an accident: skip senior external hires almost entirely and promote from within. Ramp did it. Trade Republic did it. One company he board-sits uses the chief of staff role as a deliberate talent factory — the current CMO was the last chief of staff, the incoming head of product is the current one.

Operating tempo between board meetings is the clearest early signal that a company will win

Roloff Botha joined the Square board at Series B and told Rabois six months in — two board meetings — that he hadn't seen this kind of tempo since the PayPal days. After nine years as a VC, that was his read. The specific pattern: identify a problem or opportunity at board meeting X, ship a solution and measure its impact before board meeting X+1, every time, consistently.

Delian Asparouhov shadowed Rabois at two consecutive Fair board meetings and said it was the one company in Silicon Valley that would make him leave venture. Same observation: something slightly off, root causes identified, solution shipped, impact measured — before the next meeting.

Ramp was the precipice-of-shipping example. Shipping a corporate card typically takes nine to twelve months of program managers, sponsoring banks, and regulatory scaffolding. Ramp was on the precipice in three months. Rabois led the seed in May 2019 and gave a term sheet to preempt the Series A in September. Tempo was one of two inputs.

The compounding effect is the point. Teams that close the identify-ship-measure loop in days rather than quarters build an execution advantage that slower competitors cannot close. Ramp now opens every board meeting with a slide: Day 1,184.

The CMO — not the CTO — is the most powerful AI user in the best organizations right now

The number one consumer of tokens at some of the best organizations Rabois works with, including OpenDoor, is the CMO. Not the engineering team. Not the CTO. The CMO.

The explanation is intellectual curiosity meeting sudden access. Marketing executives have always needed analytics, campaigns, and creative — but historically had to route requests through other teams, wait in queues, and accept whatever came back. AI collapsed that dependency chain. Now the intellectually curious executive can do it with their own hands: "Wow, there's all these cool things I can do now." They don't need deputies and deputies and deputies to get actual work product. They're just shipping.

Rabois's frame for surviving the AI era isn't working harder — it's intellectual curiosity. The business-person who leans in and becomes the top token consumer on their team doesn't just save time. They restructure how work gets done at the top of the org chart entirely, eliminating layers of coordination that previously slowed every decision.

The gap between barrels is where the next decade gets decided

Every framework Rabois laid out — barrels vs. ammunition, undiscovered talent, operating tempo, public criticism — points at the same underlying dynamic: the bottleneck in building great companies has always been the density of people who can independently drive outcomes, and AI is now making that bottleneck starker, not softer. Execution is getting cheaper. Judgment is getting more expensive. The organizations that figure out how to identify, develop, and retain the people with genuine strategic instincts — and ruthlessly clear everything else out of their way — will compound into something their competitors cannot replicate. The team you build is still the company you build.

---

Topics: talent, hiring, AI, product management, leadership, startups, venture capital, team building, operating, career advice