The Company That Told People Not to Come Won the AI Talent War

Deep thoughts on AI and aspirations —— ByteDance Deep Thinking Circle

When Meta launched its astronomical poaching campaign, two leading AI companies responded in two different ways. OpenAI followed market convention: counter offers for targeted employees, retention bonuses, removing vesting cliffs for new hires so stock vested faster. Anthropic’s response was unusual. Management told employees: you came here first for the mission, not to continuously bid up your price in external auctions. We won’t give you ten times the salary of equally excellent colleagues just because Zuckerberg happened to target you. That’s not fair. If you want to leave, leave.

The results speak for themselves: OpenAI lost dozens of people, Anthropic lost only 2—and those two were already former Meta employees who’d worked there for 6 and 11 years. That same year, Anthropic’s ARR grew from $9 billion to $45 billion, and its secondary market valuation reached $1 trillion, overtaking OpenAI.

The confidence to tell astronomical offers “leave if you want” doesn’t come from money—it comes from an organizational design that outsiders find hard to understand. This design is worth examining because it may be the most important organizational case study in the AI industry over the past three years.

Recruiting Starts with One Thing: Screening Out the Wrong People

Anthropic’s hiring logic is the opposite of industry standard. The mainstream approach is to hire the strongest people possible; theirs is to screen out the wrong people as early as possible.

Interviews include a dedicated culture round—one hour with 15 to 20 scenario questions. The most典型 leaked question: if the company ultimately decides not to release a model because it cannot guarantee safety, would you accept your stock going to zero? This round screens for three things: whether mission truly comes before profit; whether you’re a kind person with low ego who can admit ignorance and mistakes; whether you can handle complexity and reason through how one decision affects other areas.

This filter has costs. They genuinely turn away engineers with stronger backgrounds and capabilities. Before Stripe’s former CTO Rahul Patil decided to join, he had a long conversation with Anthropic’s then-CTO, who not only didn’t recruit him but spent two or three weeks repeatedly discussing why he shouldn’t join—unless you truly align on culture and mission, it’s not worth coming. The head of growth summed up the strategy: we’re very good at weeding out people who come for money and prestige.

The Place Most Prone to Silos in the World Doesn’t Have Them

Frontier AI labs are the most prone to internal politics. Researchers are among the smartest and highest-ego people, naturally wanting to propose different solutions and establish their own domains, while compute and resources are extremely limited—conflict is almost inevitable. People who’ve jumped from elsewhere describe other model companies internally like separate fiefdoms, each doing their own thing and quietly competing.

Anthropic’s approach is to write values directly into governance structure. The company has 7 co-founders, and Dario insisted on equal equity for all. Everyone advised him this was a disaster—ambiguous authority, misaligned incentives, the company would easily splinter through infighting. His understanding: the company revolves around the mission, not around any single founder, and equal equity is the most unforgeable proof of this principle. Seven people means seven cultural replication nodes; as the company scales rapidly, the original culture is less likely to be diluted.

Another move targets titles. Below executive level, no title distinctions—everyone is called MTS, Member of Technical Staff, deliberately erasing identity lines between researcher and engineer, senior and junior. The contrast is stark. OpenAI has a recognized internal hierarchy: researchers above research engineers, above software engineers; product often subordinate to research, with research teams only notifying product teams after new results, then looking for nails with their hammer.

Maintaining cultural density requires grunt work. Dario himself says he spends a third to 40% of his time ensuring the culture is good—this is his highest-leverage work. Every two weeks there’s an all-hands, where he speaks for an hour on everything from company direction to industry changes, taking live questions. He records in his own channel without filter what he’s worried about and how he sees issues people care about. Transparent to the point of being challengeable: someone hears the all-hands, disagrees, goes directly to his channel to publicly dispute a judgment, and debate unfolds on the spot. This atmosphere produces: in complex, changing circumstances, everyone knows how decisions are made, enabling relatively consistent distributed judgment.

Why This Matters Especially in AI

Culture isn’t just atmosphere—it directly creates moats. The hardest moat in AI competition, at the deepest level, is data: the tasks themselves, execution traces, evaluation, verification systems. These are all unglamorous work—critically important when done well, yet unlike publishing a paper or launching a new product, they don’t become personal highlight moments. From feedback across companies, one of OpenAI’s organizational challenges is getting hundreds of people with excellent backgrounds and high ambitions to seriously work on data. They won on paradigm breakthroughs, so naturally people want to make their own next bet rather than clean up the last one’s mess.

Anthropic does the opposite. Co-founder Jared Kaplan personally leads data review daily, with extremely meticulous cleaning. Low-ego, mission-driven organizations have an advantage amplified by this era: they can organize the smartest people around unglamorous work. The results correspond: OpenAI’s models remain strongest on competition-level programming challenges—more of a research problem; on everyday agentic tasks Claude often leads—more of an engineering problem, testing data, systems, and execution details. Researcher Shunyu Yao put it more directly: the era of individual heroism is over; the most important qualities are reliability and attention to detail.

This also explains the opening poaching story. People who come for money and prestige are precisely the ones who can’t be organized to do unglamorous work. Screening them out isn’t moral purity—it’s competitive strategy.

What You Can Copy, What You Can’t

At this point it’s tempting to copy. First distinguish three things.

Mission can’t be copied. Anthropic has people who truly put mission before company survival. An early employee said at an all-hands: if the mission is ultimately achieved while the company itself fails, this would still be a good outcome. Most companies don’t have such a mission. Forcing it only produces team-building rhetoric, and employees can tell the difference better than anyone.

Structure can be copied. Reverse-screening scenario interviews, unified titles to suppress silos, equal founder equity as proof of commitment, high-frequency transparent writing from leadership, distributing trade-offs to the front line rather than escalating to executives for judgment. These are mechanisms that fit any company, requiring no world-saving narrative first.

Origins can’t be copied, and are most easily overlooked. This culture is largely the reaction force of founder trauma. Dario saw control struggles tear apart teams at Baidu’s AI lab, experienced high-level trust collapse at OpenAI firsthand, and ultimately left to start a company with the core team. Having seen how political maneuvering exhausts trust, he emphasizes authentic transparency. Having seen political struggles dissolve organizations, he encourages surfacing conflict early. Values come half from what you believe, half from what you’ve deeply detested—the latter often has stronger shaping force, and disgust can’t be imitated.

Finally, a calibration to prevent this narrative from being sold as a cure-all. This is a three-year sample, not a permanent answer. Programming is already a known card, with developers showing signs of migrating from Claude Code to Codex. Compute is becoming the new battleground, and OpenAI locked in far more early compute than Anthropic. Cult-like organizations pivot slowly; they may not have the advantage in the next paradigm leap, and high retention may mean people who should leave haven’t. Anthropic’s own people say culture is the most defensible thing. That’s true, but it defends against talent wars, not technology wars.

One action item for founders: check what your filter is screening for. If it screens for the brightest resumes and highest ambitions, in an era competing on unglamorous work, the direction may be exactly backwards.