In the summer of 1948, a 32-year old mathematician working at Bell Labs published a paper in The Bell System Technical Journal. Its title was characteristically modest: A Mathematical Theory of Communication. Its consequences were anything but.
His name was Claude Shannon.
Shannon had spent much of his career thinking about a problem that had occupied Bell Labs for decades: how do you transmit a message from one place to another, reliably, through an imperfect communications system?
His answer was to strip communication down to something more fundamental. A message could be treated not primarily according to what it meant, but as information. Information could be quantified. It could be encoded, transmitted through a noisy channel and reconstructed at the other end. The implications escaped the telephone network almost immediately. Information theory would become part of the intellectual foundation of the digital age. Computers, compression, cryptography, telecommunications and eventually the internet all developed in a world Shannon had helped make mathematically legible.
Nearly eighty years later, one of the world's most capable artificial intelligence systems carries his first name.
Claude.
The name is deliberate. Anthropic (which is Greek for “human being”) named its family of AI models after Claude Shannon. It is a fitting inheritance. Shannon spent his life trying to understand how information could move through machines. Claude exists in a world where machines can increasingly interpret that information, transform it, write with it, reason across it and act upon it.
But the name is not the only thing that makes the lineage interesting.
The more closely we look at Anthropic, the more that a curious resemblance starts to emerge.
Anthropic does not describe itself simply as the company that makes Claude. On its homepage, the company leads with “AI research and products that put safety at the frontier.” Elsewhere, it describes itself as an AI safety and research company building reliable, interpretable and steerable systems. Its research operation now extends across alignment, interpretability, economics, societal impacts and frontier red teaming.
At the same time, an extraordinary collection of people has begun to congregate there.
Researchers are arriving from the world's leading AI laboratories. Founders who have already built generation defining technologies are joining. Anthropic is moving beyond software into chips, data centres and the energy infrastructure beneath them. Its researchers are studying not only how to make artificial intelligence more capable, but how it works internally, how it changes economies, how it interacts with institutions and what happens when increasingly capable machines begin participating in research themselves.
The deeper Anthropic goes into the problem of intelligence, the larger the institution required to pursue it seems to become. Bell Labs began with a similarly practical problem: how to build a better communications system. Following that problem would eventually take its researchers far beyond the telephone. Into mathematics. Physics. Materials science. Electronics. Computing. Along the way, they would produce the transistor, information theory, Unix, C, and a remarkable collection of discoveries upon which much of modern technological life was eventually built.
But the inventions were only the visible outputs. Behind them was something perhaps just as important: an institution capable of repeatedly attracting exceptional people and creating the conditions in which they could do important work.
Perhaps the greatest invention any lab can produce is the lab environment itself.
According to SignalFire's 2025 State of Tech Talent Report, Anthropic had the highest two-year employee retention rate of the major frontier AI labs it analysed. Eighty percent of employees who had joined at least two years earlier were still there at the end of year two, compared with 78% at Google DeepMind, 67% at OpenAI and 64% at Meta. Its analysis also found that engineers were eight times more likely to move from OpenAI to Anthropic than in the opposite direction, and nearly eleven times more likely to move from DeepMind to Anthropic than the reverse.
Not only is retention sky high. The density of talent is unrivalled too.
In 2024, Mike Krieger, the co-founder of Instagram and its former CTO, joined Anthropic as Chief Product Officer. Two years later, Andrej Karpathy, an OpenAI founding member who had previously led AI at Tesla, joined its pretraining team. Tom Blomfield, the co-founder of Monzo and GoCardless and a former partner at Y Combinator, took a leave from YC to join Anthropic's compute team. Even Rishi Sunak, the former Prime Minister of the United Kingdom, now serves as a senior adviser to the company on strategy, macroeconomics and geopolitics.
Anthropic has recruited energy and data-centre executives from Google, chip engineers as it develops its own silicon capabilities, senior figures to expand its operations internationally, and researchers from across the small universe of people capable of working at the frontier of artificial intelligence. Any one of those appointments would make a good press release. Taken together, they suggest something more unusual: Anthropic is becoming a gathering place for exceptional people across technology, science, infrastructure, entrepreneurship and public life.
That gives us an opportunity to ask a question that is ultimately much more useful than whether Anthropic will produce AGI:
Why are exceptional people choosing Anthropic?
Why, in perhaps the most competitive market for technical talent in the world, are people who could work almost anywhere choosing to go there? Why do they appear unusually likely to stay? What kind of mission attracts them? What kind of leader holds them together? What happens when enough of them begin working on the same problem?
And, perhaps most importantly for the rest of us:
What can we learn from it?
1. Mission as a talent strategy
Most companies compete for talent with some combination of money, status and opportunity. Anthropic has plenty of all three. But there is something else it can offer that is much harder to manufacture: a problem that feels important enough to devote a career to.
Anthropic's language is revealing. The company does not describe its mission as building the world's most capable chatbot, winning the AI race or even creating artificial general intelligence. Instead, it says that it builds “AI to serve humanity's long-term well-being.” It describes itself as a public benefit corporation dedicated to securing the benefits of artificial intelligence while mitigating its risks.
Corporate mission statements are easy to dismiss. Most deserve to be. They are often broad enough to mean almost nothing, written after a company has already decided what it wants to sell and applied as a layer of purpose over a fundamentally commercial objective.
Anthropic's appears to be more consequential because the mission is reflected in what the company chooses to work on. Alongside building Claude, Anthropic has invested heavily in interpretability, alignment, cybersecurity, economic research, model welfare, societal impacts and methods for evaluating potentially dangerous capabilities. The definition of the problem extends well beyond the product.
To understand why, it helps to understand the worldview of the person leading it.
In October 2024, three years after founding Anthropic, Dario Amodei published a 14,000-word essay titled Machines of Loving Grace. At the time, Claude was becoming a serious competitor to ChatGPT, but many of the people now joining Anthropic had not yet arrived.
The essay was unusual because Amodei had become one of the technology industry's most prominent voices warning about the potential dangers of advanced AI. Machines of Loving Grace was his attempt to articulate the other side of that conviction: what exactly are we taking all of this risk for?
“There has to be something we're fighting for,” he writes.
His answer is enormously ambitious.
Amodei asks us to imagine millions of copies of an AI system, each more capable than a Nobel Prize winner across fields including biology, mathematics, engineering and programming, able to work many times faster than humans and collaborate with one another. His shorthand for this hypothetical system is memorable:
“A country of geniuses in a datacenter.”
But what is more revealing is what he wants this country of geniuses to do.
The essay is not primarily concerned with better search engines, more productive knowledge workers or cheaper software. Amodei works through five areas in which he believes powerful AI could fundamentally alter the human condition: biology and physical health, neuroscience and mental health, economic development and poverty, peace and governance, and work and meaning. He imagines AI compressing decades of biological progress into years, helping eliminate diseases, accelerating development in poorer countries and strengthening democratic institutions.
Whether those predictions turn out to be prescient or wildly optimistic is almost beside the point for our purposes.
What matters is the scale of the project being offered to the people Anthropic wants to recruit.
“Build the best chatbot” is an interesting technical challenge. “Help determine whether extraordinarily powerful machine intelligence improves the long-term trajectory of humanity” is something closer to a life's work. That distinction matters when the people you are trying to recruit have options almost everywhere.
Someone like Andrej Karpathy did not need another prestigious job. Mike Krieger did not need another successful technology company on his CV. The researchers Anthropic is competing for can choose between OpenAI, DeepMind, Meta, startups, academia, their own companies and compensation packages that would have seemed absurd even by Silicon Valley standards a decade ago. At that level, recruitment becomes about more than employment. It becomes a competition over which problem deserves someone's finite working life.
This may help explain something about Anthropic's gravitational pull. A sufficiently ambitious mission expands the psychological surface area of a company. It creates room for people with ambitions that exceed any individual product roadmap. It also creates room for more kinds of people.
If Anthropic's mission were simply to build Claude, the organisation required to accomplish it would be relatively obvious: machine learning researchers, software engineers, product designers and the commercial functions required to distribute what they make.
But “serve humanity's long-term well-being” is a much larger problem.
Suddenly you need researchers capable of making models more intelligent and researchers capable of understanding what is happening inside them. You need people studying how AI changes labour markets, how it might be exploited by malicious actors, how governments should respond to it and how increasingly capable systems should behave. As the models become more powerful, you begin encountering questions of silicon, energy, data centres, biology, economics and geopolitics.
The mission becomes a kind of organisational gravity. The deeper you pursue it, the more disciplines it pulls towards itself, which creates a flywheel of talent acquisition in order to fulfil the scale of it.
Amodei himself has alluded to it when he talks about Anthropic’s future growth model as expanding to areas that are “complementary to intelligence.”
Of course, we should be careful not to pretend that Machines of Loving Grace is some secret blueprint from which every subsequent Anthropic hire was derived. There is no evidence of that. But there is a striking coherence between the problem Amodei described and the institution Anthropic is becoming. And that gives us the first useful lesson from Anthropic for companies operating nowhere near the frontier of artificial intelligence.
Exceptional people are attracted to problems and visions large enough to contain their ambitions. Missions that can be aligned with their life’s work.
A company's product is not necessarily its purpose.
Products are temporary expressions of a problem. They change as technologies change, markets move and organisations learn. A sufficiently important problem can remain interesting for decades. Bell Labs did not keep extraordinary people together because the telephone was an endlessly fascinating product. Communication was an endlessly deep problem. Anthropic may be discovering the same thing with intelligence.
The larger lesson here is not to manufacture a bigger mission. It is to find the biggest truthful description of the problem you are actually trying to solve.
Because if you want exceptional people to spend a meaningful portion of their lives working with you, the work has to be large enough to deserve it.
2. Designing an organisation for exceptional people
A compelling mission may help explain why exceptional people join Anthropic.
But it does not explain why they stay.
Silicon Valley has never struggled to attract ambitious people to ambitious ideas. The harder problem is creating an environment in which those people continue to believe they can do their best work after the excitement of joining has worn off. Anthropic appears unusually good at this.
SignalFire's retention data is worth returning to here. Eighty percent of Anthropic employees in its analysis remained at the company two years after joining, the highest rate among the frontier AI labs it studied. In a labour market where the most capable researchers can command extraordinary compensation packages and rival laboratories routinely attempt to recruit one another's staff, retaining four out of every five people is striking.
The question is why.
Part of the answer may lie in an unusual assumption Anthropic seems to make about the people it hires: if you recruit exceptionally capable people, you should design the organisation to let them behave like exceptionally capable people.
You can see this most clearly in the way Anthropic thinks about technical roles.
At many technology companies, the boundary between research and engineering is relatively clear. Researchers discover. Engineers build. One produces papers, the other products. Anthropic deliberately makes that boundary porous.
“We generally don't distinguish between engineers and researchers,” the company tells prospective employees. “Engineers here do lots of research, and researchers do lots of engineering.” Engineers frequently appear as first authors on Anthropic papers, while researchers contribute directly to the systems behind its products.
Even the title is revealing. Rather than constructing an elaborate hierarchy of research and engineering designations, much of Anthropic's technical organisation operates under the deliberately broad title Member of Technical Staff.
Titles tell people where they sit in a hierarchy. Job descriptions tell them which problems belong to them. Organisational boundaries determine which questions someone is permitted to ask. Status is itself an organisational design tool.
What an institution chooses to make prestigious tells ambitious people what kind of person they should become. If status comes primarily from managing the largest team, talented people eventually become managers. If it comes from proximity to leadership, they learn to become political. If it comes only from shipping the most visible product, less visible but potentially important research becomes harder to justify.
Anthropic's broad technical titles do not eliminate status, hierarchy or politics. No title can. But they suggest an attempt to prevent the distinction between researcher and engineer from becoming a distinction between different classes of technical person.
Anthropic makes another unusual claim on its careers page: roughly half of its technical staff had no prior machine learning experience before joining the company. Instead of treating a conventional AI resume as the primary evidence of someone's ability to contribute, Anthropic says it cares about what candidates can do rather than where they learned to do it.
There is a philosophy embedded in this.
Hire for intellectual capability, then give that capability room to move.
That seems particularly important in a field changing as quickly as artificial intelligence. If the important problems three years from now are not the important problems today, recruiting people exclusively for today's neatly defined specialisms may become a liability. Intellectual range becomes valuable precisely because the frontier keeps moving.
Most organisations respond to growth by adding structure.
More managers. Clearer reporting lines. Narrower responsibilities. Increasingly precise definitions of who owns what. Some of that is inevitable. A thousand people cannot coordinate themselves in the same way as ten.
But organisations built around unusually capable people face a particular tension. They have to add enough structure to coordinate people without removing the agency that made those people exceptional in the first place. Great research environments therefore seem particularly concerned with preserving agency.
Anthropic's blurred boundaries between research and engineering, broad technical roles and willingness to let people move across problems enables its members to organise over agency rather than rigid and slow structures.
Dario Amodei's leadership style appears designed around a similar premise.
His role is not simply to set strategy and allocate resources. He spends a surprising amount of time explaining to Anthropic's employees how he thinks the future might unfold and why the company is making the choices it is making. Company-wide talks have reportedly become known internally as “Dario vision quests”, extended attempts to reason through the technological landscape with the organisation rather than simply hand it a quarterly plan.
As an organisation becomes filled with increasingly capable people, coordination through instruction becomes less useful. You cannot personally tell hundreds or thousands of researchers what to think. Nor would you want to. What a leader can do is give them a sufficiently coherent model of the world that they can make good decisions independently.
Instead of:
Here is what you need to do.
It becomes:
Here is what we believe is happening. Here is what we are trying to accomplish. Use your judgement.
The stronger the shared context, the less every decision needs to travel upwards. As such, this is a culture of context rather than command. Context creates routes for exploration that go in a million directions to find unexplored horizons. Command is linear and limiting.
Anthropic's culture of long-form internal writing appears to serve a similar purpose. Employees use internal channels for extended arguments and technical thinking, including writing by Amodei himself. Ideas can be challenged, refined and absorbed by people beyond the meeting in which they originated.
This matters because talent density without information density has limited value.
It is not enough to employ brilliant people. They need access to one another's thinking.
Moreover, Anthropic now has another mechanism for expanding that access and capability that Bell Labs could never have had.
Claude itself.
In 2025, Anthropic studied how its own engineers and researchers were using Claude Code. It surveyed 132 technical employees, conducted 53 interviews and analysed internal usage data. One of the patterns it found was that employees were increasingly able to work outside their previous areas of expertise. Engineers described becoming more “full-stack”, tackling unfamiliar parts of the codebase and completing work that previously might have required assistance from another specialist.
This creates an unusual organisational possibility.
Normally, as companies grow, expertise becomes more specialised. Specialisation creates departments. Departments create boundaries. Boundaries create coordination costs.
Anthropic is developing a technology that may push partially in the opposite direction.
A security researcher can become more capable at software engineering. An engineer can navigate an unfamiliar codebase. A researcher can prototype an idea without waiting for another team. Claude effectively increases the range over which each person's existing intelligence can operate.
The organisation is therefore not simply accumulating talented people.
It is potentially amplifying the effective range of each person it accumulates.
And that brings us back to retention.
Exceptional people tend to dislike environments in which their ability is unnecessarily constrained. They dislike spending their time navigating bureaucracy, waiting for permission or being prevented from following an interesting problem because it belongs to someone else's department. Compensation can persuade someone to tolerate those conditions for a while. It is harder to persuade them to build a career around them.
Anthropic's emerging model appears to offer something different: an important mission, unusually capable colleagues, substantial intellectual agency, access to powerful tools and the opportunity to work on problems that regularly escape the boundaries of a job description.
If you want exceptional people to stay, do not merely give them important work. Build an environment that makes them more capable.
Give them context. Give them agency. Give them access to one another. Give them tools that expand what they can do. Be deliberate about what your organisation gives status to. Remove unnecessary boundaries around where their intelligence is allowed to travel.
The organisation itself becomes part of the compensation.
And once enough exceptional people begin experiencing that in the same place, another force starts to take over. They begin attracting one another.
3. The Talent Flywheel
There is another advantage to assembling exceptional people: they begin attracting one another. At a certain point, the people already inside an organisation become part of its recruiting strategy.
An ambitious researcher is more likely to join a team containing researchers they admire. A great product builder wants to work alongside other great builders. The arrival of someone like Andrej Karpathy or Mike Krieger therefore matters beyond the individual contribution either might make. Their presence changes the calibre of person for whom Anthropic becomes an interesting place to work.
This creates a flywheel.
Mission attracts exceptional people. Exceptional people produce important work. Important work builds reputation. Reputation attracts more exceptional people. A snowball effect ensues.
Each strong hire changes the set of people you can plausibly recruit next, which also introduces a signalling effect. People with exceptional ability tend to have exceptional optionality. When enough of them independently choose the same organisation, their decisions become information for everyone watching.
4. Follow the problem until you find the constraint
As mentioned earlier, something interesting happens when talented people are given a sufficiently large problem: the boundaries of the problem begin to expand.
Anthropic started as an AI research company. But pursuing the problem of intelligence seriously has pulled it into considerably more than building models. If you want to make intelligence more capable, you need better training methods and more computation. More computation requires chips. Chips require data centres. Data centres require enormous quantities of energy. Each layer eventually exposes the constraints of the one beneath it.
The same thing happens in the other direction. Powerful intelligence has to become useful and widely available. Deploy it across industries and you encounter questions of cybersecurity, labour and economics. Deploy it across countries and you encounter different languages, cultures, regulations and infrastructure. As the technology becomes more consequential, governments and geopolitical institutions inevitably become part of the problem too.
Anthropic's recent hiring increasingly maps onto these constraints. Sana Ouji, recruited from Google, was brought in to work on global infrastructure and energy strategy as access to power and data centres becomes increasingly important to frontier AI. Anthropic has also been recruiting chip engineers as it develops its own silicon capabilities, moving deeper into the physical infrastructure beneath its models. Steve Jarrett, previously Chief AI Officer at Orange, joined to lead the adaptation of Anthropic's products across Europe and Africa, where deploying the same underlying intelligence means navigating very different markets and institutions.
Then there is the layer above the technology itself. Anthropic brought in Mariano-Florentino Cuéllar, a former California Supreme Court justice and international affairs leader, to work across policy and global affairs. Rishi Sunak, the former British Prime Minister, joined as a senior adviser on strategy, macroeconomics and geopolitics. These appointments look very different from hiring another machine learning researcher, but that is precisely the point. As the capabilities and consequences of the technology expand, the range of expertise required around it expands too.
Anthropic keeps following the problem until it finds the constraint.
This gives us a useful way to think about how organisations expand. Companies often ask, what else could we do? Perhaps the better question is, what is now preventing us from solving the problem?
The distinction matters. The first question can lead to adjacent products, services and business lines simply because they represent opportunities for expansion. The second produces new capabilities because they have become necessary to solving the original problem more completely.
Sometimes that means moving down the value chain. Sometimes it means moving up it. Sometimes it means moving sideways into a field the organisation never expected to enter. The direction matters less than the constraint. If compute constrains intelligence, silicon becomes relevant. If energy constrains compute, energy becomes relevant. If geography constrains adoption, local market expertise becomes relevant. If governments constrain deployment, policy and geopolitics become relevant.
Follow the problem until you find the constraint. Then decide whether that constraint needs to become part of your business.
The second half of that idea is important. Not every constraint should be internalised. A company can partner, outsource, integrate or simply accept certain limitations. Trying to own everything your business touches is more likely to produce distraction than advantage. The strategic judgement is determining which constraints are important enough that you can no longer afford to leave them entirely to somebody else.
This is how an organisation can become multidisciplinary without becoming unfocused. It isn't expanding because it wants to do more things. It is expanding because solving one problem properly increasingly requires more kinds of expertise.
This is also why mission matters beyond recruitment. A sufficiently clear mission gives talented people a compass when they reach the edge of what the organisation currently knows how to do. Instead of defining the company exclusively by its present capabilities, it defines the problem those capabilities exist to solve.
The company's eventual scope can therefore be discovered rather than entirely predetermined. Great organisations don't become multidisciplinary because interdisciplinarity sounds good. They become multidisciplinary because solving an important problem eventually forces them beyond the boundaries of their existing expertise. Choose the problem carefully, then let the constraints show you where to go next.
5. Ideas escaping the building
Eventually, an institution has to be judged by what comes out of it.
Anthropic's most obvious output is Claude, but some of its more interesting work sits beneath the product. The company developed Constitutional AI, a method for training models to follow a written set of principles rather than relying exclusively on human feedback. Its interpretability researchers are attempting something even more fundamental: understanding how concepts and behaviours are actually represented inside neural networks.
Then there is the Model Context Protocol, or MCP.
Anthropic introduced MCP in late 2024 as an open standard for connecting AI systems to external tools and sources of data. Instead of every developer building a bespoke connection between an AI model and every database, application or service it needs to use, MCP proposed a common interface. By 2025, OpenAI had adopted MCP across its products, followed by Google DeepMind, Microsoft and a growing ecosystem of software companies. Anthropic subsequently donated the project to the Linux Foundation's Agentic AI Foundation. MCP had escaped the company that created it and started becoming infrastructure for the wider industry.
This distinction is important, because most companies want to build products. Research institutions sometimes produce something stranger: primitives. Ideas, methods and standards that other people begin building upon.
Claude Code may eventually prove to be another example. It began as a tool for Anthropic's own engineers before becoming a commercial product, and its increasingly agentic model of software development is now influencing how developers think about the relationship between programmers and machines. Anthropic has also created a dedicated Labs organisation to explore products at the edges of Claude's capabilities, while simultaneously moving deeper into the physical infrastructure required to build those capabilities. As such, Anthropic is producing internal ideas at a remarkable rate, and the best ideas eventually escape the building.
They become papers other researchers extend, methods other laboratories adopt, standards competitors implement and primitives upon which people the original inventors will never meet begin to build.
Anthropic is only five years old. We cannot yet know whether any of its work will achieve that kind of durability. But MCP offers an early glimpse of what it looks like when an idea stops belonging entirely to the company that created it. And perhaps that is the output worth watching most closely.
6. The Researcher Inside The Machine
There is one final dynamic inside Anthropic that may prove more consequential than any individual hire. Anthropic's people are not merely producing Claude. They are increasingly using Claude to produce Anthropic.
The clearest example is software engineering. Anthropic's own research into Claude Code found that its technical staff were using the system to navigate unfamiliar codebases, automate routine work and take on tasks beyond their previous areas of expertise. The result is not simply faster engineers. It is engineers whose effective range has expanded.
Now the same dynamic is beginning to move closer to the research itself. It could be said that part of what attracts talent to Anthropic's team is the possibility of using Claude to accelerate the process of improving the very same tool. The model becomes not simply the object being researched, but increasingly a tool used by the researchers doing the research. That creates a strange recursive loop.
A talented researcher helps build a more capable Claude. A more capable Claude increases what that researcher can do. The researcher can run more experiments, explore unfamiliar areas and move more quickly from an idea to an implementation. That increased capability contributes to better research, which contributes to a more capable Claude.
Anthropic is already studying this process inside its own workforce. That matters because the company is not simply adopting AI and hoping productivity increases. It can observe how its own technology changes the people building it, learn from those changes and feed what it learns back into both the product and the organisation. If this loop continues, talent density becomes even more interesting.
Traditionally, an organisation becomes more capable by adding capable people and helping them work effectively together. AI introduces another possibility: the tools produced by the organisation can increase the effective capability of the people already inside it.
A ten-person team does not literally become a twenty-person team. Something subtler happens. Each person's range expands. The engineer can attempt work that previously required a specialist. The researcher can explore an adjacent discipline without starting from zero. Small groups can test ideas that previously required larger ones.
And the better the underlying system becomes, the stronger that effect could become. This may ultimately be the most important thing to watch at Anthropic. Not simply whether Claude becomes more intelligent, but whether improvements in Claude accelerate the rate at which Anthropic itself can discover and build the next Claude. The product begins improving the institution that produces the product. And it is that kind of compounding, that creates the next Bell Labs.
What we can take away?
Now, most of us are not building frontier models. We do not have billions of dollars of compute, Nobel Prize winners joining our research teams or former prime ministers advising us on geopolitics. But that may be precisely why Anthropic is useful to study.
Extreme environments have a way of making certain principles easier to see. Anthropic is competing for some of the most sought after people in the world, in one of the fastest-moving industries in history, against companies with almost unlimited resources. The mechanisms it uses to attract those people, organise them and keep them working together are exaggerated versions of problems almost every organisation faces.
How do you convince exceptional people to choose you? How do you give them room to do exceptional work? How do you keep them? And how do you create an organisation that becomes more capable as those people spend more time together?
The first lesson is to choose a problem bigger than your product.
Anthropic does not ultimately define itself by Claude. Its mission gives the organisation a problem broad enough that the product can change while the direction remains coherent. This is particularly important for small companies because products are fragile. Technologies change. Markets move. What customers want today may be irrelevant five years from now. A sufficiently deep problem gives an organisation somewhere to go.
This doesn't mean inventing a grandiose mission statement. “Changing the world” is not a strategy. The challenge is finding the largest truthful version of the problem you are actually trying to solve. If you make accounting software, perhaps the deeper problem is giving small businesses control over their finances. If you run a restaurant, perhaps it is creating a place where people feel they belong. The product is simply your current answer to the problem.
That leads to the second principle: give exceptional people a reason to choose you beyond compensation.
Most small companies cannot win a bidding war against Google. Fortunately, compensation is not the only thing exceptional people optimise for. They also choose problems, colleagues, autonomy, learning and the possibility of doing work that will matter.
This means a smaller organisation can sometimes offer something a larger one cannot: disproportionate agency. The chance to own something consequential. Direct access to the problem. The ability to see their fingerprints on what eventually gets built.
But agency only works if you actually give it to them.
If you recruit someone because you believe they are exceptional and then tell them exactly how to do their job, you have paid a premium for intelligence you are refusing to use. Hire people whose judgement you trust, give them context, then create enough space for that judgement to surprise you.
That requires being deliberate about status too. Every organisation teaches people what behaviour gets rewarded. If status accrues to whoever manages the most people, ambitious employees will seek headcount. If it accrues to whoever is closest to the founder, they will seek proximity. If the organisation visibly rewards original thinking, excellent craft, useful disagreement and important work, ambitious people will orient themselves towards those things instead.
Culture is partly the answer to the question: what does this group make prestigious?
Then there is talent density itself.
Your first exceptional hires do more than increase the capability of the company. They change who you can attract next. Great people want colleagues from whom they can learn. Three extraordinary people working together can therefore become the recruiting argument for the fourth.
This is why hiring slowly and maintaining a high bar can matter disproportionately at the beginning of a company. You are not merely filling today's roles. You are determining the quality of the room that tomorrow's candidate will be asked to enter.
Once those people arrive, the objective shifts from accumulation to compounding.
Give them access to one another. Make knowledge easy to share. Preserve institutional memory. Give people tools that expand what they can do. Avoid creating organisational boundaries faster than you need them. Most importantly, give good people reasons to stay. Because talent density sustained over time is where the real advantage lies.
A group of exceptional people who have spent five years learning how one another thinks possesses something that cannot simply be recruited from the market. Relationships have formed. Context has accumulated. Failed experiments are remembered. People know who to call. The organisation begins carrying intelligence in the connections between its people.
And as the organisation becomes more capable, follow the problem until you find the constraint.
Do not expand because expansion sounds impressive. Do not add services because competitors have them. Do not become multidisciplinary because being multidisciplinarity is in fashion.
Ask what is preventing you from solving the problem more completely.
Maybe the constraint is technology. Maybe it is distribution. Maybe it is capital, design, regulation, infrastructure or talent. Sometimes you should build the missing capability yourself. Sometimes you should partner. Sometimes you should leave it alone. But let the problem, rather than the category you started in, determine where you look next.
Finally, use your own products.
Anthropic offers an unusually literal version of this principle because Claude is becoming a tool used to build Claude. But almost every company can create a smaller version of the same loop. Use what you make internally. Become your own demanding customer. Allow the weaknesses your team encounters to inform what gets built next.
The ideal is a company in which the product improves the organisation, and the improved organisation produces a better product.
None of these ideas require billions of dollars.
You can choose a deeper problem with five people. You can make excellent work prestigious with ten. You can give people agency before you have managers. You can hire one person who raises the standard of everyone around them. You can document what your team learns. You can follow constraints rather than opportunities. You can build tools that make the people already inside the company more capable. The scale is different, but the underlying question is the same.
Founders spend enormous amounts of time designing the thing their company sells. The product gets a roadmap, a strategy, a design system, metrics and endless iteration.
Perhaps we should apply the same attention to the company itself.
Because over a long enough period, the organisation may be the most consequential product you build.
