Well-Rounded Was a Compliment Once
It is Father’s Day as I write this, and it has me thinking about what we train the next generation for. So bear with me, I will get there.
Start with two financial analysts. Both are good. Both have spent years learning to build a model, read a filing, and tell a real signal from a distracting one. In the world that existed until recently, both did fine. There is enough work and enough demand that two solid analysts can each make a living without ever thinking about the other.
Now give each of them an AI agent built from their own knowledge and judgment. One teaches an agent how she works through a valuation. The other teaches an agent how he works through the same kind of problem.
The first loss is the one people reach for first. For their whole careers, these analysts were paid for the same knowledge over and over, once per deal, a thousand times across a thousand cases. The moment the work is teachable to an agent, the repetition stops paying. You hand the knowledge to the agent, and the agent takes the cases that look like the thousands before them. But the next case is never quite the same. Each one arrives with a wrinkle the agent has not seen, and someone has to work the wrinkle and feed it back. So the teaching is not one-and-done. It is continuous, and what you are paid for quietly shifts from repeating what you know to handling the part of each case that is genuinely new. For now that residual is large. As the agent absorbs more, it shrinks, and how fast it shrinks depends on how cheaply the work can be simulated, which is a thread I will pick up later. Either way, the job is no longer doing the analysis. It is teaching the thing that does the analysis, at the edge of what it cannot yet do.
But now look at the two agents side by side. If we are honest, most of what these analysts know is the same. They read the same references, internalized the same frameworks, learned the same hard lessons in roughly the same order. So the two agents come out nearly identical. Same strengths and same blind spots. When the agents are replicas, one of them is redundant. Not because either analyst got worse, but because the second copy taught the machine nothing the first had not already taught it. The first loss changed their job. This second one decides whether the new job is even theirs to have, and it is the one I want to follow.
The same is true for a tax attorney, a radiologist, and a translator. Anywhere the expertise is real but largely shared and largely written down, the work becomes teaching an agent, and then two people teaching the same thing collapse into one.
The risk of being different has always been real. If you go your own way, you might be wrong. But the risk of being the same just went up a lot, and most people have not repriced it. If you know exactly what everyone else knows, your incremental value to the world is close to zero. The crowd used to be a safe place to stand. It was a hedge. Now it is the one position that is guaranteed to be copied and competed away, because common knowledge is precisely what a model trained on the commons already has.
The 99 percent
In 2015, researchers at UC Davis and the University of Iowa ran a study on medical image reading. They took a group of subjects with no medical training at all, sat them in front of a screen, and showed them magnified biopsy slides, malignant tissue versus benign. Every time a subject got one right, a small reward. Slide after slide, day after day, until the learning curve flattened.
It worked. After training, a single subject could tell malignant from benign about 85 percent of the time. Respectable, roughly what you would want from someone learning to read pathology.
Then they did something clever. Instead of trusting one reader, they showed the same slide to four of them and took the majority vote. The group hit 99 percent. Higher than any individual in it.
The subjects were pigeons.
Four birds, voting, beat the best single bird and landed near the ceiling of the task. Not because any pigeon was a genius. The obvious reading is “more birds, better answer.” That is not it. The vote only works because the birds make different mistakes. When one bird’s blind spot fires on a slide, the other three are looking at it differently and outvote the error. If all four had identical strengths and identical weaknesses, the vote would buy you nothing. Four copies of the same bird is just one bird. The entire gain comes from the birds being uncorrelated.
I first heard this told this way by the mathematician Hannah Fry, who lets you go on assuming you are hearing about medical students right up until the word pigeon does its work.
A crowd can do the same thing. In 1906 the statistician Francis Galton watched fairgoers guess the weight of an ox. The individual guesses were wild, but the middle of them landed within a pound of the true 1,198. People quote this as proof that crowds are wise. They skip the condition: the errors only cancel if the guesses are independent and varied. Let everyone see everyone else’s guess first and independence collapses, and the crowd gets dumber, not smarter.
Hold that thought next to the world we are building. A market full of AI agents all fine-tuned from the same handful of foundation models is a crowd that already peeked at each other’s answers. Their errors line up. They are the pigeons that all share a blind spot. Voting across them gains you nothing, and being one more of them gains you nothing either.
So far this is the small case for being different: your mistakes and mine do not line up, so a group of us is steadier than the best one of us. Useful, but defensive. The bigger case is composition. Genuinely different strengths combine into things neither person could build alone. Put a marketer next to an engineer, and you do not get a mediocre generalist. When they actually gel, you get a company. The pigeons show why difference is robust. Composition is why difference is creative, and the creative part is where the real money is.
It comes with a price the pigeons do not pay. You can average four birds for free; they do not have to get along. People have to fit. The marketer and the engineer have to share enough language and trust to combine, or the difference turns into friction instead of a company. Composition is rarer than cancellation precisely because it asks for that fit.
So the goal, whether you are a person or a team, is to be a useful member of the ensemble rather than a redundant approximation of it. To be differently right, not just different. Being randomly contrarian is its own trap. An ensemble of noise is still noise. What has value is a real signal that the others do not have.
The job moves to the edge
If teaching the agent is the new job, the next question is what you teach it once the obvious cases are handled. The answer is that the human work moves to where the knowledge runs out. The agent handles the settled cases, the ones that have been seen and written down a thousand times. You handle the case, it fumbles, the strange one nobody has dealt with yet, and then you fold what you learned back in so the agent is better next time. You stop doing the work and start improving the thing that does the work. You live one step up, at the edge where the map ends.
Being different stops being a slogan here and becomes the actual job description. The esoteric case is the uncorrelated signal. The combination nobody has tried is the frontier. The tactics below are not decoration; they are how you get to that edge and stay useful once you are there.
There is a catch here: If improving the agent is valuable work, can that be automated too? In part, yes. What protects the edge is not that machines are weak there; it is that the edge is made of experience that does not exist yet, and you cannot learn from data you do not have. New signal comes from new lived experience, the experiment actually run, the situation actually survives. The only way to manufacture that without living it is to simulate reality faithfully enough to stand in for it. Where you can do that cheaply, protein structures, chip layouts, anything with a tight, checkable loop, the machine can generate its own experience and the human edge there erodes fast. Where you cannot, which is most of what humans actually do, reality stays the only reliable oracle, and it is slow and expensive to ask.
None of this makes the edge eternal. You can automate the work, then automate the improvement of the work, and if every level, including the making of genuinely new experience, becomes cheap to simulate, this whole argument has a horizon. But that horizon sits a long way off for most real fields, and even short of it, there is one more thing the edge does to you: it keeps moving. The strange case you handle today becomes known tomorrow, the agent absorbs it, and the frontier slides forward. So being different is not a place you arrive at and rest. It is a practice. The combination that is rare this year is training data, the next, and the work is to keep walking the edge outward.
How do you actually become different?
This is where most advice is bad. The usual answer points inward: find your passion, find your innate genius. I think that is mostly wrong, or at least incomplete. Difference is not something you find sitting inside you. It is something you construct, and most of the raw material is around you, not in you.
A few things that have actually worked, for me and for people I have watched up close.
Find the thing that is easy for you and hard for others, but find it by watching other people, not yourself. What comes easy to you is invisible to you. You cannot see your own water. You assume everyone knows the shortcut you know. So the signal is not introspection; it is the friction other people hit with things you breeze through. The thing colleagues keep asking you specifically to do. The explanation that makes a room go quiet because it had not occurred to them. Track what people come to you for, and you will find your edge faster than any amount of soul searching.
Stack two or three things you are merely good at, and pick them so they multiply. Scott Adams, who drew Dilbert, put this well. He was not the best artist, but he was better than most. He was not the funniest person alive, but he was funnier than most. Add a few years in a corporate office, and the combination produced a comic strip almost no one else could have made. Being in the top quarter of three things is achievable. Their intersection is rare to the point of being yours alone. But pick the pieces so they reframe each other, not at random. A mortgage broker who spent years as a director at real estate firms sees the whole transaction in a way other brokers cannot, because the two halves talk to each other. That is the difference between a real combination and a longer resume. And it is the thing an agent struggles to hold, because your specific intersection was never written down anywhere to be trained on.
Do the opposite of best practices on purpose. If something is genuinely a best practice, the agent already has it. The standard path produces standard people, because everyone walking it absorbs the same blind spots. So it is worth asking what happens if you learn the thing backward, or start where everyone else finishes. The point is not to be different for its own sake. It is that the well-trodden route is exactly the route that has been fully captured already, and the untrodden one is where signal nobody else has is still lying around.
Audit your unfair advantages of place, not just talent. The community you are part of, the people who will take your call, the data you happen to have access to, the specific rooms you can walk into. These are differentiators you did not earn through talent and that a competitor with your exact skill set still cannot copy. Most people optimize the one input they cannot change, their innate ability, and ignore the inputs they can move toward. Make a list of what is available to you that is not available to people otherwise just like you. That list is a map.
Go collect experiences that do not exist in any training set. The failed project, the strange client, the year in a market nobody pays attention to. Lived experience is, almost by definition, the one dataset the model was not trained on. The way to make these affordable is to structure them so you win even if they fail. Take risks you can absorb, and pick the ones where the skills and relationships you build transfer regardless of the outcome. Then a failure still pays you in something proprietary.
There is a tension running through all of this. Combining things that have not been combined before means stepping off the map, and off the map you might just be lost. The discipline that keeps it from being reckless is depth. Combine domains where you actually have real signal in each piece, so the bridge rests on something solid at both ends. Difference grounded in mastery is an edge. Difference grounded in nothing is just noise walking back into the ensemble.
On strengths and weaknesses
The standard career advice is to shore up your weaknesses. In this frame, that advice is mostly backward. Every hour you spend turning a weakness into mediocrity is an hour spent moving toward the average, toward the thing the crowd and the agents already cover. It makes you more complete and less distinct, and completeness is exactly what gets commoditized first.
A lopsided profile is more useful than a well-rounded one. A person who is exceptional at one thing and merely fine at the rest is legible and rare. A person who is evenly good at everything is the easiest thing in the world to replicate, because “good at everything, great at nothing” is a fair description of a foundation model. Two well-rounded people are nearly interchangeable. Two lopsided people are almost certainly lopsided in different directions, and those different directions are the whole game. The spikes are where the uncorrelated signal lives. The roundness is where everyone overlaps.
This does not mean ignore your weaknesses entirely. Some weaknesses are not interesting quirks, they are liabilities that sink the whole boat. The brilliant specialist who cannot be in a room with another person never gets to deploy the specialty. So the rule is narrow. Raise a weakness only to the point where it stops vetoing your strength, and not one hour further. Fix weaknesses to par. Invest your strengths to the frontier.
Where this leaves you
For most of history, you could earn a decent living by being a competent, unremarkable instance of a known thing. That option is quietly disappearing. The arithmetic has changed underneath us. When the price of common knowledge falls to zero, being a clean copy of the common knowledge is not the safe choice anymore. It is the risky one. It just does not feel risky yet, which is exactly why so few people are moving.
I am not telling you to be brave. I am telling you the conservative position became the dangerous one, and the differentiated position became the conservative one, and the labels have not caught up. Look hard at what comes easy to you that is hard for others. Look at what is around you that is not around anyone else. Combine the things you have real depth in that no one has thought to combine. Build the one spike that no agent and no peer happens to share.
Which brings me back to where I started. Almost every instinct I have as a father pulls the other way. Do well in the standard subjects. Clear the standard milestones. Get on the known path. Be well-rounded. I have one daughter, I think she is wonderful, and the pull I feel is to make her the best at the things everyone competes over. For my whole life that was the safe bet. But the same repricing applies to her, and it applies harder, because the urge to smooth out her spikes is strongest exactly when I love the person I am smoothing. Making her a little more like everyone else, a little more of a champion at the known game, is the thing that would make her replaceable. The hardest place to let someone stay lopsided is when you love them and want them protected. That is also, now, the place where the old definition of protected has quietly flipped into its opposite.
So this is what I actually want for her, and for myself, and for you. Be one useful bird in the flock. Not a copy of the others.