IQ, genes, and practice: what actually predicts who becomes a strong developer

Every few months a thread appears that wants a single cause. Either great developers are born with a number, or anyone can become one if they grind 10,000 hours, or it is “in the genes” in the way people say that when they mean they have stopped thinking. I have spent a career reading developmental and cognitive work, and then watching how those papers get flattened into hiring folklore. The flattening is the interesting part. The papers are smaller than the folklore.

I am not going to tell you your SAT score is your career. I am also not going to tell you practice is a magic equalizer. Both of those sentences are popular because they let someone stop looking at the rest of the file: opportunity, feedback, the kind of practice, the kind of job, and the fact that “strong developer” is not one trait.

This is what I think the evidence, such as it is, actually licenses. The next essay in this pair can argue that most of what we call developer intelligence is training. Here I want the prediction question: if you had to bet on who becomes strong, what should you put money on, and what should you stop treating as destiny.

The outcome is not one thing

Psychometrics likes variables you can put on a page. Software work does not. The person who is strong at a compiler is not the same person who is strong at a messy product negotiation, and both get called senior. Studies that correlate “programming skill” with a test usually measure something like closing a set of puzzles or a CS1 exam. That is a slice. It is not on-call judgment, not taste, not the ability to say no to a rewrite.

When I hear “what predicts a strong developer,” I ask which strength. If you mean speed at well-specified algorithmic tasks, cognitive measures will show up. If you mean five years later, still shipping, still trusted, still learning a new stack without a personality crisis, the predictors look more like practice quality, feedback, and whether anyone let you own a system. Mixing those outcomes is how both the IQ camp and the grind camp get to be right in a thread and useless in a hiring meeting.

What IQ-type measures actually show

General cognitive ability — the statistical factor people shorten to IQ — correlates with performance in a lot of trained jobs. That is an old, replicated, unsexy finding. Programming is a trained job with a symbolic load. It would be surprising if there were no relationship between measures of reasoning/working memory and early programming performance. There is a relationship. It is not a destiny coefficient.

The typical pattern, across skilled work, is that cognitive measures predict training success better than they predict later expert performance once people have had years in the domain. The first semester of CS sorts on how quickly you can hold a state machine in your head. Year seven sorts on whether you have seen this failure before and whether you still care. Those are not the same sort. Using a puzzle test to hire a staff engineer is a category error. Using “I never needed the puzzles, therefore cognition is irrelevant” is the opposite error.

Working memory and spatial/abstract reasoning show up in some programming-education studies. So does prior math. So does whether the student has been allowed to be wrong in public. If you only quote the first, you are doing PR for a test vendor. If you only quote the last, you are doing PR for a bootcamp.

Genes: a word people use as a door

Heritability estimates for cognitive ability in adulthood are not zero. They are also not a hiring tool, and they are not a story about a “developer gene.” Heritability is a population statistic. It says something about variance in a place and time. It does not say your colleague was fated. It does not say practice is pointless. It does not license the comments that appear under these posts.

What genes-and-environment work actually pushes against is the slogan that the starting line is identical. People arrive with different ease at certain kinds of symbol-manipulation. They also arrive with different schools, different time, different permission to look stupid, different health, different English, different access to a machine. If you flatten that into “it’s genetic,” you have chosen the least actionable and most socially radioactive slice of a messy causal pile. If you flatten it into “it’s all practice,” you have chosen the most marketable slice.

I will not pretend we have a clean map from DNA to a GitHub graph. We do not. Anyone selling that map is selling a vibe.

Practice: the word that hides the mechanism

Deliberate practice research — Ericsson and the arguments that followed — is not “hours in the chair.” It is hours with a specific difficulty, feedback, and a goal just past the current skill. A lot of software “practice” is not that. It is tutorials that stay easy, tickets that stay familiar, and a year of Copilot that finishes the line you were about to learn to finish.

The 10,000-hour slogan was always a magazine object. The useful residue is: people who become strong usually spent a long time in contact with problems that could prove them wrong, with someone or something that told them they were wrong. That can look like a brutal CS program, a messy job, an open-source maintainer who is honest, a pairing culture. It does not look like a streak counter.

Practice also predicts because it is how opportunity shows up. The intern who got the scary service learned more than the intern who got the ticket queue. In 2026 the ticket queue is thinner. That is a prediction problem: the old practice path is closing. If you only hire people who already look strong, you are measuring a history of practice you did not provide.

What I would actually bet on

If I had to rank predictors for “this person will be trusted with a system in three years,” I would not start with an IQ score. I would start with evidence they have finished hard, ugly work and updated when it broke. That is practice you can see. I would look at whether they can explain a trade-off they lost. That is cognition in the wild, not a matrix. I would look at whether they still seek the problem that can embarrass them. That is not a gene test. It is a habit, and habits have causes that include personality, health, and whether anyone made the environment survivable.

If I had to rank predictors for “this person will pass a brutal algorithms screen next month,” I would put more weight on recent puzzle practice and on the cognitive slice those puzzles sample. That is a different bet. Companies pretend those bets are the same. They are not. A lot of the IQ-and-coding discourse is a fight about which bet the industry should run, dressed up as biology.

Interest and time-on-task still eat most of the variance people want to give to mystery. People who like this kind of problem stay in contact with it. Staying in contact is how you get the 3 a.m. pattern match that looks like talent. Talent is a word we use when we did not watch the hours. The hours are not equally available. That is not a footnote. It is most of the prediction once you leave the lab.

What I would stop doing

Stop using “high IQ” as a compliment that means “agrees with me quickly.” Stop using “anyone can” as kindness that means “your barriers were character.” Stop talking about genes in a hiring channel. You do not have the measurement, you do not have the right, and you will use it as a story about people you already wanted to exclude.

Start measuring the practice you can offer. Start noticing which screens select for recent leetcode and which select for ownership. Start telling the truth about the slice. Strong is many jobs. Predictors are many, smaller than the thread, and mostly visible in work rather than in a number someone got at sixteen. If your process cannot see the work, it will keep reaching for the number. That reach is a habit, not a science.

If you want a one-line bet: put your money on high-quality practice in a domain that can fail, plus enough cognitive horsepower to hold the state, plus a life that lets the person stay. Leave the rest to people selling books. The books will still be there after the next study. The person who got to own the ugly service will already be better. That is a prediction you can act on this quarter. A number from a teenage test is not.