Vector 1
AI Strategy & Economics

From Idea to Artifact: Why AI Is the Best Thing That Ever Happened to Polymaths, Divergent Thinkers and Builders

Brian J. Curry
Brian J. Curry
October 18, 2026 · 20 min read

For people who think in connections and learn by making: the cost of turning an idea into a working first version is falling fast, and the controlled evidence says who gains most. A model, the data, code you can run, and the places the argument fails.

Every creative project has the same hard part. Not the idea, and usually not the polish. The hard part is the first working version: the prototype that runs, the draft that holds together, the mockup a customer can react to. Going from nothing to something is where most ideas die, because it demands a stack of skills that rarely live in one head. A product idea needs code, design, copy, a pricing guess and a way to reach users. A research idea needs the literature, the method, the statistics and the write-up.

For a century the answer was specialization and teams. Each person owned one layer of the stack, and the idea moved between them through meetings, handoffs and budgets. That system works, but it taxes a certain kind of person: the polymath with five half-finished projects across three fields, the divergent thinker who sees a connection nobody asked for, the builder who learns by shipping. Their ideas cross boundaries, and every boundary crossed meant another specialist to recruit or another skill to spend a year learning.

Generative AI cuts that tax. A capable model gives you a competent first pass in statistics, front-end code, contract language or 3D modeling, in minutes. Depth still matters, and the model’s first pass is not expert work. But the fixed cost of an attempt in an unfamiliar field has dropped, and the people who generate the most cross-field ideas are the ones who benefit most when attempts get cheap.

This piece makes that case with evidence. It starts with a simple model of 0 to 1 as a series of attempts. It then walks through the controlled studies on creativity and idea generation, the older research on why range produces outsized work, and the data on people building alone. It ends with a simulation you can run and an honest account of where the thesis breaks.

A model of 0 to 1: attempts, not genius

The first working version of anything is rarely the one that succeeds. Founders, researchers and artists describe the same pattern: many attempts, most of them dead ends, and the occasional hit. So treat 0 to 1 as a numbers game. If each attempt has probability p of becoming something real, the chance of at least one hit in n attempts is:

The number of attempts you can afford is your time budget B divided by the cost of one attempt. This is where range comes in. An attempt that sits inside one field costs a base amount c₀. Every unfamiliar field it touches adds a crossing cost d: the hours spent learning enough of that field, or finding and paying someone who already knows it. For an idea that crosses m unfamiliar fields, I propose:

This is a proposed model, not an established result, but it makes the mechanism visible. Take a builder with B = 400 hours (a quarter of evenings and weekends), c₀ = 20 hours, and ideas that cross m = 2 fields they don’t know. Before capable AI, assume each crossing costs d = 80 hours. One attempt costs 180 hours, so they get about 2.2 attempts. With p = 0.05, their chance of a hit that quarter is 1 − 0.95².22 ≈ 10.8%.

Now let AI cut the crossing cost to d = 10 hours: enough to get a competent first pass at the unfamiliar parts, not to master them. An attempt now costs 40 hours, the builder gets 10 attempts, and the chance of a hit rises to 1 − 0.95¹⁰ ≈ 40.1%. Same person, same ideas, nearly four times the odds.

Compare a specialist who stays inside one field (m = 0) with a lower per-attempt hit rate of p = 0.03, since single-field ideas compete in a more crowded space. They get 20 attempts and a 45.6% chance either way, because the crossing cost never applied to them. Before AI, the specialist’s odds were about 4.2 times the generalist’s. After, the ratio is about 1.1.

AI lowers c₀ for the specialist too, and the toy numbers are assumptions you should replace with your own. The structural point survives any reasonable choice of numbers: a cheaper crossing cost helps in proportion to how many crossings your ideas need. The people with the most boundary-crossing ideas gain the most.

What the experiments say about AI and creative work

The model only matters if AI actually raises the quality of creative first attempts. Four controlled studies say it does, with a pattern that favors people working outside their comfort zone.

Start with idea generation. In a Wharton study, Karan Girotra, Lennart Meincke, Christian Terwiesch and Karl Ulrich pitted GPT-4 against MBA students at generating new consumer product ideas, then measured purchase intent with real consumers. GPT-4 produced 200 ideas in roughly 15 minutes, where one person produces about five in the same window. Its ideas also scored higher: an average purchase probability of 47% against 40% for the students, rising to 49% when the model was shown examples of good ideas first. Most of the best ideas in the pooled sample came from the model. For a builder, that means the blank page is no longer the bottleneck.

Next, creative quality. Anil Doshi and Oliver Hauser ran an experiment published in Science Advances in 2024 in which 300 people wrote short stories, some with access to AI-generated story ideas, and 600 evaluators rated them. Access to up to five AI ideas raised rated novelty by 8.1%. The gains were largest for the writers who scored lowest on a baseline creativity test: their novelty rose 10.7% and usefulness 11.5%. The same study found the cost we will return to later. Stories written with AI help were more similar to each other.

Third, consulting work at Boston Consulting Group. Fabrizio Dell’Acqua and colleagues gave 758 consultants a realistic task: develop new product concepts for a footwear company, then plan how to bring them to market. With GPT-4, consultants completed 12.2% more tasks, 25.1% faster, at more than 40% higher quality. Below-average performers improved 43%, above-average performers 17%. The tool closed most of the gap between weaker and stronger people, which is exactly what someone working outside their home field needs.

The fourth study is the most direct test of the polymath thesis. In 2024, 776 professionals at Procter & Gamble worked on real product innovation challenges, randomly assigned to work alone or in two-person teams (one R&D, one commercial), with or without AI. Individuals with AI improved solution quality by 0.37 standard deviations, matching the 0.39 gain of teams with AI and the level of human teams without it. They did it in 16.4% less time.

Two details matter for this argument. Without AI, R&D people proposed technical solutions and commercial people proposed commercial ones; with AI, both produced balanced solutions regardless of background. And employees working outside their core job, alone with AI, reached the performance of teams that included a specialist. That is the crossing cost d from the model, measured in the field and shown to fall. Teams with AI were also 9.2 percentage points more likely to produce a top-decile solution, against a control rate of 5.8%: roughly three times the chance of a standout idea.

One caution on reading these together. Each study uses a different task, measure and population, and two of them (Girotra and BCG) involve tasks chosen to suit the models of 2023. They agree on direction, not on a single effect size.

Why range was always the source of new things

If cheap crossings only let generalists do ordinary work faster, this would be a productivity story. It is a creativity story because the evidence has long said that unusual combinations produce the most important work, and that the people who make them tend to have wide interests.

Brian Uzzi, Satyam Mukherjee, Michael Stringer and Ben Jones analyzed 17.9 million scientific papers across every field, published in Science in 2013. They measured how typical each paper’s combination of cited sources was. The highest-impact papers were mostly grounded in conventional combinations of prior work, with an intrusion of unusual pairings on top. Papers with that profile were twice as likely to be highly cited. In the language of the model, cross-field attempts carry a higher p for a hit, which is why the crossing cost was always worth paying for the few people who could pay it.

The people who make those combinations have a recognizable profile. Robert Root-Bernstein and colleagues compared the avocations of Nobel laureates in science with those of other scientists and the public, reported in 2008 in the Journal of Psychology of Science and Technology. Laureates were about 2.85 times as likely as members of the scientific society Sigma Xi, and about 2.69 times as likely as the general public, to have a serious arts or crafts avocation. Against members of the Royal Society and the US National Academy of Sciences, the ratio was still about 1.6. The most decorated scientists are disproportionately people who paint, play, write or build things outside their field.

Why, then, did the world drift toward specialists? Ben Jones gave the answer in a paper published in the Review of Economic Studies in 2009. As knowledge accumulates, the amount an innovator must learn to reach the frontier grows. His model predicts, and his data on inventors support, rising education, narrower specialization and more teamwork as fields deepen. He called it the death of the Renaissance man. The Renaissance person did not stop being valuable. Reaching the frontier of a second or third field simply became too expensive for one lifetime.

Put those three findings together and the case is direct. Breadth produces the combinations that matter most. The people who carry it are overrepresented among the top creators. And the reason the economy stopped rewarding them was cost, specifically the cost of getting to working competence in a field you were not trained in. That is the cost AI is now cutting.

The doer’s shortcut: shipping outside your field

Ideas are cheap; working artifacts are not. The builder’s advantage has always been the willingness to make something real before it is fully understood. What held builders back was the stack of skills a real artifact requires. Here the evidence is about execution, and it points the same way.

Code is the clearest case, because so many first versions are software. In a 2023 controlled experiment by Sida Peng and colleagues at Microsoft Research and GitHub, developers asked to build an HTTP server in JavaScript finished 55.8% faster with GitHub Copilot than without it. The authors noted that the heterogeneous effects suggest AI pair programmers can help people transition into software work. For someone whose main field is design, biology or finance, that transition is the 0 to 1 step.

The largest field study of AI at work shows the same skill-compression pattern. Erik Brynjolfsson, Danielle Li and Lindsey Raymond followed 5,179 customer support agents as an AI assistant rolled out, published in the Quarterly Journal of Economics in 2025. Issues resolved per hour rose 14% on average and 34% for novice and low-skilled workers, with minimal effect on the most experienced. The tool spread the tacit knowledge of top performers to everyone else. A polymath entering a new field is, by definition, a novice there, and novices are who these tools help most.

You can see the result in how companies get started. Carta’s Solo Founders Report, published in December 2025, found that the share of new startups on its platform founded by a single person rose from 23.7% in 2019 to 36.3% in the first half of 2025. In March 2025, Y Combinator’s Garry Tan said that for about a quarter of its current batch, 95% of the code was written by AI, and that companies were reaching up to $10 million in revenue with fewer than 10 people. Carta’s data covers only companies on its platform and shows correlation, not cause, and Tan’s figures are his own characterization of YC companies. But both fit the model: when crossing costs fall, one person can cover more of the stack.

The practical shift is in what a first attempt looks like. A builder used to write a spec and look for a cofounder. Now the first attempt can be the thing itself: a running prototype, a landing page, a data analysis, a 3D-printable part. Feedback arrives on the artifact rather than the pitch, and each loop takes days instead of months.

Divergent thinkers: more ideas was never the problem

People who think divergently rarely lack ideas. Their bottleneck sits on either side of the idea: choosing which one to pursue, and then carrying it through the dull middle where the novelty has worn off and the work is administrative. Many people with ADHD and other neurodivergent profiles describe exactly this split, quick to connect and slow to start or finish.

AI helps on both sides, in different ways. On the selection side, a model is a cheap sparring partner. You can dump twenty half-formed ideas, ask for the strongest objection to each, and get a ranked shortlist in minutes. The Wharton result above showed a model can also generate and rate ideas at scale, which makes it useful as a filter for a person who already has too many. On the execution side, the model takes over the steps that drain attention: breaking a project into tasks, drafting the boilerplate email, writing the test harness, turning notes into a structured outline. The interesting part stays with the person; the friction moves to the machine.

The direct evidence here is thinner than for productivity, and I want to be plain about that. The best official data point is the UK Department for Business and Trade’s evaluation of Microsoft 365 Copilot, published in August 2025, covering a pilot that ran from October to December 2024. Neurodiverse respondents were more likely than other respondents to recommend the tool, a difference significant at the 95% level, and somewhat more likely to be satisfied, though that result only cleared the weaker 90% level. One interviewee said it had leveled the playing field for neurodiverse colleagues. The data is self-reported, the diary study had a 32% response rate, and the report did not find robust evidence of productivity gains overall. Treat it as a signal worth testing, not a settled finding.

The Doshi and Hauser result adds a second, quieter signal. The writers who gained most from AI story ideas were the ones who scored lowest on a baseline divergent-thinking test. Read alongside the other studies, the tool appears to help most where a person is weakest, whether that weakness is generating options, evaluating them or executing on them. For a divergent thinker, the weak link is usually execution. That is now the cheapest part to buy.

Run the model yourself

The closed-form numbers above assume fractional attempts and fixed costs. Real attempts come in whole units and vary in size: some prototypes take a weekend, others a month. The script below runs the same model as a Monte Carlo simulation with noisy attempt costs, for a specialist and a generalist, across a range of crossing costs. It needs only NumPy.

Monte Carlo of the 0-to-1 attempts model.

n = B / (c0 + m * d), P(hit) = 1 - (1 - p)^n
Attempt costs are noisy (lognormal), so the number of attempts per quarter varies.
"""
import numpy as np

RNG = np.random.default_rng(42)
B = 400            # hours available per quarter
C0 = 20            # base hours per attempt inside a familiar field
SIGMA = 0.5        # spread of attempt cost (lognormal sigma)
TRIALS = 100_000   # simulated quarters per scenario

PROFILES = {
    # name: (unfamiliar fields crossed per idea m, hit probability per attempt p)
    "specialist": (0, 0.03),
    "generalist": (2, 0.05),
}

def simulate(m: int, p: float, d: float) -> tuple[float, float]:
    """Return (mean attempts per quarter, share of quarters with at least one hit)."""
    mean_cost = C0 + m * d
    # lognormal with the requested mean: mu = ln(mean) - sigma^2 / 2
    mu = np.log(mean_cost) - SIGMA**2 / 2
    attempts = np.zeros(TRIALS, dtype=int)
    hit = np.zeros(TRIALS, dtype=bool)
    spent = np.zeros(TRIALS)
    active = np.ones(TRIALS, dtype=bool)
    while active.any():
        cost = RNG.lognormal(mu, SIGMA, size=TRIALS)
        fits = active & (spent + cost <= B)
        spent[fits] += cost[fits]
        attempts[fits] += 1
        hit[fits] |= RNG.random(TRIALS)[fits] < p
        active = fits
    return attempts.mean(), hit.mean()

if __name__ == "__main__":
    print(f"{'crossing cost d':>16} | {'specialist P(hit)':>17} | {'generalist P(hit)':>17} | ratio")
    for d in (160, 80, 40, 20, 10, 5):
        results = {name: simulate(m, p, d) for name, (m, p) in PROFILES.items()}
        s, g = results["specialist"][1], results["generalist"][1]
        print(f"{d:>14} h | {s:>17.1%} | {g:>17.1%} | {s / g:5.2f}")

Two choices are worth explaining. Attempt costs are lognormal because project durations are right-skewed: most attempts take about as long as planned, and a few take far longer. The mean is held at c₀ + m·d by setting μ = ln(mean) − σ²/2, so the noise changes the spread without changing the expected cost. The loop stops each simulated quarter at the first attempt that would overrun the 400-hour budget, which is why results come out slightly below the closed-form values.

At a crossing cost of 80 hours, close to the cost of picking up enough of an unfamiliar field the old way, the specialist’s odds of a hit in a quarter are about five times the generalist’s. At 10 hours the ratio is 1.17, and at 5 hours the generalist’s odds pass the specialist’s. The specialist’s line stays flat because nothing in their work crosses a boundary. Change p, m or B to match your own situation; the crossover point moves, but the shape does not.

Where it breaks

The case above is strong on direction and weak in four specific places. Anyone building a creative practice on AI should know them.

The first is sameness. In the Doshi and Hauser study, stories written with one AI idea were 10.7% more similar to each other than stories written without AI. Each writer got better while the collection got narrower. A 2024 study by Barrett Anderson, Jash Shah and Max Kreminski, presented at ACM Creativity and Cognition, found the same pattern in brainstorming: 36 participants using ChatGPT produced more ideas and more detailed ones, but ideas across users were less semantically distinct, and participants felt less responsible for what they produced. For a polymath, this is the risk that matters most. Your value comes from unusual combinations, and a model trained on the average of everything pulls toward the average. Use it to execute your strange idea, not to supply it.

The second is the jagged frontier. In the BCG study, consultants given a task that sat just outside the model’s capabilities were 19 percentage points less likely to reach the correct answer with AI than without it. The model sounded equally confident on both sides of the line. A newcomer to a field is the person least able to spot when the model has crossed that line, which is exactly the person this article says gains most. Cheap crossings come with a duty to verify.

The third is that experts do not always speed up. In METR’s randomized trial, published in July 2025, 16 experienced open-source developers working in their own large repositories took 19% longer on tasks when AI was allowed, while believing it had made them about 20% faster. METR’s follow-up, posted in February 2026, could not produce a new estimate because developers increasingly refused to work without AI in the control condition. The honest reading is that gains are largest where your skill is thinnest and smallest, possibly negative, where it is deepest. That is good news for 0 to 1 work in a new field and a warning against assuming the same gains in your home field.

The fourth is taste. When generating a version costs almost nothing, the scarce skill becomes choosing which version deserves the next hour. No study above measures that skill directly, and the Wharton result shows that a model can now outgenerate a room of MBA students. The person who wins is not the one who prompts most, but the one who can tell a good first version from a plausible one. That judgment still comes from depth somewhere, which is why the best generalists have always been deep in at least one thing.

A final note on the evidence base. Most of the controlled studies used 2023 and 2024 models on tasks lasting hours, not projects lasting months. Two of the headline numbers come from vendor or investor sources (Microsoft and GitHub for the Copilot study, Y Combinator for the batch figures). None of the studies follows a cohort of polymaths over years to see whether they actually build more successful things. That study does not exist yet.

The path forward

If the model is right, the move for a polymath or builder this quarter is to raise n, the number of real attempts, while protecting the two things AI cannot supply: unusual ideas and the judgment to pick among them.

Start by pricing your own crossing cost. Pick an idea you shelved because it needed a field you don’t know, and time how long it takes to get a working first version with AI help. If it took 80 hours to learn enough of a field before and now takes 10, your attempt budget for cross-field ideas has gone up by a factor you can actually measure. Run the simulation with your own B, m and p and see where your crossover sits.

Ship artifacts, not plans. The P&G result says one person with AI can match a two-person team on solution quality. Use that to put a working prototype, a dataset, a draft chapter or a printable part in front of real people every few weeks. A quarter that ends with ten tested attempts teaches more than one that ends with one polished deck.

Keep the strange part yours. Write down your idea before you open the model, then use AI to build it, critique it and find what breaks. Asking the model for ideas first is how you end up with the same idea as everyone else. If you do brainstorm with it, ask for the options it considers least likely and push from there.

Verify at the boundary. Every time you cross into an unfamiliar field, find one way to check the model’s work that doesn’t depend on the model: a domain expert’s 30-minute review, a test suite, a known benchmark result, a primary source. The jagged frontier is real, and newcomers are the ones who can’t see it.

And go deep in one place. The generalists who benefit most will be the ones with a home field deep enough to supply taste. Range without depth produces many plausible attempts and no way to tell them apart. Range built on depth is what the Nobel laureates in Root-Bernstein’s data had, and AI just made the range part cheaper.

About the Author

Brian James Curry is a Kansas City AI researcher and data scientist, the founder of Vector1 Research, an independent applied AI lab, and of KC AI Lab, which began in 2016 as the Kansas City Machine Learning Group and grew to more than 1,500 members. He has more than twenty years of experience leading analytics, data science and machine learning at enterprises including Hallmark, Garmin, Koch Industries, Tractor Supply Company and Vail Resorts. His open-source work, including Papilon, PyCausalSim, MeaningFlow, Daedalus and tincast, is on GitHub. He grew up in a small Kansas farming town and attended Kansas State University. He writes about agentic AI, causal inference and the economics of AI on Medium and Substack. Contact: brian@vector1.ai.

Sources

Sources

  1. Brynjolfsson, E., Li, D. and Raymond, L. Generative AI at Work. NBER Working Paper 31161; Quarterly Journal of Economics 140(2), 2025.
  2. Dell’Acqua, F. et al. Navigating the Jagged Technological Frontier. Harvard Business School Working Paper, 2023.
  3. Dell’Acqua, F., Ayoubi, C., Lifshitz, H., Sadun, R., Mollick, E. et al. The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise. NBER Working Paper 33641, 2025. Full paper (PDF).
  4. Doshi, A. and Hauser, O. Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances 10(28), 2024. ScienceDaily summary.
  5. Girotra, K., Meincke, L., Terwiesch, C. and Ulrich, K. Ideas are Dimes a Dozen. Wharton working paper, 2023. Knowledge at Wharton summary.
  6. Uzzi, B., Mukherjee, S., Stringer, M. and Jones, B. Atypical Combinations and Scientific Impact. Science 342, 2013.
  7. Root-Bernstein, R. et al. Arts Foster Scientific Success: Avocations of Nobel, National Academy, Royal Society, and Sigma Xi Members. Journal of Psychology of Science and Technology 1(2), 2008.
  8. Jones, B. F. The Burden of Knowledge and the ‘Death of the Renaissance Man’: Is Innovation Getting Harder? NBER Working Paper 11360; Review of Economic Studies 76(1), 2009.
  9. Peng, S., Kalliamvakou, E., Cihon, P. and Demirer, M. The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. Microsoft Research, 2023.
  10. Carta. Solo Founders Report 2025. December 2025.
  11. CNBC via NBC New York. Y Combinator startups are fastest growing, most profitable in fund history because of AI. March 2025.
  12. UK Department for Business and Trade. Microsoft 365 Copilot Evaluation. August 2025.
  13. METR. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, July 2025. Summary by Simon Willison. METR uplift update, February 2026.
  14. Anderson, B., Shah, J. and Kreminski, M. Homogenization Effects of Large Language Models on Human Creative Ideation. ACM Creativity and Cognition, 2

Related research