The Explore-Exploit Dilemma: When to Try New Things and When to Double Down

TL;DR. Choosing between a known good option and an untested one is an explore-exploit problem, and the research points to one variable: how long you have left to use what you learn. In a laboratory task, people explored more when 6 choices lay ahead than when only 1 did. In US data, people born in 1957-64 held 5.6 jobs at ages 18-24 and 2.2 at ages 45-54, and a classic study found that job changes account for at least a third of wage growth in the first ten years of a career. Across 2,128 artists, 4,337 film directors and 20,040 scientists, “explore, then focus” was the only sequence linked to a higher chance of a hot streak starting. AI makes trials cheaper: in one working paper, developers given GitHub Copilot access increased their exposure to new programming languages by 21.79%. The CEO move: set a fixed exploration budget sized to your horizon and decide in advance what result would make you commit. The student move: treat every trial as a cheap experiment with one question, a cost cap and a stop date.

This piece applies the Personal Decision Stack to personal strategy. It sits next to optionality as a career strategy, about keeping doors open, and the antifragile individual, about gaining from shocks. The question here is narrower: when to try new things, and when to double down.

What is the explore-exploit dilemma, in plain terms?

It is the choice between a reward you know and an option you know less about that might be better. The formal version is called the multi-armed bandit problem, after slot machines. In a 2006 Nature study, Nathaniel Daw and colleagues asked 14 healthy subjects to choose repeatedly among four slot machines whose average payoffs drifted from trial to trial. Brain imaging showed the frontopolar cortex and intraparietal sulcus were more active during exploratory choices, while parts of the striatum and ventromedial prefrontal cortex were involved in value-based exploitative choices.

In management research, James March’s 1991 paper in Organization Science framed organizational learning as a tension between “the exploration of new possibilities and the exploitation of old certainties”. Its abstract warns that adaptive processes, because they refine exploitation faster than exploration, “are likely to become effective in the short run but self-destructive in the long run”. The same trap applies to one career: the skill that pays today gets the practice, so it crowds out the search for the skill that would pay later.

A 2015 synthesis in Decision by Katja Mehlhorn and colleagues traced the tradeoff across human and animal research and showed it has no fixed answer: it depends on definitions, on environmental, social and individual factors, and on time scale. The concept reached a general audience through the 2016 book Algorithms to Live By by Brian Christian and Tom Griffiths.

How is this different from the 37 percent rule?

The 37 percent rule belongs to a different problem, optimal stopping. In the classical secretary problem, candidates arrive one at a time, a rejected candidate cannot be recalled, and the only success is picking the single best. Thomas Ferguson’s lecture notes on optimal stopping show that, with many candidates, it is approximately optimal to pass over the first 36.8% (a share of 1/e) and then take the next candidate who beats everyone seen so far.

Explore-exploit differs: you choose repeatedly, collect rewards along the way, and can usually return to an earlier option. Splitting your learning hours each quarter between your core skill and new ones is a bandit problem, and that is the subject here.

Why does your time horizon decide most of the answer?

Because information only has value if there is time left to use it. Robert Wilson and colleagues (Journal of Experimental Psychology: General, 2014) summarise theory going back to John Gittins’ work on bandit problems in the 1970s: “information has real value for long horizons”. Their Horizon task tested this. Each game began with four forced-choice trials, after which participants made either a single free choice (horizon 1) or a sequence of 6 (horizon 6). With the longer horizon, the 30 participants in the main analysis showed both forms of exploration: a larger bonus for the less-known option (directed exploration) and noisier choices (random exploration). A parallel experiment found no difference between horizons of 6 and 11, consistent with an effect that switches on once any future exists to use the information.

Alison Gopnik (Philosophical Transactions of the Royal Society B, 2020) argues that the long human childhood is an evolved period of broad search “before the demands of goal-directed exploitation set in”, and notes that solutions to the dilemma share one feature: “the learner should begin by exploring and gradually converge on exploiting”.

At the other end of life, Rui Mata, Andreas Wilke and Uwe Czienskowski (Frontiers in Neuroscience, 2013) compared younger and older adults in two foraging tasks, including a Fishing task with 150 participants (75 younger, 75 older). They report “a trend toward reduced exploration with increased age”: older adults stayed longer in a patch before moving on. The authors weigh an adaptive explanation (less future left) against cognitive decline without settling which matters more.

The practical lesson: horizon is set per domain, not by birthday. A 50-year-old who expects to use a new analytics tool for 15 years has a long horizon in that domain. A 25-year-old who plans to leave a field next year has a short one inside it.

What does career data say about exploring early?

The Bureau of Labor Statistics follows the same people for decades in the National Longitudinal Survey of Youth 1979 (NLSY79): 9,964 men and women born in 1957-64, last interviewed in 2022-23 at ages 57 to 66. Table 1 lines up its latest figures with the newest Employee Tenure release.

Table 1. The career curve: job changes, early exits, earnings growth and tenure by age (US)

Age band	Jobs held in band (NLSY79)	Jobs per year of age (CEOtudent calculation)	Jobs started in band that ended within 1 year	Average annual growth in real hourly earnings	Median tenure with current employer, January 2026
18-24	5.6	0.80	60.7%	6.5%	1.5 years (ages 20-24)
25-34	4.5	0.45	41.7%	3.3%	3.0 years
35-44	2.9	0.29	25.3%	1.9%	4.7 years
45-54	2.2	0.22	21.4%	0.1%	7.0 years
55-58	1.3	0.33	23.8%	0.7%	9.6 years (ages 55-64)

Sources: U.S. Bureau of Labor Statistics, Number of Jobs, Labor Market Experience, Marital Status, and Health for Those Born 1957-1964 (NLSY79; released August 26, 2025, reissued July 21, 2026), highlights and Tables 2 and 5; U.S. Bureau of Labor Statistics, Employee Tenure in 2026 (September 24, 2026), Table 1. Jobs per year of age divides each band’s job count by the years of age in the band (7 for ages 18-24, 10 for the middle bands, 4 for ages 55-58). The tenure column is a 2026 snapshot of all US wage and salary workers, not the NLSY79 cohort, and its youngest and oldest bands differ. The BLS release text gives the 0.1% earnings growth figure for ages 45 to 55.

Three patterns stand out. This cohort held 12.9 jobs from ages 18 to 58, over 40 percent of them before age 25; per year of age, people aged 18-24 started about 3.6 times as many jobs as people aged 45-54. Early jobs end fast: 60.7% of jobs started at ages 18-24 ended within a year, against 21.4% at ages 45-54. And real hourly earnings grew 6.5% a year at ages 18-24 against 0.1% at ages 45-54.

A younger cohort shows the same shape. The NLSY97 release of May 2026 covers people born in 1980-84: they held 9.4 jobs from age 18 through age 38, 5.1 of them at ages 18 to 23, and 61 percent of jobs started at ages 18 to 24 ended in less than a year, against 32 percent of jobs started at ages 25 to 38. In the January 2026 tenure data, median tenure for workers aged 55 to 64 (9.6 years) was more than three times that of workers aged 25 to 34 (3.0 years).

Robert Topel and Michael Ward followed young men’s careers for up to 15 years using the Longitudinal Employee-Employer Data (NBER Working Paper 2649, 1988, later published in the Quarterly Journal of Economics). During the first 10 years in the labor market, “a typical young worker will work for seven employers”, about two-thirds of all the jobs he will hold, and wage gains at job changes “account for at least a third of early-career wage growth”. In their decomposition, job transitions accounted for 40 percent of cumulative log wage growth over the first ten years. Early job changing looks haphazard, they conclude, but it is how workers reach stable matches.

One caution: layoffs, relocations and family events sit inside these numbers too, and averages say nothing about any single move.

Does it pay to explore before you specialize?

Ofer Malamud used a natural experiment in British universities. English students choose one field before entering university; Scottish students are typically admitted to a broad faculty and study several fields in their first two years. If university mainly built field-specific skills, Scottish late specializers should switch fields more after graduating, because they give up less training. The data showed the opposite. In the preferred estimate, graduates with Scottish degrees were approximately 6 percentage points less likely to switch to an occupation unrelated to their field, against a Scottish switching rate of about 42 percent (NBER Working Paper 15522). Sampling fields helped students find a better match.

A companion paper priced the switch. Graduates who moved to an unrelated occupation took a first-year wage loss of around 7 percentage points. More switching occurred in England, where students specialize early, but “higher wage growth among those who switch eliminates the wage difference after several years” (NBER Working Paper 15943). Together: exploring before committing reduces mismatches, and correcting a mismatch later has a real but temporary cost for people with enough horizon left.

What comes right before a career hot streak?

Lu Liu, Dashun Wang and colleagues (Nature Communications, 2021) traced the work of 2,128 artists (over 800,000 images of artworks), 4,337 film directors (79,000 films) and 20,040 scientists, using deep learning and network methods to measure how varied each person’s output was over time. In all three domains, people tended to explore diverse styles or topics before a hot streak and became notably more focused once it began.

The key result is about sequence. Exploration not followed by focus was linked to a lower chance of a hot streak starting than the baseline, and so was exploitation not preceded by exploration. Exploration followed by exploitation was the only combination with a positive lift in the probability of a hot streak starting: 20.5% over the baseline for artists, 13.8% for directors and 19.2% for scientists. The authors suggest that exploration, a risky strategy that widens the range of outcomes, raises the chance of finding a promising idea, and exploitation then develops it.

The authors stress the evidence is correlational. It does not show that focusing causes success, but it gives a shape: the move that matters is the switch from exploring to focusing.

How does AI change the cost of exploring?

Exploration is expensive mainly because the first weeks in a new area are slow. Early evidence suggests generative AI shrinks that cost. Manuel Hoffmann and colleagues (Harvard Business School Working Paper 25-021, not peer reviewed; two of the five authors work at Microsoft and GitHub) used a program eligibility threshold in a regression discontinuity design to study free GitHub Copilot access among 50,032 top open-source developers. Eligible developers shifted toward coding, up 5.4 percentage points as a share of all activity, and away from project management, down 10 percentage points. They also explored more: they engaged with an additional 15 new repositories on average relative to ineligible peers and increased their cumulative exposure to new programming languages by 21.79%, while revisits to known repositories declined. The authors call it “an increase in exploration activities rather than exploitation”.

A second study suggests why. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied an AI assistant for customer-support agents. The published version (Quarterly Journal of Economics, 2025), covering 5,172 agents, reports that access raised issues resolved per hour by 15% on average, with less experienced and lower-skilled workers improving both the speed and quality of their output. In the 2023 working paper version, treated agents with two months of tenure performed as well as untreated agents with more than six months. For that curve the authors say they cannot distinguish learning from following the tool’s recommendations, but data from software outages, when the tool gave no suggestions, still showed gains over workers’ pre-AI baseline, which the authors read as evidence of durable learning.

The implication: AI lowers the cost per trial and of the novice phase, so the same budget buys more trials. It does not lengthen your horizon or tell you when to stop, which makes a stop rule more important.

What does the evidence say about when exploration pays?

Table 2. CEOtudent analysis of eleven studies and statistical releases: when exploration pays

Setting (source)	Sample	Verified finding	What it suggests about timing
Lab choice task (Wilson et al., 2014)	30 participants in main analysis	More directed and random exploration with 6 choices ahead than with 1; no difference between 6 and 11	Explore whenever there is a future in which to use the information
Childhood (Gopnik, 2020)	Theory and review of evidence	Solutions tend to begin with exploring and converge on exploiting	Broad search belongs at the start of any learning curve
Aging (Mata et al., 2013)	150 adults in the Fishing task	Trend toward reduced exploration with age	Exploration falls as the horizon shortens, possibly also through decline
Early career (Topel and Ward, 1988 working paper)	Young men followed up to 15 years	Seven employers in the first 10 years; job changes at least a third of early wage growth	Early job shopping is a main source of early wage growth
US cohort data (BLS NLSY79 and NLSY97)	9,964 people born 1957-64; about 9,000 born 1980-84	5.6 jobs at ages 18-24 vs 2.2 at 45-54; earnings growth 6.5% vs 0.1% a year	Most switching, and most earnings growth, happens early
University specialization (Malamud, 2009 and 2010 working papers)	British graduates	Late specializers about 6 points less likely to switch fields; switchers’ first-year wage gap closes after several years	Sampling before committing reduces mismatch
Creative careers (Liu et al., 2021)	26,505 careers	Only exploration followed by exploitation lifted the chance of a hot streak starting (13.8% to 20.5% over baseline)	The switch to focus is the move that matters
AI tools (Hoffmann et al., working paper; Brynjolfsson et al., 2025)	50,032 developers; 5,172 agents	21.79% more exposure to new languages; novices gain most from AI help	Cost per trial is falling; the horizon is not

Sources: as cited in each row; full references below. The timing column is CEOtudent editorial interpretation, not a finding of the studies.

Three conditions recur: exploration is worth most when the horizon is long, when a trial reveals real information about fit, and when a deliberate switch to focus follows. The tool below builds on them.

The CEOtudent explore-exploit allocation tool

This is a CEOtudent editorial framework, not a validated instrument, with no success rates. Rate any candidate skill, role, side project or tool on three questions:

1. Horizon remaining. How long could you use what this trial would teach you? Long means several years of use after reaching competence.
2. Reversibility. If it fails, can you undo it within weeks at a cost you can absorb?
3. Information per trial. Would one small, bounded trial produce a signal that changes your next decision?

Table 3. CEOtudent editorial framework: explore-exploit allocation tool

Horizon	Reversible?	Information per trial	Stance	Example
Long	Yes	High	Explore now with small trials; the main home of your exploration budget	Testing a few AI tools on real weekly tasks
Long	Yes	Low	Redesign the trial for faster feedback before spending on it	Turning a course with no project into a course plus a shipped mini-project
Long	No	High	Explore through a reversible proxy first	A short contract or side project before a career switch
Long	No	Low	Hold; buy information elsewhere and keep the option open	Talking to several practitioners before a multi-year degree
Short	Yes	High	Quick test only if it can pay back inside the horizon	A one-week trial of a workflow change
Short	Yes	Low	Exploit; treat it as curiosity, not strategy	Sampling a trending tool you will not use at work
Short	No	High	Exploit, unless kill criteria on the current option have fired	Changing employers shortly before a planned exit
Short	No	Low	Exploit firmly and double down	Finishing a nearly complete certification

Horizon decides whether exploring is worth it, information decides whether a trial is worth running, and reversibility decides how to run it. To choose which skill gets your next block of focused effort, use the four-factor skill decision matrix; this tool decides how much effort goes to trying versus deepening.

How big should your exploration budget be?

No study gives a correct ratio for an individual career, so this rule and its numbers are editorial starting points, not research-derived values.

1. Budget time first, money second. Choose an amount you could lose entirely without harming your core work.
2. Scale the share with your horizon in the domain. Editorial bands: long horizon (more than 5 years of expected use), 25-35% of discretionary development time; medium (2-5 years), 15-25%; short (under 2 years), 0-10%.
3. Split the budget into trials, each with one question, a cost cap and a stop date.
4. Review on a fixed date. Each trial ends in commit, extend once, or stop.
5. Graduate winners. A committed trial moves into core work and the budget refills with new trials.

Worked example (illustrative, not data). A product analyst has 8 hours a week for development and expects to work in analytics for many years, so she takes the lower end of the long-horizon band, 25%: 2 hours a week, or 26 hours over a 13-week quarter. She runs three trials of about 8.7 hours each: an AI agent workflow for weekly reporting, porting one recurring analysis from spreadsheets to Python, and a short customer-interview sprint. The other 6 hours a week, 78 hours a quarter, deepen her core skill. At review, the agent workflow saved time every week, so she commits. The Python port gave no clear signal because its question was vague, so she extends it once with a narrower question. The interview sprint showed nothing that would change her role, so she stops it.

The student side of the rule is a trial log: question, cost cap, stop date, result and decision.

The commit-or-stop test

Run these five questions at each trial’s stop date. They turn the explore-exploit logic into a decision made in advance, in the spirit of kill criteria.

1. Did the trial answer its question? If not, was the trial badly designed (extend once, narrower) or was the signal negative (stop)?
2. Would you choose this over your current best option if starting today? Novelty does not count; only expected value against your current best does.
3. Is your remaining horizon long enough to recover the ramp-up cost? Malamud’s switchers recovered their wage gap, but only over several years.
4. Are you ready to narrow? Committing means cutting other trials; the hot-streak data links gains to the switch to focus, not to piling up options.
5. Can you commit in stages? Prefer a reversible first commitment, such as a defined project or a part-time role.

Decision rule. Commit if questions 1, 2 and 3 are all yes. Extend once if question 1 is unclear because of the trial’s design and question 2 is still plausible. Stop if question 2 is no. Record the reason, and judge it by its reasoning rather than its outcome, as described in how to think in bets.

What this does not mean

• Not “always explore when young”. The career data describes averages, and the hot-streak research links gains to exploration followed by focus, not exploration alone.
• Not that exploration ends at a certain age. Horizon is set per domain.
• Not that job changes cause higher pay. People who move may differ from people who stay.
• General information only. This describes research findings and an editorial framework, not personalised career or financial advice.

Frequently asked questions

What is the explore-exploit dilemma?
The choice between a known option with a reliable payoff and a less-known option that might be better.

Is there an optimal ratio of exploring to exploiting?
Not for an individual career. Research says exploring is worth more when more time is left to use what you learn, so the share should shrink as your horizon shortens. The bands above are editorial starting points.

Is the 37 percent rule the answer to when to stop exploring?
Only for one-shot choices where rejected options cannot be recalled. Most career choices are repeated and allow a return to earlier options.

Is it ever too late to start exploring?
Horizon in the domain matters more than age. If you will use a new skill for many years, the case for trying it remains.

Does AI mean you should explore more?
Each trial now costs less, so the same budget buys more trials. AI does not tell you when to commit, so a written stop rule matters more.

How do you know when to double down?
When a trial has answered its question, beats your current best option on expected value, and you have enough horizon left to recover the ramp-up cost.

Sources

1. Daw ND, O’Doherty JP, Dayan P, Seymour B, Dolan RJ. Cortical substrates for exploratory decisions in humans. Nature 441(7095): 876-879, 2006 (PMC author manuscript).
2. March JG. Exploration and exploitation in organizational learning. Organization Science 2(1): 71-87, 1991 (abstract only, via RePEc/IDEAS).
3. Mehlhorn K, Newell BR, Todd PM, Lee MD, Morgan K, Braithwaite VA, Hausmann D, Fiedler K, Gonzalez C. Unpacking the exploration-exploitation tradeoff: a synthesis of human and animal literatures. Decision 2(3): 191-215, 2015.
4. Christian B, Griffiths T. Algorithms to Live By: The Computer Science of Human Decisions. Henry Holt and Company, 2016 (bibliographic record only, Open Library).
5. Ferguson TS. Optimal Stopping and Applications, Chapter 2: Finite Horizon Problems. Lecture notes, Department of Mathematics, University of California, Los Angeles.
6. Wilson RC, Geana A, White JM, Ludvig EA, Cohen JD. Humans use directed and random exploration to solve the explore-exploit dilemma. Journal of Experimental Psychology: General 143(6): 2074-2081, 2014 (PMC author manuscript).
7. Gopnik A. Childhood as a solution to explore-exploit tensions. Philosophical Transactions of the Royal Society B 375(1803): 20190502, 2020.
8. Mata R, Wilke A, Czienskowski U. Foraging across the life span: is there a reduction in exploration with aging? Frontiers in Neuroscience 7: 53, 2013.
9. U.S. Bureau of Labor Statistics. Number of Jobs, Labor Market Experience, Marital Status, and Health for Those Born 1957-1964 (NLSY79). News release USDL-25-1322, August 26, 2025, reissued July 21, 2026, with Tables 2 and 5.
10. U.S. Bureau of Labor Statistics. Labor Market Experience, Education, Partner Status, and Health for Those Born 1980-1984 (NLSY97). News release USDL-26-0684, May 5, 2026.
11. U.S. Bureau of Labor Statistics. Employee Tenure in 2026. News release USDL-26-1532, September 24, 2026, with Table 1.
12. Topel RH, Ward MP. Job Mobility and the Careers of Young Men. NBER Working Paper 2649, 1988; published in Quarterly Journal of Economics 107(2): 439-479, 1992. Figures cited from the working paper.
13. Malamud O. Discovering One’s Talent: Learning from Academic Specialization. NBER Working Paper 15522, 2009; published in ILR Review 64(2): 375-405, 2011. Figures cited from the working paper.
14. Malamud O. Breadth vs. Depth: The Timing of Specialization in Higher Education. NBER Working Paper 15943, 2010; published in Labour 24(4): 359-390, 2010. Figures cited from the working paper.
15. Liu L, Dehmamy N, Chown J, Giles CL, Wang D. Understanding the onset of hot streaks across artistic, cultural, and scientific careers. Nature Communications 12: 5392, 2021.
16. Hoffmann M, Boysel S, Nagle F, Peng S, Xu K. Generative AI and the Nature of Work. Harvard Business School Working Paper 25-021, 2024 (working paper).
17. Brynjolfsson E, Li D, Raymond LR. Generative AI at Work. Quarterly Journal of Economics 140(2): 889-942, 2025 (abstract only, via Crossref); and NBER Working Paper 31161, April 2023, revised November 2023.

Note on original analysis: Table 1 combines Bureau of Labor Statistics tables and adds a CEOtudent calculation (jobs per year of age). Table 2 is a CEOtudent analysis of eleven studies and releases. Table 3, the exploration budget rule and its bands, the worked example and the commit-or-stop test are a CEOtudent editorial framework with no measured success rates. Deliberately not used: the 37 percent rule applied to careers or partners without its assumptions; popular claims about how many jobs or careers a person has in a lifetime that could not be traced to a primary source; the 14% headline from the working paper version of Brynjolfsson, Li and Raymond, replaced by the published 15%; and anything from the full text of March (1991), which was not opened.

---

This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.

This post is also available in: