Beyond the Nearest Peak

Cheap Exploration Changes What Leadership Owes

The first workable answer used to deserve more respect than it does now because alternatives were expensive. A second architecture, second strategy, or second market motion meant days or weeks of work before you knew whether the path had any headroom. Teams stayed on local maxima because leaving them required crossing real valleys: budget, coordination, political capital, broken prototypes, and the shame of throwing away work. LLMs change the cost of inspecting those valleys. They let us sketch alternatives, expose assumptions, compare trade-offs, and attack weak paths before the organization commits depth. They also help after commitment by turning observed failures into tests, comparing fixes, summarizing messy evidence, and making small corrections cheaper to inspect. That does not make the model wise. It changes what is cheap.

The useful distinction is search versus compounding. Search chooses where depth should land. Compounding is what repeated correction does after you commit. Leadership work shifts toward keeping those motions separate enough to manage: search broadly before the first idea hardens, deepen only where evidence earns it, and look up again when the current path stops improving.

Search Chooses. Compounding Climbs.

A landscape is the option space: nearby hills, distant peaks, and valleys between them. In strategy or design work, those hills are possible architectures, operating models, product bets, or market motions. Some are close because the organization already has tools, habits, customer expectations, political agreements, and scars around them. Others may have more headroom, but they start out looking worse because no one has climbed them yet.

A learning curve is what happens after the choice. Costs fall, quality rises, and hidden constraints become legible because people keep correcting errors in the same direction. An S-curve is the shape of those returns: slow while the system is being understood, steep while improvements reinforce one another, and flat when the architecture runs out of headroom. The landscape asks where progress might be possible. The curve asks whether repeated effort is still compounding.

Feedback connects the two. If each correction makes the system easier to operate, explain, change, or sell, keep climbing. If each gain requires more coordination for less return, if workarounds are becoming the architecture, or if customers are asking for something the current system resists, look up again.

Why Nearby Peaks Used to Win

In the pre-LLM era, moving across the landscape was punishingly expensive. Switching approaches meant throwing away days or weeks of work. Prototypes were slow to build and painful to discard. Because teams could not afford to explore many valleys, they rationally optimized for safety. They picked the nearest promising hill and defended it, even if a much better path might have existed nearby.

Escaping that local maximum required unusual breadth. You needed someone with enough domain depth, enough cross-domain memory, and enough metis — tacit, situational competence — to guess which valley was worth crossing. This was the hero model of design: reliance on the intuition of a few because empirical testing of many options was too slow.

The incumbent path still deserves respect. It has accumulated learning behind it. A new path may have a higher ceiling and still look worse at first because it has not earned the tools, habits, supply chain, stakeholder trust, and scar tissue that make the incumbent feel obvious. That initial inferiority is the valley.

LLMs Make More Curves Inspectable

LLMs do not replace intuition or deep skill. They change the inspection cost. Before, testing a different architecture or strategy required building enough of it to feel the pain. Now you can sample five divergent approaches in an afternoon, compare the assumptions, and find the failure modes before the concrete sets.

The output is shallow, and that is fine. Shallow breadth gives you more surfaces to inspect before you commit. You can prompt for ten approaches, scan them for viability, and discard nine. Depth still emerges after selection and feedback. The model accelerates the work of discovering what might be worth deepening; reality decides whether the deepening compounds.

LLMs can accelerate both sides of the search/compounding loop. While looking up, they sketch competing paths, expose assumptions, compare trade-offs, and attack weak branches. While climbing, they turn production failures into tests, compare explanations, sketch small refactors, and make repeated corrections cheaper to inspect. They are neither necessary nor sufficient. A disciplined team can climb and look up without them; a sloppy team can use them everywhere and learn nothing. At current costs, refusing to use them for inspection, comparison, and correction makes the loop slower than it needs to be.

They still do not supply the compounding. Production data, customer contact, failure analysis, process discipline, and thousands of small corrections come from contact with reality. The model can sketch a new operating model. It cannot pre-install the judgment that comes from operating it under stress. It can identify a hill. It cannot climb it for you.

Some Curves Need Time to Reveal Themselves

Solar PV makes this concrete. In the module-price series compiled by Our World in Data from IRENA, Nemet, and Farmer/Lafond, photovoltaic modules fall from roughly $128 per watt in 1975 to about $0.26 per watt in 2024. That collapse was not one decisive invention. It was better silicon, better cells, better inverters, better manufacturing, better installation, better financing, and better coordination. None of those improvements alone looked like the revolution. The revolution emerged because many small improvements compounded along a curve with room left to climb.

That example cuts both ways. The lesson is not to believe every new path deserves faith. Most do not. The lesson is that some curves only reveal their advantage after enough correction has been allowed to accumulate. Leaders need to inspect more candidate curves before committing, then protect the chosen curve long enough for real learning to show up.

Two failure modes follow. Climbing without renewed search becomes lock-in: the organization keeps improving a curve whose ceiling is already visible, treating every incremental gain as validation while the opportunity cost of staying grows. Search without sustained climbing becomes churn: the organization keeps changing branches before any learning has time to compound, producing a portfolio of promising beginnings and no accumulated advantage.

The leadership task is to alternate deliberately between the two: search broadly enough to choose a curve with real headroom, commit long enough for cumulative improvements to reinforce one another, and reopen the search when marginal gains flatten, constraints change, or a new branch becomes credible.

Use Tree Search Instead of Slot Machines

Too many people use an LLM like a slot machine. They toss in a prompt — “Write a strategy for entering the SMB market” — pull the lever, and accept whatever comes out. If it feels off, they tweak language around the edges instead of questioning the direction. That is still linear work. It just has a faster typist.

The better workflow is tree search: widen the candidate set, score the branches, prune aggressively, and reserve expensive commitment for survivors.

• Fan out. Do not ask for a plan. Ask for options. “Propose five distinct strategies for entering the SMB market: partnerships, self-serve product-led, inside sales, ecosystem integrations, and a hybrid. For each, outline benefits, risks, time-to-impact, and required org changes.”
• Audit. Read the critique. Apply your constraints. Kill the options that do not fit your runway, risk appetite, or brand.
• Scaffold. For the survivors, draft comparable scaffolds: high-level roadmaps, org charts, investment theses, or test plans for the one or two options that remain.
• Deepen. Take over where the work becomes expensive. Pressure-test with your team, model the P&L, surface execution risks, talk to customers, and negotiate trade-offs the model cannot see.
• Verify. Decide what evidence would prove the selected path is compounding: fewer repeated failures, faster rejection of weak paths, clearer reasons for the surviving path, or measurable improvement in the customer or operator workflow.

Use this as the operating discipline: search broadly enough that the first idea loses its monopoly, score fast enough that weak branches die before they become projects, and deepen few enough that contact with reality can still compound.

Do Less, Judge More

The practical adaptation is to invert the ratio of effort. Do less production work on the first plausible answer. Do more judgment work before and after commitment. Stop measuring value by the volume of activity. Measure it by the learning you create, the paths you reject, and the speed with which reality can correct the path you chose.

Good judgment is not heroic intuition. It is a discipline built on three commitments:

• Do not settle on the first workable option. A workable idea proves only that one path exists. Generate contrasting alternatives before choosing. Treat the earliest idea as a reference point, not a winner.
• Score options before deepening them. Depth is costly. Before investing, compare options against the same constraints: runway, risk, time-to-impact, differentiation, and feasibility. Let the scoring filter the field so depth lands where it matters.
• Keep cycle time visible and short. Track how long it takes to move from question to options, from options to selection, and from selection to a testable draft. Short cycles prevent overcommitment to weak paths and make learning harder to fake.

A portfolio lens helps because it keeps the allocation problem visible. Options are candidate investments. Scoring is diligence. Deepening is capital allocation. Pruning is cutting losses early before they compound. The aim is maintaining a healthy pipeline of alternatives and investing only in the ones that survive the filter.

Search chooses the curve. Compounding climbs it. Judgment decides when to look up again.

---

Appendices

Appendix A: Tunneling Across Rugged Landscapes

In Stuart Kauffman’s NK Model, a “rugged” fitness landscape traps agents on local maxima. To find a higher peak, an agent usually has to step “downhill” (worse performance) to cross a valley.

• Pre-LLM: Crossing the valley required expensive refactoring. The cost of the “downhill” move often exceeded the team’s political or temporal capital, trapping them on mediocre hills.
• Post-LLM: LLMs allow for “tunneling.” Because we can generate a fully formed alternative architecture in parallel, we can assess the potential of a distant peak without painstakingly traversing the valley of broken code in between.

Result: We no longer walk the landscape linearly; we sample it discontinuously.

Appendix B: The "Shallow-to-Deep" Protocol (Cognition)

The Shallow-to-Deep protocol is the cognitive version of tree search. It widens the branch factor early, applies a simple evaluation function, and prunes most of the space before committing depth to the remaining branches.

1. Explosion (shallow): Prompt for N divergent approaches; maximize variance, ignore syntax errors.
2. Selection (the filter): Apply the scoring function (cycle time, modularity, correctness); reject N−1 options immediately.
3. Deepening (the investment): Inject metis—tacit knowledge, constraints, edge cases—into the surviving option.
4. Verification: Run it. If it fails, do not debug linearly; loop back to step 1 with new constraints.

Key point: Spend deepening effort after selection, not before.

Appendix C: Generation Gets Cheap; Verification Does Not

The economic shift is simple: the cost of generation collapses, the cost of verification persists. Writing variants is cheap; reading, testing, and rejecting them still takes human time. The bottleneck moves from producing to judging.

• Old rationality: productivity was lines written because writing was scarce.
• New rationality: productivity is bad paths rejected because judgment is scarce.

David Deutsch’s lens applies: progress is the rate at which we correct errors. Faster generation only helps if we raise our rejection and correction rate to match.

Appendix D: Curve Selection vs. Curve Climbing

The body of the essay uses learning curves and landscapes together, but they answer different questions. The landscape asks where a better path might be. The curve asks what repeated correction can compound once you commit to a path.

Use the distinction as a diagnostic.

Keep climbing when:

• Repeated effort is making the system easier to operate, explain, change, or sell.
• The same class of failure is becoming rarer or cheaper to fix.
• New improvements reinforce earlier improvements instead of fighting them.
• The team can name the next bottleneck from production, customer contact, or repeated use.

Look up again when:

• Each gain requires more coordination for less return.
• Workarounds are becoming the architecture.
• Reviews keep catching the same class of problem.
• Customers are asking for something the current system resists.
• A new branch can now be inspected cheaply enough to compare honestly.

LLMs can accelerate both motions. They are neither necessary nor sufficient: a disciplined team can climb and look up without them, and a sloppy team can use them everywhere and learn nothing. But at current costs, refusing to use them for inspection, comparison, and correction is increasingly hard to defend. While climbing, they can help turn observed failures into tests, summarize messy evidence, compare candidate fixes, and make small corrections easier to inspect. While looking up, they can sketch alternatives, expose assumptions, and test whether another curve has enough headroom to deserve depth. They still do not replace contact with reality. Production data, customer contact, failure analysis, and repeated correction decide whether a curve compounds.

The mistake is treating either motion as virtue by itself. Climbing without search becomes lock-in. Search without climbing becomes churn.