Alex Komoroske on Strategy and Complexity

Guest: Alex Komoroske — founder building a startup to reimagine the web for the AI era; formerly 13 years at Google (Search, DoubleClick, Chrome Open Web Platform team for 8 years, AR in Google Maps, bottom-up strategy toolkit); Head of Corporate Strategy at Stripe. Creator of the “Bits and Bobs” weekly reflection practice and the Compendium tool (17,000+ working notes).
Host: Lenny Rachitsky
Source: Lenny’s Podcast. Recorded ~2023–24.

Overview

Alex Komoroske presents a set of frameworks for navigating complexity in product and organisations: LLMs as magical duct tape (disruptive cost-structure shift), the gardener vs. builder mindset (farming for miracles), organisational kayfabe (systemic compounding optimism), the slime mold principle (coordination costs scale with n²), adjacent possible (small reachable actions + North Star), and strategy salons (emergent idea generation). The thread connecting all of them: top-down plans cannot outperform the energy put into them; emergence-based approaches can.

Key ideas

LLMs are magical duct tape. Distilled intuition of all society; operates at a cost between human and plain computing. Undermines the industry’s foundational assumption that “software is expensive to write, cheap to run.” Current failure mode: treating LLMs as oracles. Correct question: given magical duct tape, what can I build that was previously impossible?
Squishy computer. LLMs do roughly what you meant, not exactly what you said. 99% accuracy is still not viable if the 1% failure mode punches the user in the face. Product design must account for fundamental squishiness rather than trying to reduce it to zero.
Gardener vs. builder mindset. Builder: plan, execute, value — capped at energy put in. Gardener: find compounding loops, plant cheap seeds, water what grows. Looks like luck; is actually farming for miracles. Requires accepting illegibility to the organisation during execution.
Organisational kayfabe. Every organisation systematically diverges from ground truth as status reports propagate upward. Single root asymmetry: you cannot make your boss look dumb. This alone creates compounding distortion, zombie organisations, and leaders operating many orders of magnitude from reality.
Adjacent possible + North Star. The adjacent possible is small — only actions within arm’s reach are genuinely available. Take one, the world reconfigures, get the next set. Combine with a low-resolution 3–5 year North Star to avoid random walking. The risk: false precision at great expense on numbers that don’t matter.

LLMs as disruptive technology

The industry presupposes: software is expensive to write, cheap to run. LLMs break both sides simultaneously:

Writing shitty software becomes much cheaper (good software remains hard).
Inference costs mean running LLM-powered software is relatively expensive; advertising cannot clear inference costs for consumer apps.

Analogy: a room tilts five degrees on its axis. Everything looks the same, but gravity now pulls differently. All the intuitions built when gravity was fixed are now wrong.

Individual amplification is stealth: productivity gains from LLMs accrue individually and stay invisible to organisations. Workers report no gains to avoid headcount reduction. This means AI impact is being systematically underreported in aggregate.

The squishy computer

Classical computing: does exactly what you said, not what you meant. LLMs: does roughly what you meant, not exactly what you said.

Design implication: stop asking “how do I reduce the face-punch rate from 5% to 1%?” Start asking “given that it will sometimes punch users in the face, what products can I build where that failure mode is acceptable or manageable?” The answer is different from trying to use LLMs as high-precision oracles.

Gardener vs. builder

Dimension	Builder	Gardener
Direction	Plan → execute	Find compounding potential → plant
Value ceiling	Capped at effort	Uncapped if loop takes hold
Legibility	High (visible work)	Low (looks like luck)
Failure mode	Expensive failure still looks like effort	Success attributed to luck, not design
Organisation risk	Low	High (requires trust / cover fire)

Cover fire model. 70% of team effort on clearly legible, unambiguous value work — eliminates “what does that team do anyway?” risk. Remaining 30% is seed-planting protected by the 70% credibility.

Ecosystem qualification. Gardening works when the thing, if it works, would self-accelerate — network effects, compounding adoption, increasing returns. Truffle-hunting: this structure is more common than people assume.

Organisational kayfabe

Kayfabe (carny/professional wrestling term): a thing everyone knows is fake and yet acts as if real.

Mechanism:

IC makes yellow status green to avoid management review (rational, local).
Manager does the same upward.
After several layers: leadership is operating far from ground truth.

Why it can’t be solved by speaking truth:

Surfacing ground truth identifies the truth-teller as the cause of the problem.
Senior response: “help me fix it, don’t announce it.”
Cognitive load of holding split-belief eventually resolves by genuinely believing the kayfabe.
Result: zombie organisation — all individuals privately agree it cannot work; collective continues.

This is emergent, not evil. The single asymmetry (“can’t make your boss look dumb”) is sufficient and necessary. No bad actors required.

Vehicle metaphor. Small company = sports car; can be steered quickly. Large company = big rig; must be driven accordingly. Apple: one big rig, perfect coherence illusion. AWS: swarm of sports cars, visible incoherence, anti-fragile aggregate capability.

Slime mold principle

Core thesis: coordination costs grow with the square of the number of people working on a shared project. This is not a management failure — it is a mathematical property of the system.

Companies that ignore this: grind against the force. Companies that fight it: burn out. Companies that embrace it: leverage slime molds’ surprising problem-solving capacity (emergent pathfinding, anti-fragility).

Slime mold acknowledgement does not mean “go slow.” It means: pivot less, invest more in programme management, give more slack in planning, drive the vehicle you have.

Adjacent possible + North Star

Adjacent possible (from design thinking): the small set of actions genuinely reachable right now. Each action reconfigures the world and reveals the next set. The mistake is assuming the adjacent possible is large and jumping to the end-state.

Why incremental-only fails: gradient descent into a local maximum. Need a North Star to provide direction.

North Star requirements:

Low resolution — not precise targets.
3–5 years out.
Plausible to all relevant domain experts.
High-five-worthy if achieved.
Updates slowly (slides across the sky, does not jerk around).

Combined model: find the adjacent action that has the steepest gradient toward the North Star. Take it. Repeat.

Strategy salons

Also called “nerd clubs.” Self-organising emergent idea generation groups.

Setup requirements:

Start with the smallest seed of people who already find conversations generative together.
Establish yes-and norms explicitly. If you disagree, don’t engage — nothing is at stake.
Dribble in new perspectives (1–3 per week); prioritise diversity over volume.
Feed momentum: ensure quorum at live events; send FOMO summaries afterwards.
Prune bad actors early — one jerk can kill the culture.

What emerges: ideas that lots of diverse people build on; these are strong signals of large potential audience and real insight. Cannot be steered toward a predetermined outcome — that kills emergence.

Komoroske has started 8–9 such groups over his career. Several still run at Google without him.

Taste as differentiation

In an era of cheap content production (LLMs), standing out requires genuine perspective. Komoroske’s operational definition: differentiate from what the LLM would have written given the same prompt. How distinctive is your specific point of view? Taste is individual and also compelling to others — resonance with diverse audiences is the test.