Notes — Alex Komoroske on Strategy and Complexity

Four questions [Adler frame]

Q1 — What is it about? A practitioner’s guide to navigating complexity in organisations and product strategy, using emergence-based thinking rather than top-down plans. Komoroske argues that (1) LLMs are a genuinely disruptive technology that changes the cost-structure assumptions underlying all existing software playbooks; (2) organisations are more like emergent slime molds than like machines, and fighting that is fatal; (3) the way to create outsized value is to garden rather than build — find compounding loops and plant cheap seeds. The episode is a live demonstration of his “Bits and Bobs” note-taking and idea-refinement practice.

Q2 — How is it argued? By analogy and first-principles observation. Komoroske does not cite studies; he derives frameworks from personal observation over 13 years at Google and a stint at Stripe. He uses evocative metaphors (slime mold, magical duct tape, acorn, squishy computer, sports car vs. big rig) deliberately — he explains his metaphor-generation process as finding frames that resonate across diverse audiences, indicating large potential reach. Each framework is grounded in a mechanistic explanation: kayfabe arises from the single asymmetry “don’t make your boss look dumb”; adjacent possible arises from the small size of the action set in an unknown environment.

Q3 — Is it true? The kayfabe diagnosis rings true structurally; it is consistent with Goodhart’s Law and principal-agent literature, though Komoroske does not frame it that way. The gardener/builder distinction is a useful but incomplete heuristic — gardening requires an environment where compounding is possible, which is not always the case. The adjacent possible framework is loosely borrowed from design thinking and origin unclear (Komoroske attributes it to designers he has worked with). The LLMs-as-magical-duct-tape framing is idiosyncratic but analytically useful for understanding why existing SaaS playbooks misfit. The squishy computer / 99%-punches-you-in-the-face observation is practically important and underappreciated.

Q4 — What of it? The frameworks are most useful to PMs in large organisations navigating misalignment and for product builders thinking about what kind of value AI can reliably deliver. The gardener mindset reframes “planting seeds and watching what grows” as rigorous strategy, not laziness — this is a legitimate counterweight to overplanning. Kayfabe awareness is a diagnostic tool for understanding why orgs underperform despite capable individuals. The adjacent possible is a constraint on strategic ambition that reduces risk without sacrificing direction.

Glossary

Magical duct tape — Komoroske’s label for LLMs. Duct tape because they are highly versatile, work in many unexpected places, and are made of distilled societal intuition rather than formal specification. “Magical” because the capability is surprising relative to how they work. Undermines the assumption that “software is expensive to write, cheap to run.”

Squishy computer — LLMs do roughly what you meant, not exactly what you said. Contrast with classical computing which does exactly what you said, not what you meant. 99% accuracy still fails as a product if the 1% failure mode is catastrophic or humiliating to users.

Gardener vs. builder mindset — Builder: make a plan, execute the plan; can’t create more value than effort put in. Gardener: find things that grow on their own; direct and curate; can create far more value than effort put in because compounding does the work. Farming for miracles — on a systemic basis, one seed will grow into an oak tree; the trick is to plant cheaply and water only what starts growing.

Organisational kayfabe — From professional wrestling: “a thing everyone knows is fake and yet everyone acts as if real.” In organisations: the accumulated optimism and status reporting that diverges from ground truth as it propagates up the hierarchy. Single root asymmetry: you cannot make your boss look dumb. This alone creates systemic compounding distortion. Zombie organisation: an org in which no individual believes the official narrative but the collective still executes it.

Slime mold principle — Coordination costs grow with the square of the number of people working on a shared project. Organisations that ignore this grind; those that fight it burn out; those that embrace it (accepting that they are more like emergent swarms than machines) can leverage the emergent problem-solving capacity of slime mold. Sports car vs. big rig: steer proportionally to your actual size. Swarm of sports cars: accept visible incoherence in exchange for anti-fragile autonomy (e.g., AWS product suite).

Adjacent possible — The small set of actions that are genuinely reachable right now. Take one, the world reconfigures, and you get the next set. Avoid jumping to an end-state; slice decisions into small steps that each clearly pay for themselves. Combine with a North Star (low-resolution, 3–5 year direction everyone can believe in) to avoid random walking. The risk of long-horizon planning is false precision at great expense; the risk of pure incrementalism is gradient descent into a corner.

Strategy salons / nerd clubs — Self-organising idea generation groups. Key properties: opt-in (only people who intrinsically want to be there), yes-and norms, curated momentum (FOMO after live sessions, gardener pruning bad actors), novelty search through diverse perspectives. Cannot be steered toward a predetermined outcome — that destroys them. Spin off insights stochastically. Komoroske has started 8–9 over his career.

Hallmark card fallacy — The observation that deep personal insights, once acquired, sound like clichés to others who have not yet been ready to hear them. The phrase has been shared many times because it is true; the sharing predates the listener’s readiness. Seeds can be planted even into unready soil; they grow when the right experience opens a crack.

LLMs as disruptive technology

The underlying cost-structure assumption of the entire software industry: software is expensive to write, cheap to run. LLMs break both sides:

Writing software becomes cheaper (shitty software very cheaply; good software remains hard).
Running software becomes relatively expensive (inference costs; consumer startups cannot be ad-funded at current inference cost levels).

The implication: gravity has tilted five degrees. All the existing playbooks — vertical SaaS, ad-supported consumer, API-first — were built assuming a different gravitational direction. Most incumbents do not yet notice this.

Current failure mode: using LLMs as oracles — “formulate the answer and output a fully fledged thing.” The squishy computer nature means this produces a 5% face-punch failure rate that is not viable in most products. The correct design question: given magical duct tape (versatile but imprecise), what can I now build that was impossible before? Not “how do I apply AI to my existing thing?”

Individual amplification is stealth: AI productivity gains accrue at the individual level, often invisibly to organisations. Workers are reluctant to report productivity gains that could lead to headcount reduction. This means aggregate AI impact is being systematically underreported.

Gardener vs. builder in practice

Builder mindset: plan → execute → value. Linear. Fully legible to the organisation. Defensible even if it fails (“we worked hard”). But capped at the energy put in.

Gardener mindset: find ecosystem with compounding potential → plant cheap seed → respond to growth signals → water only what’s working → stop when signals cease. Non-linear. Looks like luck after the fact. Illegible to the organisation during execution. Looks like “we got lucky” even when it succeeds.

The problem: designing for emergence looks unserious. Komoroske’s solution: 70% of team effort on clearly legible, unambiguous value work (minimises “what does that team do anyway?”). The 30% residual creates space for seed-planting that the 70% credibility protects.

Ecosystem qualification: gardening works when the thing, if it works, would accelerate — network effects, compounding loops, self-reinforcing adoption. Lots of things have this structure if you look for it (truffle-hunting analogy).

Kayfabe mechanics

Single asymmetry → cascade:

Individual contributor makes a status yellow green to avoid management review (rational local decision).
Manager does the same upward.
Multiple layers later: leadership is operating on information many orders of magnitude from ground truth.

Ground truth is dangerous to surface because:

The person who tells the truth will be seen as the cause of the problem.
Leadership rational response: “don’t hit the ground truth button; help me fix it instead.”
If fixing proves impossible, holding the split belief (truth + kayfabe) is cognitively costly → easiest resolution is to genuinely believe the kayfabe.
Result: zombie organisation — individuals privately agree it cannot work; collective continues to execute.

This is not cynicism — it is emergent dynamics. Nobody wants it. Nobody caused it. The single asymmetry (“can’t make boss look dumb”) is sufficient to produce the full pathology.

Adjacent possible + North Star

North Star requirements: low resolution; 3–5 years out; plausible to all relevant domain experts (“I could see how that could work”); high-fives if you get there. Updates slowly (slides across the sky, does not jerk).

Adjacent possible: the small set of cheap, safe actions in front of you. Take one, world reconfigures, get next set. The steepest gradient action that also pulls toward your North Star is the right move.

Analogy: Buddhist retreat — point the cart in the right direction and start walking. Don’t pre-plan the path.

False precision trap: organisations spend enormous effort deriving whether the five-year number will be 93 or 95, rather than acknowledging that the right question is “will it be zero or 1,000?” Compounding bets don’t require that precision.