Core Principles for Builders

Thinking
in Systems
— AI Age

A field guide to understanding the invisible architecture behind how products scale, AI models improve, teams fail, and creators compound — told through the systems that actually govern your work today.

System LOOP STOCK FLOW DELAY FEEDBACK
01
01 · Foundation

What is a System?

"A system is a set of elements interconnected in such a way that they produce their own pattern of behavior over time."

Most people look at a failing product launch and blame the marketing team. Or look at a buggy codebase and blame one engineer. I've found that you're usually looking at events, not the system. Events are the visible surface. The system is the invisible engine underneath.

Every system has three things: stocks (what accumulates), flows (what changes the stock), and feedback loops (information that controls the flows). Once you can see these three things in any situation — a startup, an AI pipeline, your creative output — you stop reacting to events and start designing the system.

Mental Model

Think of a stock as a bathtub. The inflow tap fills it; the drain empties it. The level of water is the stock. The rate of filling/draining is the flow. The thermostat (sensor + goal) is the feedback loop. Simple. Powerful. Universal.

INFLOW Accumulates ↑ Stock + STOCK What Accumulates Current Level OUTFLOW Depletes ↓ Stock FEEDBACK SIGNAL Information Flow ⏱️ System Core STOCK + FLOW + FEEDBACK EXAMPLE: Users → Churn → Retention
Fig 1.1 — The universal building block of any system
WATER LINE EVENTS What we see PATTERNS Trends over time STRUCTURE The System Design CAUSALITY
Fig 1.2 — The Iceberg Model: Events are only the tip of the system
🤖
AI Model Training
Stock: Model accuracy. Inflow: Quality training data. Outflow: Model drift, data staleness. Feedback: Evaluation metrics (BLEU, F1) that determine next training run. The model "knows" how good it is and self-corrects.
📱
SaaS Product Growth
Stock: Active subscribers. Inflow: New sign-ups (via SEO, ads, word of mouth). Outflow: Churn rate. Feedback: NPS scores, support tickets → triggers retention features. Classic leaky bucket.
🎬
YouTube Channel
Stock: Subscribers + watch hours (trust). Inflow: New videos, shares, algorithm pushes. Outflow: Unsubscribes, irrelevant content drift. Feedback: CTR + watch time signals that reshape your content strategy.
💡
Developer Knowledge
Stock: Technical capability. Inflow: Learning, shipping, code review. Outflow: Skill obsolescence (React → RSC → AI-native). Feedback: Compilation errors, PR comments, prod incidents — all signals that adjust learning.
02 · Dynamics

The Two Loops

There are only two kinds of feedback loops in the universe. Everything you've ever seen grow, crash, stabilize, or spiral — came from these two.

Reinforcing Loop (R+)
A change in stock A causes more of the same change. Growth amplifies growth. Collapse amplifies collapse. Also called a "positive feedback loop" — not positive as in good, positive as in amplifying.

Real example: LLM Adoption. More users → more revenue → better GPUs → smarter model → more users. OpenAI's growth loop is a textbook R+ loop. Once it tips, it compounds.

Dangerous example: Technical Debt. More shortcuts taken → more bugs → more firefighting → fewer hours to write clean code → more shortcuts. This is also an R+ loop. It amplifies in the wrong direction.

Balancing Loop (B–)
A change in stock A creates pressure to bring it back toward a goal. It resists change. All goal-seeking behavior — thermostats, immune systems, KPI dashboards — is a balancing loop.

Real example: Alert Fatigue. Error alerts spike → engineers get paged → they fix issues → alerts drop → fewer engineers monitoring → errors spike again. Every on-call rotation is a B– loop.

Real example: Content Moderation. Bad content increases → moderation ramps up → bad content decreases → moderation scales back → bad content increases. Platforms are stuck managing this oscillation permanently.

REINFORCING (R+) USER GROWTH PRODUCT QUALITY + + R BALANCING (B–) ACTUAL PERF GOAL / TARGET GAP – corrects B
Fig 2.1 — Reinforcing loop amplifies change · Balancing loop resists change toward a goal
"The world is filled with powerful reinforcing loops. But there is no such thing as unlimited growth. Every reinforcing loop eventually runs into a balancing loop."
Core Principle of System Dynamics

Why this matters for AI builders: Most AI product teams only manage R+ loops (growth, model improvement). They neglect the B– loops — rate limiting, content filters, user fatigue, regulatory pressure — until those balancing forces crash the party. Design both, or the system designs you.

03 · Time Gaps

Delays — The Silent Killer

A delay is the gap between cause and effect. It's the most misunderstood component of any system — and the root cause of most overreactions.

t=0 CAUSE DELAY ··· t+n EFFECT IMPATIENT INTERVENTION OSCILLATION · CAUSED BY DELAY
Fig 3.1 — Delays create oscillation when decision-makers act before the effect arrives

When you can't see the effect of your action immediately, you do one of two things: you over-apply the fix (because nothing seems to be working) or you abandon the strategy (because you don't see results). Both are wrong. Both are caused by ignoring delays.

I believe most political, organizational, and personal disasters come from people not accounting for delays — pushing harder on a solution that was already working, or giving up on one that just hadn't kicked in yet.

Delay Types in Tech

Perception delay — time before you notice something changed (data pipeline lag). Response delay — time to decide and act (approval chains, sprint planning). Effect delay — time before your action produces results (SEO, model fine-tuning).

🧠
LLM Fine-tuning Delay
You inject new training data today. The model improvement shows up in eval metrics in 2–3 weeks. Teams that don't track this keep changing the data pipeline before the first intervention even had a chance to work. Result: chaotic, untraceable model degradation.
📈
SEO Content Lag
Google indexes a page in days, but domain authority takes 3–6 months to materialize in rankings. Startups that stop publishing after "nothing happened in 60 days" kill their own compounding loop. The cause happened. The effect is just delayed.
🏗️
Hiring Decisions
You hire 5 engineers to meet current demand. By the time they're onboarded and productive (90–180 days), demand has already shifted. Every engineering org in a hypergrowth phase has hired for yesterday's problem because they couldn't see the delay.

The fix: Don't increase the rate of action — increase the accuracy of your information. Know your system's delay time. Set a waiting period before changing strategy. A pipeline with a 6-week delay needs a 6-week patience window before you call it a failure.

04 · Patterns

System Archetypes

I've identified patterns that repeat across industries, centuries, and contexts. Once you see them, you see them everywhere. Here are four that are actively destroying products, teams, and careers right now.

Limits to Growth
The R+ loop that always hits a wall

A system grows, then slows — not because of bad strategy, but because it hit an unnoticed constraint. Removing the constraint restores growth. Ignoring it and pushing harder makes it worse.

Tech example: Startup scales its AI API — users grow fast, then DAU plateaus. Teams double down on marketing. But the real limit? Response latency hit 8 seconds at scale and users silently churned. The constraint was infrastructure, not demand.

// Fix: find the limiting constraint, don't push the accelerator
Shifting the Burden
Symptomatic fixes that make root causes worse

A problem creates pressure. You apply a quick fix (symptomatic solution) instead of a hard fix (fundamental solution). The quick fix works temporarily, reducing urgency for the real fix — so the root cause grows.

Tech example: App crashes under load. Team adds more caching (quick fix). This masks the problem — the real bug is an N+1 query that's getting worse. The more caching they add, the less motivated they are to find the actual query. One day caching can't save them.

// Fix: invest in fundamental solutions, not just symptomatic patches
Tragedy of the Commons
When individual rationality destroys collective resources

Each individual rationally exploits a shared resource. Collectively, the resource is destroyed. No single actor is "wrong" — the system design is wrong.

Tech example: API rate limits shared across a team. Every developer adds just a few more calls. Nobody individually exceeds their "fair share." Together they hit the ceiling, break production, and spend a sprint debugging. The shared GPU cluster, the shared CI/CD pipeline, shared Slack channels — all commons tragedies in waiting.

// Fix: define individual quotas + make the shared resource visible
Escalation
Two R+ loops in opposition

Party A acts. Party B responds with more. A escalates further. Each party is just reacting — but the system spirals. The most common outcome: both lose.

Tech example: Google and OpenAI racing to release models faster. Each announcement makes the other accelerate. Safety tests get shorter. Alignment work gets delayed. Neither company wanted a race — but the system structure created one. Also visible in: ad bidding wars, VC valuation inflation, open-source one-upmanship.

// Fix: one party must unilaterally de-escalate or introduce a governor
LIMITS TO GROWTH SHIFTING BURDEN COMMONS ESCALATION growth → plateau fix → bounce → worse overshoot → collapse A escalates · B responds
Fig 4.1 — Behavioral signature of the four most common system archetypes
05 · Intervention

Leverage Points

One of the most counterintuitive insights I follow: the places where small changes produce large effects are rarely where you'd expect to find them.

I rank leverage points from least to most effective. Most organizations spend all their time at the bottom of this list — tweaking numbers and parameters — when the highest leverage is in changing the goal of the system or changing the paradigm the system operates under.

PARADIGMS Highest Leverage GOALS & RULES Structural Change NUMBERS & PARAMETERS Easiest but Weakest INTERVENTION POWER
Fig 5.1 — The Leverage Hierarchy: Where small changes produce large effects
12
Constants & Parameters (Numbers)
Tech example: Changing your API rate limit from 100 to 150 req/min. Tweaking a model's temperature from 0.7 to 0.9. These feel like work. They rarely change outcomes significantly. Almost all A/B testing lives here.
9
Feedback Loop Strength
Tech example: Adding real-time model eval dashboards. Shortening the sprint review cycle from 2 weeks to 1 week. Making feedback loops faster and stronger is dramatically more impactful than adjusting the numbers they measure.
6
Information Flows (Who gets to see what)
Tech example: Making production error logs visible to every engineer, not just the on-call team. Publishing your API cost breakdown to all internal stakeholders. Information visibility is one of the most underused leverage points in most companies.
3
The Goal of the System
Tech example: Changing what an AI product optimizes for — from "engagement time" to "user-reported wellbeing." Instagram optimized for time-on-app and got a mental health crisis. Changing the goal from engagement to connection would restructure every sub-system beneath it. This is why mission statements actually matter.
1
The Mindset — The Paradigm of the System
Tech example: The shift from "software is a product you sell" → "software is a service you continuously operate" restructured the entire SaaS industry. The shift from "AI is a research project" → "AI is infrastructure" is restructuring every company today. Paradigm shifts are the most powerful leverage point. They change everything, all at once.
The Counterintuitive Warning

I often note that high-leverage points can work in the wrong direction. Optimizing hard for a misaligned goal (engagement → addiction) is worse than having no goal at all. The power of the leverage point amplifies whatever direction you push it. This is why "move fast and break things" breaks real things when the goal isn't carefully designed.

06 · Application

Your Work as a System

You don't manage tasks. You operate a system. Once you see it that way, you stop asking "how do I get more done?" and start asking "where is the bottleneck, and what's the feedback signal I'm missing?"

Take any creator, engineer, or founder managing multiple workstreams simultaneously — content creation, client delivery, learning, health. The natural instinct is to make a better to-do list. That's tweaking a parameter (Leverage Point #12). The systems thinker asks different questions entirely:

What is the stock I'm building? (reputation, skill, audience, codebase) What are the inflows and outflows? (output rate, consistency vs. distraction, skill decay) What's the feedback loop telling me? (engagement metrics, client satisfaction, test results) What's the delay I need to respect? (audience building takes 12–18 months; don't read last week's analytics as signal)

REPUTATION + Output Quality DEEP WORK CONSISTENCY DISTRACTION SKILL DECAY SIGNAL: ENGAGEMENT ⏳ DELAY: 6–18 MONTHS TO COMPOUND
Fig 6.1 — Personal output as a stock-and-flow system
01
Map your stocks
What are you actually accumulating? Audience trust, code quality, domain expertise, client relationships. Be specific. These are your real assets — not tasks completed.
02
Identify your inflows and outflows
What fills your stock? What drains it? For a developer: inflow = shipping, learning, code reviews. Outflow = context switching, meetings without decisions, technical debt accumulation.
03
Design your feedback loops
What signals tell you when something is off? If you have no feedback mechanism, you're flying blind. Weekly reviews, engagement metrics, peer feedback — these aren't optional. They're the nervous system of your work system.
04
Respect delays — build patience into your system
Know the delay for each system you operate. Content: 6 months. Skill building: 90 days minimum. Product-market fit: 12–18 months. Build decision rules: "I will not change strategy until I've waited [delay period] after the last change."
05
Find your leverage point — and act there
Don't optimize tasks. Find the one structural change that changes everything. For most builders: it's information flow (who sees what, when). For most creators: it's the goal (create for learning vs. for validation).
⚙️
Think in Structures
Events are symptoms. Patterns are clues. Structure is the cause. Stop managing events — design structures.
🔁
Close the Loop
Every system without feedback is running open-loop. An open loop cannot self-correct. Every process needs a signal and a response.
Honor Delays
Patience is not passive. It is the active recognition that cause and effect are separated in time. React to the right time horizon.
∞ · Closing

The Real Shift

Systems thinking isn't a tool you deploy on Fridays. It's a lens you can never fully take off once you've put it on.

I strongly believe in the principle of staying humble and staying a learner when navigating these systems. The most radical advice I follow: "Stay humble. Stay a learner." Because systems are complex. Our models of them are always incomplete. The best systems thinkers are those who hold their understanding loosely — continuously updating as new information arrives.

In the AI age, this is more urgent than ever. The systems we build (products, models, organizations) now move faster than our ability to observe them. The feedback loops are shorter. The delays are smaller. The reinforcing loops are more vicious.

But the core insight remains unchanged: you cannot manage what you cannot see the structure of. Learn the structure. Build feedback. Respect delays. Intervene at leverage points. And when you find yourself tweaking a parameter for the third time without results — stop. Look for the system.

systems thinking feedback loops AI products leverage points delays complexity mental models
Suggested Reading

Many of these principles are detailed in the foundational work "Thinking in Systems" by Donella H. Meadows. It is highly recommended for anyone looking to go deeper into the invisible architecture of the world.

PARADIGM SHIFT highest leverage · hardest to change GOALS OF THE SYSTEM what the system is optimizing for INFORMATION FLOWS who knows what, when FEEDBACK STRENGTH speed + quality of correction signals DELAYS · BUFFERS timing & resilience in the structure PARAMETERS · NUMBERS
Fig 6.2 — Leverage hierarchy: where most teams act vs. where they should