Complexity Is Just a Phase: The Pattern That Explains Why Everything Gets Messy First
A Physics Insight That Might Save Your Startup from Its Own Chaos
The 30-Second Version
MIT’s Scott Aaronson proposed something useful: in closed systems, apparent complexity tends to rise, peak, then fall.
Coffee + milk: Clean split → Beautiful swirls → Boring brown.
Not a law. A lens. And it might explain why your startup feels like chaos right now.
Why you should care: Ilya Sutskever (OpenAI co-founder) put this near the top of the reading list he gave John Carmack, calling it essential for understanding how learning systems evolve.
Why Ilya Sutskever Thinks This Matters for AI
In 2020, when John Carmack (creator of DOOM and Oculus) asked Ilya Sutskever for an AI reading list, Sutskever gave him around 30-40 papers with this claim:
“If you really learn all of these, you’ll know 90% of what matters today.”
The First Law of Complexodynamics appears near the top of reconstructed versions of this list.
Why would an AI pioneer include a physics blog post about coffee mixing?
Because learning systems show similar dynamics:
Simple initialization → Complex training dynamics → Compressed learned representations
Training is essentially “mixing” a model with data—early layers look messy before compression finds stable features. Recent work like Anthropic’s “Toy Models of Superposition” shows neural networks developing complex geometric feature arrangements during training before settling into more stable configurations.
The pattern appears in many places. The question is what to do about it.
The Coffee Cup Insight
In September 2011, at a physics conference, Sean Carroll posed a question:
“Why does complexity rise and then fall, while entropy only increases?”
Scott Aaronson proposed what he called (somewhat tongue-in-cheek) the First Law of Complexodynamics.
The Visual That Started Everything
Picture this sequence:
Coffee and milk, separated → Simple to describe
Swirling, mixing tendrils → Hard to describe precisely
Uniform light brown → Simple again
In 2014, Aaronson, Carroll, and Lauren Ouellette published a paper exploring this with cellular automata. They showed that under their specific measure (”apparent complexity”), the pattern held in their toy model.
Credibility note: This isn’t a universal law. It’s a model shown in specific toy simulations using a specific definition of “apparent complexity.” Use it as a lens, not a deity.
The insight: In closed systems drifting toward equilibrium, complexity tends to peak at intermediate times.
The Pattern in Nature (With Important Caveats)
The Universe Story (Simplified and Speculative)
According to cosmological models:
13.8 billion years ago (Big Bang):
Hot, dense, relatively uniform
Simple to describe at large scales
Now (Present day):
Galaxies, stars, planets, life
Complex structures everywhere
Far future (10^100+ years):
If current models hold, eventual heat death
Particles dispersed, no structure
Simple again (but cold and empty)
Important: This is extrapolation, not established fact. The crucial distinction: this pattern is clearest in closed systems heading toward equilibrium. Businesses are open systems—they import energy, capital, and information. That changes everything.
Why This Matters for Your Business
The Startup Complexity Curve
Based on common patterns observed in tech companies:
Phase 1: Simple Beginning (0-6 months)
Clear vision
Minimal features
Small team
Easy to explain
Phase 2: Peak Complexity (1-3 years)
Feature creep sets in
Process proliferation
Team scaling challenges
Hard to explain simply
Phase 3: Simplification or Death (3-5 years)
Either: Refined core product
Or: Failure
What we know from data:
65-90% of startups fail overall (various studies)
Top failure reasons: No market need (42%), cash problems (29%), team issues (23%) - CB Insights
Feature creep is a known killer but not precisely quantified
My observation: Most failures cluster in the complexity peak zone (years 2-5) when scope creep, org sprawl, and cash pressure collide.
Real Examples (Simplified Narratives)
Amazon’s Arc:
1994: “Online bookstore” (simple)
2000-2010: Expanding into everything (complex)
Now: “We deliver everything” plus AWS (modularized complexity)
Apple’s Journey:
1976: “Personal computers” (simple)
1990s: Dozens of confusing products (complex)
Post-1997: Streamlined product lines (simplified)
Tesla’s Path:
2008: “Electric sports car” (simple)
2015-2020: Production hell, multiple models (complex)
Now: More streamlined production (but adding new complexity with robots/energy)
Note: These are simplified stories. Real businesses often cycle through multiple complexity peaks.
Are You in Peak Complexity? (Self-Diagnosis)
Check yourself against these symptoms:
☐ Your pitch is getting longer, not shorter ☐ Roadmap has more “new stuff” than “fix/kill stuff”
☐ Teams spend more time coordinating than shipping ☐ Customers ask “so what do you actually do?” ☐ You can’t explain your product in one sentence ☐ Every feature feels “critical” ☐ Meetings about meetings are happening
If you checked 3+ boxes: You’re in the swirl. That’s normal. Now stop adding and start cutting.
The Counter-Intuitive Insights
1. Complexity Tends to Be Temporary (In Closed Systems)
For systems drifting toward equilibrium, complexity tends to decrease after peaking. The system either:
Finds a stable configuration (success)
Collapses (failure)
But remember: businesses are open systems. You can sustain complexity if you keep feeding it resources.
2. Adding Features During Peak Complexity Often Backfires
Common pattern observed:
Companies in crisis add features to “innovate out”
This usually increases complexity without solving core issues
Simplification often works better than addition
3. Successful Survivors Often Embrace Radical Simplification
Companies that navigate complexity peaks often:
Cut significant portions of their product line
Reduce organizational layers
Focus on core value propositions
This is observation, not rigorous data. But the pattern appears repeatedly.
The Investment Angle
How to Think About Complexity When Evaluating Companies
Pre-complexity companies (early stage):
Simple story
Limited traction
High risk, potentially high reward
Peak complexity companies (danger zone):
Can’t explain what they do simply
Multiple pivots happening
High burn rate on scattered initiatives
Post-complexity survivors (potentially attractive):
Clear value proposition emerged from chaos
Streamlined operations
Proven ability to navigate difficulty
Investment consideration: Companies that successfully navigate complexity peaks have demonstrated adaptability. Those still in peak complexity face uncertain outcomes.
The challenge: Identifying whether a company is temporarily complex (and heading toward simplification) or permanently complicated (and heading toward failure).
Practical Applications
For Founders
Recognize your phase:
If everything’s simple → Prepare for complexity
If everything’s chaos → Stop adding, start cutting
If emerging from chaos → Double down on what worked
Survival tactics for peak complexity:
Freeze new features
Focus on core metrics only
Cut ruthlessly
Wait it out
For Investors
Red flags (peak complexity):
“We’re a platform for platforms”
10+ product lines
Can’t explain business in one sentence
“AI-powered blockchain for X”
Green flags (emerging simplicity):
Recently cut product lines
Clear focus after pivot
“We only do X now”
Simplified pricing
For Employees
Career timing:
Early phase: High learning, high risk
Complexity phase: Chaos but opportunity
Simplicity phase: Stability but less upside
The trade-off: Post-complexity hires usually get less equity but far more stability.
The Science Behind It (Simplified)
Why This Pattern Emerges
According to Aaronson’s blog post and subsequent paper:
Initial state: Low entropy, low complexity (organized, simple)
Mixing begins: Entropy increases, complexity rises (interesting patterns emerge)
Equilibrium approaches: Entropy maximizes, complexity falls (boring uniformity)
The key insight: Complexity ≠ Entropy
Entropy (disorder) only increases in closed systems. Complexity (interesting structure) can rise and fall.
The Technical Definition
Kolmogorov Complexity: The length of the shortest computer program that can describe the system.
Separated coffee/milk: Short program (draw two rectangles)
Mixing: Long program (can’t compress the swirl patterns easily)
Mixed: Short program again (fill with uniform color)
The pattern depends on the system being closed and drifting toward equilibrium. Open systems (like businesses) can sustain complexity with continuous inputs.
What Nobody Tells You
The Universe’s Complexity Arc (Maybe)
Some cosmologists suggest:
The universe might be near peak apparent complexity
We exist in an era of maximum interesting structure
Future eras might see declining complexity
But this is highly speculative and depends on:
How you measure complexity
What happens with dark energy
Whether new physics emerges
It’s a compelling story, not established fact.
Your Business Will Peak Too (Unless You Feed It)
In closed systems, the pattern is clear. But businesses are open systems—they can:
Import new energy (capital)
Bring in new information (talent, data)
Create new gradients (markets, products)
The real insight: If you stop feeding your business new inputs, it becomes a closed system. Then complexity peaks and falls.
The question isn’t IF complexity will peak. It’s WHETHER you’ll navigate it and WHAT you’ll do next.
The Bottom Line
The First Law of Complexodynamics isn’t a proven physics law. It’s a useful model for thinking about how systems evolve.
Three practical insights:
Complexity often peaks - Don’t panic when things get messy
In closed systems, it’s temporary - But you need to navigate it
Open systems can sustain complexity - But only with continuous inputs
While others panic during peak complexity, you’ll recognize: this might be a phase, not a failure.
The mess isn’t necessarily a bug. It might be part of the process.
But unlike coffee mixing with milk, you can influence the outcome.
Action Items
First, determine your system type:
Closed system (no new inputs) → Complexity will peak and fall
Open system (continuous inputs) → You can sustain or restart complexity
If you’re pre-complexity:
Document everything while it’s simple
Build reserves for the chaos phase
Prepare for 2-3 years of increasing complexity
If you’re in peak complexity:
Stop adding features immediately
Start identifying what to cut
Focus on surviving, not optimizing
Consider: Can you add new inputs to remain open?
If you’re post-complexity:
Double down on what survived
Resist urge to re-complexify without purpose
Scale the simplified solution
Stay open to prevent stagnation
The meta-strategy: Engineer your downslope. Don’t wait for equilibrium to find you.
Peak-Complexity Survival Checklist
If you’re in the swirl right now, here’s your playbook:
□ Freeze new features for 6-8 weeks (no exceptions)
□ Kill one “pet project” per sprint (the ones no customer asked for)
□ Make every team ship ONE metric, not five (focus beats breadth)
□ Rewrite your one-sentence pitch weekly (if it gets longer, you’re still climbing)
□ The 12-year-old test: If you can’t explain your product to a smart 12-year-old in 30 seconds, you’re still in peak complexity
□ Cut meeting time by 50% (complexity feeds on coordination overhead)
□ Ask weekly: “What would this look like if it were simple?”
Remember: You’re not failing. You’re in a phase. The goal is to navigate it, not to win it.
One More Thing
Remember when everyone said remote work would simplify everything?
We went from:
Simple: “Everyone comes to office”
To complex: “Hybrid async-first distributed collaboration with virtual presence”
Heading toward: “Work from anywhere” or “Everyone back in office”
Even our work models follow the law.
The pattern is everywhere once you see it.
Sources & Further Reading
The Original Post: Aaronson’s “The First Law of Complexodynamics” (September 2011)
The Scientific Paper: Aaronson, Carroll & Ouellette - “Quantifying the Rise and Fall of Complexity in Closed Systems” (2014)
Ilya’s Reading List: The legendary 30 papers that Sutskever told Carmack would teach him “90% of what matters”
Sean Carroll’s Book: “From Eternity to Here: The Quest for the Ultimate Theory of Time” - The book that inspired the original question
Follow-up Research: Aaronson’s update on the Coffee Automaton with corrections and improvements
Related Concepts:
Anthropic’s Toy Models of Superposition - showing complexity in neural networks
Business Applications:
Startup failure rates: CB Insights Startup Failure Report
Thanks for writing this, it clarifies a lot. This insight on complexity rising and falling, especially for AI learning systems, perfectly frames training dinamics. Makes so much sense.