Back to Reports
Deep Dive Report

The 90% Trap

Why Every Generation of Engineers Thought They Were Five Years Away — And Why They Were All Right

Topic ID #6276
Comments 8,000+
Years 18
Range 2007-2026

The Hidden Pattern

The conventional wisdom says:

"Self-driving cars are always five years away because the technology is harder than we thought."

The data tells a different story.

Every generation of engineers DID solve 90% of the problem. The issue is that each generation solved a DIFFERENT 90%.

90% × 90% × 90% × 90% = 65%

Four generations of "almost done" — and we're still not there.

The "last 10%" isn't a finish line — it's a horizon that moves as you approach it.

Four Generations, Four "90% Done" Moments

Each era's engineers genuinely solved most of their problem. Then a new problem appeared.

2007-2012

The DARPA Dreamers

348
comments
The Problem
"Can a car drive itself at all?"
What They Solved
Yes — in controlled environments
The New Problem
But what about traffic? Pedestrians? Weather?

HN Sentiment: Pure optimism. The hard part seemed done.

"The DARPA challenge seemed to prove it's possible today."
2013-2016

The Hype Singularity

1,925
comments
The Problem
"Can it work in real cities?"
What They Solved
Yes — Google drove a blind man to Taco Bell in 2012
The New Problem
But that was one route. What about everywhere?

HN Sentiment: Corporate promises peaked. "2-5 years" became mantra.

"Volvo are planning to ship their first self driving cars in 2014."
2017-2020

The Edge Case Apocalypse

2,861
comments
The Problem
"Can it handle edge cases?"
What They Solved
Some of them — Waymo launched limited pilots
The New Problem
But edge cases are infinite. Each solved reveals three more.

HN Sentiment: Reality set in. Uber crash (2018) changed the tone.

"Google drove a blind man to Taco Bell in 2012 and still hasn't released a self driving car 4 years later."
2021-2026

The Geofence Compromise

2,846
comments
The Problem
"Can it work without human backup?"
What They Solved
Yes — in Phoenix, in San Francisco, in specific zones
The New Problem
But that's not "self-driving cars." That's "robot taxis in selected areas."

HN Sentiment: Pragmatic acceptance. The goalpost moved, and we pretended it didn't.

"Getting the last 10% of edge cases to work seems to require a monumental effort still."

The Prophets & The Optimists

Those Who Saw Further

"Self driving cars are 15 years away."

2011 + 15 = 2026. We're here. Still waiting for Level 5.

"I am making a bold prediction that it will be decades before a self driving car can drive itself safely on Indian streets."

The "Indian streets test" became shorthand for real-world complexity.

"We'll never see a Google self driving car commercially. It's a science experiment."

Waymo exists, but "Google car in your driveway"? Still waiting.

Predictions That Aged Poorly

"GM's claim they would have a self driving car available for sale in 2015."

Lesson: Corporate timelines are marketing, not engineering.

"Volvo are planning to ship their first self driving cars in 2014."

Lesson: Shipping "self-driving features" ≠ shipping "self-driving cars."

"I'll bet you money there's no legal self-driving car subscription like that in DC by 2023."

Lesson: He was right. The optimist lost this bet.

So What?

What 18 years of self-driving predictions teach us about technology forecasting.

📈
For Investors

Discount all tech timelines by 3x

When engineers say "5 years," hear "15 years." When they say "90% done," understand that the last 10% IS the product. The Waymo/Cruise billions were spent on the last 10%.

⚙️
For Engineers

Demos are not products

The Google Taco Bell demo was real. It also meant nothing for general deployment. Your impressive demo is the starting line, not the finish. Plan for 10x the edge cases you can imagine.

🚀
For Founders

Solve a smaller problem completely

Waymo won by shrinking the problem (geofencing) rather than solving the original one (everywhere). Sometimes "changing the goalposts" is the right strategy. Ship something real > promise something revolutionary.

Ready to Share

"Self-driving cars are 15 years away." — HN user, 2011. It's 2026. He was right.
View original →
Every generation of self-driving engineers solved 90% of the problem. The issue: each solved a DIFFERENT 90%.
"Google drove a blind man to Taco Bell in 2012 and still hasn't released a self driving car 4 years later." — HN user, 2016. It's now 14 years later.
View original →
90% × 90% × 90% × 90% = 65%. That's why "almost done" never means done.

The Verdict

Self-driving cars weren't "always five years away" because engineers were wrong. They were five years away from solving the problem they could see.

The real lesson isn't about autonomous vehicles. It's about how complex systems reveal themselves only through contact with reality. The next "five years" isn't a lie — it's the honest assessment of a problem that grows as you approach it.

The 90% Trap: In complex systems, being "90% done" four times still leaves you 35% short. The finish line isn't a place — it's a process of discovering what "done" actually means.

HN Zeitgeist — 8,000 comments. 18 years. One pattern nobody saw.