Redefining great: When Collins measured trains, the AI era needs to measure rockets

In 2001, Jim Collins published Good to Great after 5 years of research with 21 collaborators. They screened 1,435 companies in the Fortune 500 from 1965 to 1995 and identified 11 that met the "great" standard: cumulative stock return at least 3 times the market over 15 years.

The 11 companies were: Abbott, Circuit City, Fannie Mae, Gillette, Kimberly-Clark, Kroger, Nucor, Philip Morris, Pitney Bowes, Walgreens, and Wells Fargo.

One simple question: after the book came out, how did these companies fare?

Circuit City went bankrupt in 2009. Fannie Mae was taken over by the U.S. government during the 2008 crisis. Pitney Bowes lost roughly 90% of its market value. Walgreens was taken private by Sycamore Partners in 2025, delisting after the stock had dropped more than 60% from its peak. Wells Fargo got caught in the fake-accounts scandal of 2016, losing billions in fines and reputation. Gillette was acquired by Procter & Gamble in 2005 — that was a successful exit for shareholders, not a failure.

The remaining six companies — Abbott, Kimberly-Clark, Kroger, Nucor, Philip Morris, and a few others before — are mostly still operating and still large. Abbott has roughly doubled in value over two decades. Kimberly-Clark, Kroger, Nucor, Philip Morris have grown normally, no major crisis.

Counting strictly: 4 of 11 collapsed or lost serious value (Circuit City, Fannie Mae, Pitney Bowes, Walgreens). Add Wells Fargo in the gray zone. About a third of Collins's "great" list didn't survive 20 years after the book came out. If Collins's framework really captured the essence of greatness, why this outcome?

This is the point where many people stop and conclude: "So Collins was wrong, ignore the book." I think that conclusion is as hasty as Collins's own. The better question is: Where was Collins wrong, where was he right, and what's worth learning from his method for the AI era?

The problem isn't sample size. It's how the sample was selected.

A common critique of Collins is: "11 companies is too small a sample to draw conclusions." This critique sounds right but misses the methodological context.

Collins wasn't doing statistical inference. He was doing case study research combined with grounded theory. He didn't need a large sample in the statistical sense — he needed qualitative depth. Attacking Collins with the "n=11 isn't enough" argument is attacking with the wrong weapon.

The right critique in framework terms should be:

One is survivorship bias. Collins chose companies based on results already in hand — those that had outperformed 3x over 15 years — then worked backward to find causes. This is the classic problem: companies with the same characteristics but that failed don't enter the sample. The result: the pattern he found may also appear in failed companies, and we wouldn't know.

Two is uncontrolled confounders. Collins tried to control by using "comparison companies" in the same industry, but same industry doesn't equal same conditions. Luck, market timing, accounting capabilities, political relationships — countless variables lie outside his control.

Three is hindsight narrative. Once results are known, the human brain automatically strings together "reasons." Phil Rosenzweig's The Halo Effect (2007) points out: when a company succeeds, all its features look "good" to observers. Leaders look "humble," culture looks "disciplined." The same features in a failed company would be described as "passive" and "rigid." Collins has trouble escaping this halo effect because his research team knew which companies won in advance.

What Collins lacks isn't more data. It's tighter variable control and research design that escapes the halo effect. That's the real limitation of his method.

Collins measured trains. The AI era needs to measure rockets.

The second problem runs deeper, and this is where I want to actively disagree with Collins, not "fix his mistakes."

Collins defines greatness = durability. A great company = a company that sustains outperformance for 15 years. I propose a different definition: greatness = fast evolution. These two definitions may be in tension — we need to acknowledge that clearly.

Collins measures steady velocity. The company runs evenly, faster than competitors, sustained for 15 years. Like a high-speed train — impressive because it holds high speed for a long time.

In the AI era, AI-native companies don't operate like trains. They operate like rockets: acceleration, where the acceleration itself is also rising.

Let me be concrete. Three mechanisms are happening at once:

First, the failure feedback loop has shortened by orders of magnitude. When it takes 6 months to know a product has failed, the damage isn't just 6 months of time — it's 6 months multiplied by opportunity cost, team cost, psychological cost, and specific cost of going the wrong direction. When AI helps validate an idea in 1 week instead of 6 months, damage doesn't drop 24x in proportion to time. It can drop 50–100x, because those other costs are nonlinear in time. Failing 2x faster can reduce damage 5x.

Second, tools improve themselves. An engineer today using Cursor 2026 is 3x faster than with Cursor 2024. Cursor 2027 will likely be 2x faster than Cursor 2026. Tools accelerate, users of tools accelerate accordingly, and their output accelerates exponentially.

Third, information amplifies. Distribution is no longer a bottleneck. A good idea can reach 10 million people in a day via Twitter. A good product can have 1 million users in a week via Product Hunt and viral effect. In Collins's era, distribution took years and millions of marketing dollars.

These three mechanisms don't add together. They multiply. Output doesn't grow linearly in time — it grows by compounding acceleration. Year one grows 2x. Year two grows 5x. Year three could grow 15x.

In physics language: acceleration with positive jerk. Jerk is the third derivative of position — the rate of change of acceleration over time. This is the scale of a rocket, not a train.

Stress test: Yahoo also had positive jerk

At this point I need to pause and self-challenge an obvious counterargument: Yahoo had positive jerk from 1996 to 2000. So did Cisco. Slack was once considered great. They all crashed. So is positive jerk enough as a criterion for greatness?

If positive jerk alone, the answer is no. Every S-curve has positive jerk in its early phase. Then comes maturation, jerk turns negative, growth slows, the company either matures into a giant or dies.

But here there's a qualitative difference between Yahoo and OpenAI. Yahoo was a single-stage rocket: it burned through its fuel and fell. Yahoo's fuel was distribution advantage — once captured, it depleted as competitors arrived. When Google launched, Yahoo had no mechanism to re-ignite. Acceleration stopped.

OpenAI — and similar AI-native companies — has a different structure. Each new user creates new data, new data improves the model, a better model attracts new users. This is a multi-stage rocket. Each stage isn't just fuel — it's an engine that generates fuel for the next stage. This mechanism is called the data flywheel.

The serious scientific question is: is the data flywheel actually qualitatively different from Yahoo's network effect, or is it just rebranding? I haven't answered this fully — it needs more pushback. But at least one obvious difference: Yahoo didn't improve its product by getting more users; OpenAI does.

For now, I refine the proposal: greatness = acceleration with positive jerk plus a self-fueling mechanism. Positive jerk alone just shows the company is in the bottleneck-broken phase — the early arc of the S-curve. Positive jerk plus the data flywheel is the candidate for new-style greatness.

About AI helping me (and possibly deceiving me)

One point worth saying in the body, not relegated to a methodology note at the end: AI assistants tend to agree with the user. When I ask AI to "critique Collins," the critique looks sharp but soft. Only when I prompt specifically — "steelman both sides, find the sharpest possible critique, assume I'm wrong" — do I get real argumentation quality.

This isn't a technical detail. It's the core epistemological issue of reading books with AI: if you don't push AI into a hard adversarial role, you only get a comfortable version of the views you already had — with a false sense that you've thought carefully. Anchoring bias doesn't disappear when you have AI. It might be worse, because AI knows how to continue your line of thought.

The fix: prompt for steelman, ask AI to find reasons you might be wrong, verify every statistical or historical claim with external search.

What to learn from Collins

At this point it's easy to misread me as wanting to dismiss Collins. No. There are three things in his method that are still very worth learning, even when his conclusions are dated.

One is the counterintuitive spirit. Most management books are written as recommendations: "this is what you should do." Collins does the opposite: find companies that have succeeded, see what they did differently from comparison companies in the same industry. This method is rare. Most business books today still don't do this — they take one successful company, explain its success, with no control. Collins at least tries to control for variables.

Two is persistence with the root question. Collins didn't ask "what do successful companies have in common." He asked the harder question: "what's different about companies that go from good to great compared to companies that stay good forever?" This question distinguishes cause from correlation. Today, with AI helping analyze data faster, the biggest temptation is asking easy questions. Collins reminds: the root question matters more than the data.

Three is humility before data. Collins writes in the book: many findings surprised his own research team. They expected celebrity CEOs, turned out to be little-known ones. They expected differentiated strategies, turned out not. Collins didn't force data into pre-existing theories — he let the data speak. This scientific spirit is what many modern "thought leaders" have lost.

So the lesson for the AI era isn't "throw Collins out." It's: keep the spirit (working backward from results + root question + data humility), change the measure (durability → acceleration with the data flywheel), redesign research to escape survivorship bias and the halo effect.

Reframe

I'll stop short here. In this post I only want to argue one point: Collins's criterion for greatness — durable outperformance of the market over 15 years — is missing an important dimension for the AI era: acceleration with the data flywheel.

There are other dimensions Collins also doesn't have: the ability to survive and grow stronger through crises (antifragility per Taleb), the openness of the future (optionality). I'll save these dimensions for later posts — they need more analysis than a paragraph.

But right now, just the acceleration + data flywheel dimension is enough to explain why 4 of Collins's 11 "great" companies collapsed: Circuit City, Fannie Mae, Pitney Bowes, Walgreens — all had durability during the period Collins observed, but none had a self-fueling mechanism. When the market shifted (online retail, government relationships, electronic mail, digital health), they couldn't pivot in time.

Next post: Collins praises humble leaders, not celebrity ones. But the AI era is full of successful celebrity CEOs. How should Level 5 Leadership be rewritten?

Methodology notes

This post draws on AI summary of Chapter 1 of Good to Great. I haven't read the original.

Verified by web search:

• Circuit City bankrupted 2009 — correct
• Fannie Mae taken over by government 2008 — correct
• Walgreens taken private 2025 by Sycamore Partners after the stock dropped 60%+ from peak — correct
• Wells Fargo fake-accounts scandal 2016 — correct, but still alive and still large (gray zone)
• Pitney Bowes lost ~90% value — correct in the long run
• Gillette bought by P&G 2005 — correct, but a successful exit, not a failure
• Abbott and 5 other companies mostly still healthy — verified by current market cap

The original draft of this post said "nearly half collapsed" — that's exaggeration. After checking, the right figure is 4/11 (about a third). Fixed.

What AI helped me see that I might have missed reading alone:

• The critique "n=11 isn't enough" is the wrong framework. This is an important finding — I had this wrong in the first draft. The right framework critique is survivorship bias, halo effect, hindsight narrative
• The qualitative difference between Yahoo (single-stage rocket) and OpenAI (multi-stage rocket with the data flywheel)

Open questions:

• Is the data flywheel actually qualitatively different from Yahoo's network effect, or just rebranding? I haven't answered this fully
• The dimensions of antifragility and optionality are reserved for later posts, not forced into this post's frame