The Smartest Guys in the Room — and why the smartest guys are always the last to notice

Books — what I read, and what I make of it.

I sell artificial intelligence. Not the stock — the consulting. I walk into mid-sized companies and explain to people that the machine will handle their processes faster, cheaper and better than the humans currently handling them. I believe it, too. Most days.

You should not read The Smartest Guys in the Room if this is your job.

Bethany McLean and Peter Elkind did not write a book about accounting fraud in 2003. They wrote a book about a company that became so convinced of its own brilliance that it eventually lost the ability to hear the word no — not from the market, not from its own balance sheet, not from reality. The fraud came later. The fraud was the symptom, not the disease.

Which is the uncomfortable part: the fraud is the one thing about Enron that doesn't apply to me.

What it's about

In the early nineties, Enron was a boring pipeline company out of Houston. Natural gas. Regulated markets, thin margins, men in hard hats. Then a McKinsey consultant named Jeffrey Skilling turned up with an idea that was genuinely brilliant: why sell gas when you can sell contracts about gas?

Skilling's "Gas Bank" turned natural gas into a financial product. Enron bought long, sold long, took the spread and the risk — and got paid for making a chaotic, deregulating market predictable. This was not vapour. This was a real innovation, and it worked.

From there it went almost vertically upward. Fortune named Enron America's "Most Innovative Company" six years running. For the year 2000 it reported revenues of $100.8 billion and ranked seventh on the Fortune 500. The stock hit roughly $90 in the summer of 2000. On the cover of the annual report, in enormous type, sat a single instruction:

ASK WHY.

On 2 December 2001, Enron filed for bankruptcy — the largest in American history at the time. Some 20,000 people lost their jobs, many of them their retirement savings at the same moment, because those savings were held in Enron stock. Arthur Andersen, the auditor, 85,000 employees worldwide, ceased to exist within eighteen months. The stock stopped trading at 26 cents.

Nobody had asked why.

The wrath of Achilles

The Iliad does not open with a war. It opens with a word: mēnis. Wrath. "Sing, goddess, the wrath of Peleus' son Achilles." And the whole epic is already contained in that first line: what nearly destroys the Greeks is not the Trojans. It is the wounded vanity of their best man.

Achilles is the best. Everyone knows it; he most of all. Agamemnon takes a war prize from him — an insult, not a threat — and Achilles withdraws to his tent and watches his own people die. He had a choice: a long, unremarkable life at home, or a short one with immortal glory. He takes the glory. And in the end he does not fall in single combat against an equal. He dies from an arrow in the heel, fired by Paris — a man no one in Troy took seriously.

Jeff Skilling is the Achilles of this book, and he is Achilles in every particular.

He really was the best. The most famous line in his biography comes from his admissions interview at Harvard Business School. The interviewer asks whether he is smart. Skilling, as McLean and Elkind tell it, replies: "I'm fucking smart." They took him. He graduated near the top, went to McKinsey, became its youngest partner, joined Enron, and within a few years was the man the company's cleverest people obeyed — not because of his title, but because they were convinced he thought faster than they did.

He had the wrath, too. On 17 April 2001, on an analyst call, a hedge fund manager named Richard Grubman asked for something entirely mundane: a balance sheet. Enron was the only major company that released quarterly results without one. Skilling — the chief executive of a public company, live, with hundreds of people on the line — answered:

"Well, thank you very much, we appreciate that … asshole."

That is Achilles walking into his tent. It was not a strategic decision. It was an insult absorbed. Someone had dared to ask the smartest guy in the room to show his homework.

Four months later, on 14 August 2001, Skilling resigned. Reason given: personal. The stock stood at around $40, half its peak. He later testified, under oath, that he had known nothing.

The armour was made of his own skin

Achilles gets his armour from Hephaestus, the limping smith of the gods — armour that makes him invulnerable because no mortal could have forged it.

Enron's Hephaestus was Andrew Fastow, chief financial officer, and what he forged were special purpose entities. LJM1, LJM2, Chewco — and the four that carry a name nobody would need to invent: the Raptors.

Technically, the design was elegant. Enron's books held a pile of investments whose values moved around. Moving values are bad for earnings. So Fastow built external vehicles with which Enron "hedged" those risks. The vehicle took the downside, Enron booked a stable profit, the analysts were satisfied, the stock went up.

One question: what was the vehicle funded with? What collateralised the hedge?

Enron stock.

Sit with that for a second. Enron insured itself against a falling share price — with a counterparty whose only asset was Enron shares. As long as the stock rose, this worked beautifully. It worked especially beautifully, because the insurance appreciated exactly when it needed to appreciate. But the moment the stock fell — which is to say, the exact moment you buy insurance for — the insurer's entire net worth evaporated and it could not pay.

It was armour made of the wearer's own skin. It protected against everything except the one thing it was built to protect against.

That is the heel. Not the greed, not the fraud: the structural dependence on its own share price, welded into the very instrument that was supposed to protect against a falling share price. When the stock began sliding in the summer of 2001, the Raptors collapsed, and on 16 October 2001 Enron had to announce a $618 million loss — and, mentioned almost in passing on a conference call, a $1.2 billion reduction in shareholders' equity.

The arrow was in the heel. Six weeks to bankruptcy.

Fastow is reported to have personally made around $45 million from the vehicles he ran while serving as CFO of the company on the other side of the trade. To permit this, the board had suspended the company's own code of ethics — twice. So the code of ethics existed. All 64 pages of it. They just switched it off temporarily, the way you tape over the smoke detectors during a renovation.

Mark-to-market: booking the future this morning

If one single accounting method killed Enron, it was this one. And you understand it best through a single deal.

In July 2000, Enron signed a twenty-year contract with Blockbuster: video-on-demand over Enron's broadband network. A contract for a technology that did not yet exist, serving a market that did not yet exist, with a partner who cancelled it eight months later. The service was piloted in a handful of cities. Essentially no customer ever paid for it.

Enron booked roughly $110 million in profit on that deal.

And here is the part people never want to believe: this was legal. The SEC had approved mark-to-market accounting for Enron in January 1992. The logic runs: if you sign a long-term contract, you may recognise the present value of the expected future profits immediately. For a liquid, exchange-traded commodity, that's reasonable — the market tells you what it's worth. For a twenty-year broadband contract covering a technology nobody has seen, nobody tells you what it's worth. So your own model tells you.

Internally, Enron had a name for this, and the honesty of it is almost touching: HFV — Hypothetical Future Value.

Now the mechanism this entire essay is about. Mark-to-market creates an addiction. Because you book the whole twenty-year profit in the first quarter, the second quarter arrives with nothing. The deal that made you a hero this month is an empty line next month. So you need a new deal. And a bigger one, because Wall Street wants growth, not repetition.

To hold its numbers, Enron had to dream a larger dream every quarter than the quarter before. Not out of malice. Out of bookkeeping.

That is not an accounting policy. That is a treadmill with a legal basis.

Rank and yank: how to train a company out of saying no

But the real heart of the book isn't the accounting. It's HR.

Enron ran an evaluation system officially called the Performance Review Committee and universally known as rank and yank. Twice a year, every employee was ranked 1 through 5 — not against objectives, but against each other. Whoever landed in the bottom band, roughly 15 percent, was "redeployed": two weeks to find another job internally, then out.

Consider what that does to people. It isn't merely harsh. It is a system in which your colleague's success increases your risk. In which withholding information is rational. In which the fastest route upward is not to do a good deal but to make a deal look good — because what got measured was booked profit, and booked profit, as we've established, was set by your own model.

Enron had optimised away its own early-warning system. A company where anyone raising a concern can expect to be ranked a 5 next cycle for being a naysayer will, in time, become a company where nobody raises concerns. That isn't a culture problem. It's a sensory problem. They surgically removed the nerve endings that register pain, and then wondered why the organism kept sprinting on a broken leg until the bone came through the skin.

The four corporate values, incidentally, were carved into the marble of the Houston lobby: Respect. Integrity. Communication. Excellence.

The arrow didn't come from a hero

Achilles is not killed by Hector. He is killed by Paris — the pretty, unserious brother, the archer, the man no warrior fears.

Enron's Paris was a 31-year-old Fortune journalist named Bethany McLean, and her arrow was a question requiring no finance degree whatsoever:

How, exactly, does Enron make its money?

She had read the annual report and hadn't understood it. That was the entire provocation. She called Enron. Skilling called her unethical on the phone — for daring to write a piece without having understood the company — and hung up. Shortly afterwards, Andy Fastow flew to New York with two other executives to sort it out in person. You do not dispatch three senior officers across a continent to answer a harmless question.

The piece ran on 5 March 2001. It was called "Is Enron Overpriced?", it was cautiously worded, it contained not one allegation of fraud, and its central finding was simply this: nobody she spoke to could explain how the company made money.

Nine months later Enron was bankrupt.

There were others. Jim Chanos, a short seller, had started betting against it in late 2000 — he had simply looked at the return on capital and noticed it was below the cost of capital. A company that gets back less than a dollar for every dollar it invests is not a growth story. It is a machine for destroying capital, however fast it spins. And there was Sherron Watkins, a vice president, who wrote to Ken Lay in August 2001 in a sentence that made it into every textbook:

"I am incredibly nervous that we will implode in a wave of accounting scandals."

All three were ignored or dressed down. Not because they argued poorly. Because by then Enron was an organisation in which the sentence "we no longer understand our own business" had become structurally unsayable.

And now the awkward part: the ego was mostly right

The essay could end here, morally tidy, finger raised. That would also be a lie — and a lie that teaches precisely the wrong lesson.

Because here is the truth: without Skilling's megalomania, there would have been no Enron. The Gas Bank was a genuine invention. EnronOnline, launched in 1999, was one of the first functioning electronic commodity exchanges in the world and cleared volumes the industry had considered impossible. Enron really did change energy trading. A substantial part of the innovation Fortune handed out six trophies for was not an illusion — it is the reason anyone believed the illusion that came later.

Goals that look achievable are not goals. They are plans. To state something nobody believes is possible, you need a certain insolence toward reality — an ego large enough to ignore the probabilities. Every company that ever built something new had, somewhere near its beginning, a person who was objectively wrong and carried on anyway. Moralise that away and you don't get a better economy. You get a slower one.

The problem is not the ego. The problem is that nobody ever taught it when to stop.

There is a point — and it is fiendishly hard to spot from the inside — at which the ego has finished its job. From then on it is no longer fuel. It is a bug in the control system. Skilling crossed that line somewhere around 1997 and failed to notice, because during precisely those years he was being applauded from every direction for failing to notice.

And that is the actual pattern here. It has nothing to do with energy trading. It happens to anyone who has won for ten years straight:

Stage 3 is the one nobody talks about, because it feels so harmless. It has no line on any balance sheet. It shows up as your standards for your own life quietly indexing themselves to your success curve — the house, the circle, the calendar, the ease with which you mistake your own opinion for expertise. And, most insidiously, as an entitlement about other people: how they ought to work, how they ought to think, how they ought to be. By the end, Skilling didn't merely believe he was right. He believed that people who functioned differently were a defect. He had, after all, built a performance review system on exactly that premise.

Stage 4 is then just bookkeeping. A man who mistakes his own invulnerability for a law of nature stops moving with the times, because he believes he is the times.

The memo you never want to write

In November 2025 something happened that I consider one of the funnier footnotes in financial history — and I mean funny in both senses.

Nvidia, then the most valuable company on earth, sent a seven-page memo to Wall Street analysts. Its content: a rebuttal of the claim that its accounting resembled historical accounting scandals. The press immediately gave it a name the company had not chosen:

the "We're not Enron" memo.

There is an iron law of business, and it goes like this: the moment you have to write a document explaining why you are not Enron, you already have a problem — though possibly not the one the document addresses.

And now I have to be honest, or none of this is worth reading.

Nvidia is not Enron. The AI industry is not Enron. The most prominent sceptic in the sector — Michael Burry, the man Christian Bale played in The Big Short, who has spent months accusing the hyperscalers of massaging their depreciation — put it most clearly himself:

"I am not claiming Nvidia is Enron. It is clearly Cisco."

That distinction matters, and it is why this essay is not an indictment. Enron sold things that did not exist. The GPUs exist. They are manufactured, shipped, racked, switched on, and they run flat out. Anthropic reported an annualised revenue run-rate of $47 billion in May 2026 — that is not Hypothetical Future Value, those are invoices that get paid. Nvidia's gross margin is around 75 percent. Cisco in 2000 had a real product, real customers and real profits — and the stock still fell about eighty percent, because it had been priced for a future that arrived three years late. That is a completely different accident from Enron's. It is simply not a more comfortable one.

What is transferable are the mechanisms. And that is where it gets uncomfortable again.

One: the armour made of your own skin

Enron's cause of death was the circular hedge — insurance collateralised by its own share price. So let us look at what the AI industry built in 2025 and 2026.

The clearest case is Anthropic. In November 2025, Microsoft and Nvidia jointly invested up to $15 billion in the company (Microsoft up to $5bn, Nvidia up to $10bn). In the same breath, Anthropic committed to purchasing $30 billion of compute on Azure — plus up to a gigawatt of Nvidia systems. Fifteen billion of investor money in; thirty billion of purchase commitments back to the same two investors.

The most elegant case is AMD. In October 2025, AMD and OpenAI agreed on 6 gigawatts of compute. In return, AMD granted OpenAI a warrant for up to 160 million AMD shares — roughly ten percent of the company — at an exercise price of one cent apiece. The tranches vest as OpenAI takes delivery of the chips and as AMD's stock hits escalating price targets. The final tranche requires AMD to trade at $600.

Read that again, slowly. AMD is paying its customer in AMD equity for buying AMD chips — and the value of that payment depends on AMD's stock rising, because the customer bought AMD chips. This is not fraud. It is fully disclosed, in the press release and in the annual report. It is merely reflexive: the collateral consists of the thing it collateralises.

And the best evidence sits in a filing. Nvidia's non-marketable equity securities — plainly: stakes in private companies, many of which are Nvidia customers — went from $3.24 billion in April 2025 to $42.34 billion in April 2026. Thirteen-fold, in twelve months. That figure is audited, signed and public. It is far less spectacular than the famous "$100 billion for OpenAI" — which, incidentally, never existed: that was a letter of intent from September 2025, and Jensen Huang personally buried it on 4 March 2026 ("not in the cards"). What actually moved was $30 billion of ordinary equity. Anyone still circulating charts with that $100 billion arrow is quoting a number that has been dead for four months.

The $42.34 billion is the better story anyway. Because it's true.

Two: mark-to-market is now called "backlog"

Enron booked twenty years of Blockbuster profit on the day of signature. Today there is a perfectly legitimate, cleanly regulated instrument for this: RPO — remaining performance obligations, contractually committed revenue not yet delivered. In September 2025, Oracle announced a $300 billion, five-year contract with OpenAI. The share price responded the way share prices respond to numbers like that.

Except: Oracle itself guides that it converts roughly 12 percent of total RPO into revenue over the next twelve months, and a further 34 percent across the two years after. Less than half the backlog becomes cash within three years. The rest is — and I use the phrase deliberately — hypothetical future value. Legal, audited, explained in the notes. And still a promise from a company that is itself burning billions over the same period.

In April 2026 the Wall Street Journal reported that OpenAI's CFO had warned internally that the company might not be able to pay for the capacity it had contracted. Oracle fell five percent that day, CoreWeave seven, SoftBank ten. Management called the characterisation "ridiculous". A month earlier, OpenAI had quietly cut its own compute target through 2030 from roughly $1.4 trillion to roughly $600 billion. That is the largest unwinding of a promise this industry has yet produced, and it happened almost without a sound.

Three: the argument about the heel

Enron's heel was its share price. The AI industry's heel is a number that four companies openly disagree about: how long does a GPU live?

Microsoft and Alphabet extended the assumed useful life of their servers from four years to six. Meta stretched it in stages from four to five and a half. The longer the assumed life, the smaller the annual depreciation charge, the larger the reported profit. It is the simplest earnings lever in all of accounting.

And then there is Amazon. Effective 1 January 2025, Amazon shortened useful life from six years to five — and gave its reason: "the faster pace of AI and machine-learning technology development."

This is the observation that matters, and it comes not from a short seller but from an audited filing. Four companies buy the same chips, from the same manufacturer, for the same purpose. Three say they last longer than we thought. One says they last less long. One of them is wrong, and nobody knows today which. Michael Burry calculates that the resulting under-depreciation runs to some $176 billion across 2026–2028 — his estimate, his assumptions, unaudited, and you should read it sceptically. But the disagreement itself is a fact, and it is sitting in the books.

The Bank for International Settlements — the central bank of central banks — put it rather more bluntly in its Annual Economic Report in June 2026. On circular financing, it wrote that the terms of these deals are

"typically poorly disclosed, with risks of the same asset being pledged multiple times."

The same asset, pledged more than once. That is not a blog comment. That is the BIS, in its annual report, in print.

The 20 cents I wish I had

And now the part where I have to correct myself.

I originally meant to put a number here, one that circulates in consulting circles — mine included: of every euro spent on AI, only about 20 cents is genuine productivity gain; the other 80 cents goes on cleaning up the errors the AI itself produced. It's a wonderful statistic. Punchy, counterintuitive, redolent of inside knowledge.

It exists, in fact. Exactly once.

In May 2026, a company called Entelligence published an analysis of over a million pull requests across 2,444 engineering organisations. The finding: of every dollar a team spends on AI coding tools, 18 cents becomes shipped product. The other 82 cents is consumed by the maintenance cycle those same tools accelerate. That is, to the cent, the number I wanted.

Now the catch. Entelligence sells — you're ahead of me — software that fixes precisely this problem. The error classification the entire calculation rests on was validated on thirteen organisations. There are no named authors, no confidence intervals, no peer review.

That is not a study. That is a sales deck with a p-value.

So the number stays out. An essay that condemns Enron for booking numbers that felt right cannot itself book a number because it feels right — least of all one whose sole source is selling the cure. That isn't modesty. That is the entire bloody point of the book.

What exists instead is duller and considerably sturdier.

In January 2026, Workday surveyed 3,200 employees at larger firms: roughly 40 percent of the time AI saves flows straight back into rework. Not 80. Forty — and even that is self-reported. The telemetry is harder. GitClear analysed 623 million code changes and found duplicated code up 81 percent since 2023, while the share of properly refactored code collapsed from 21 percent to 3.8. Faros AI did not ask 22,000 developers over two years — it measured them: throughput rose 33.7 percent, and simultaneously bugs per developer rose 54 percent and incidents per pull request rose 242 percent. Veracode found that 45 percent of AI-generated code samples contain an OWASP Top 10 vulnerability, and reported in spring 2026 that despite newer models there had been no improvement.

In short: AI produces more, faster — and a measurable slice of it has to be done again. The offset is real. It just isn't 80 percent, and anyone claiming it is, is quoting a vendor.

Mark-to-model

Here is where I caught myself a second time. It's embarrassing, which is exactly why it's staying in.

In July 2025, METR published a randomised controlled trial: experienced open-source developers working on their own repositories — no toy problems. Result: with AI they took about 19 percent longer. And the part that went around the world: the same developers estimated afterwards that they had been about 20 percent faster.

That figure has been every AI sceptic's favourite weapon for a year. I had it right here. It sat in this essay, presented as a current finding, in a piece about people recirculating stale numbers because they sound too good to check.

Because in February 2026, METR published a successor — and in it, effectively withdrew its own design. The pay rate fell from $150 to $50 an hour, recruitment collapsed, developers increasingly refused to work without AI at all, and 30 to 50 percent admitted to withholding tasks they didn't want to attempt unassisted. The new estimates carry confidence intervals that include zero. METR writes, in its own words, that it considers it likely developers are more sped up in early 2026 than its own 2025 figure suggests.

The "19 percent slower" gets quoted daily regardless. By consultants. By journalists. By me, a few hours ago.

This is the point where the book stops being a history book. Mark-to-model is not a disease of optimists. I booked a number that fitted my thesis, I didn't check whether it still held, and I would have printed it — in the very essay that convicts Jeff Skilling of booking numbers that fitted his thesis. The bear marks his book to model exactly as the bull does. He merely feels cleverer while doing it.

What survives from the METR study is still the most important thing in it — and it isn't the percentage. It is the gap. The developers were badly wrong about their own speed, in the direction that flattered them. The economists and ML experts polled beforehand were wrong in the same direction. Whether the true figure was −19 percent or zero: nobody in the room knew, and everybody was certain.

When a company rolls out AI and reports three months later that everyone is faster — what is that claim resting on? In the overwhelming majority of cases: self-report. A model. A feeling. That is hypothetical future value, booked in a skull instead of a ledger.

What the grown-ups say

If you want solid numbers, you have to leave the viral studies behind.

The notorious MIT study of August 2025 — "95 percent of GenAI pilots deliver zero ROI" — rests, if you read the appendix, on 52 interviews and a survey of 153 executives collected at four industry conferences. Kevin Werbach of Wharton has read the paper repeatedly and writes that he simply cannot see where the 95 percent comes from. I am deliberately not citing it as proof. I cite it as a symptom: that number went around the world because thousands of executives read it and thought "that matches what I see but am not allowed to say."

The study you should cite instead is one almost nobody has heard of. In February 2026, Nicholas Bloom, Steven Davis and colleagues (Stanford, Bank of England, Bundesbank, Atlanta Fed) surveyed roughly 6,000 senior executives across the US, UK, Germany and Australia. Findings: 69 percent of firms actively use AI — and nine in ten report no effect whatsoever on productivity or employment over the past three years.

Goldman Sachs, whose research desk was projecting seven percent of GDP from AI back in 2023, wrote in March 2026: "We still do not find a meaningful relationship between productivity and AI adoption at the economy-wide level." And supplied what I consider the single most revealing statistic of this entire cycle:

Of S&P 500 management teams, 70 percent discussed AI on last quarter's calls. 54 percent framed it as a productivity story. 10 percent quantified its impact on any use case.

One percent quantified what it earned them.

Daron Acemoglu of MIT has held the same estimate for two years and reaffirmed it in June 2026: around 0.55 percent in total factor productivity gains — over a decade. Not per year. In total.

And yet it works

Now the part that is equally true and that the sceptics turn the page on.

The best study in the field is Brynjolfsson, Li and Raymond, covering 5,172 customer-support agents, published in the Quarterly Journal of Economics — not on a blog. Result: +14 percent productivity. And the genuinely interesting finding: +34 percent among novices, and roughly zero among the most experienced staff. New agents reached in two months a level that otherwise took six. Attrition fell 8.6 percent.

Management Science published, in 2026, an analysis of three randomised field experiments with 4,867 developers at Microsoft, Accenture and a Fortune 100 firm: +26 percent tasks completed. Same gradient — the less experienced, the bigger the jump.

And the BCG/Harvard experiment with 758 consultants found, inside the model's competence, 25 percent more speed and over 40 percent better quality — and, just outside it, a 19 percentage point lower chance of getting the right answer. Because the consultants believed the model there too. The authors called it the "jagged frontier": a serrated boundary whose contour you cannot see from outside.

This does not contradict the macro data. It is the same finding, photographed twice. At task level, among novices, inside the frontier: 15 to 26 percent, measured, published, replicated. At firm and national level: close to nothing.

The gain is real. It just doesn't arrive in the accounts.

AI delivers real gains — but reliably in the places you didn't expect, and reliably not in the places you budgeted for. And the person least equipped to judge this is the person using the AI.

Between the task and the profit-and-loss statement lies a distance nobody has measured, and about which a very great number of people are talking very, very loudly. Seventy percent of S&P 500 management teams discuss AI. One in a hundred says what it brought in.

Enron had a performance review system that made that one percentage point unsayable.

What follows from this

The difference between Enron and the AI industry is real and it is large, and anyone who flattens it for a better headline is doing exactly what they accuse the other side of. Enron invented revenue. Nvidia ships chips that actually compute. There has not been a single confirmed writedown on AI assets at any hyperscaler. There has been no credit default. In April 2026, capital expenditure was not cut but raised — by all four majors, simultaneously. That is not the behaviour of people concealing something. It is the behaviour of people convinced they are right. And they may well be right.

But Enron did not die of fraud either. The fraud was the symptom. What died was an organisation so successful for so long that it had trained itself out of the counter-argument — and therefore, when it finally mattered, could no longer read its own balance sheet. The smartest guys in the room were not the last to notice because they were stupid. They were the last to notice because they were the smartest guys in the room, and because everyone in the room knew it, they most of all.

That is the only lesson I genuinely take from the book, and it is duller than I'd like:

The worth of an organisation is measured not by how good its best idea is, but by whether anyone inside it can still say "no" without paying for it.

Enron had a 64-page code of ethics and a board that suspended it twice. It had four corporate values in marble on the lobby wall and a review system that punished anyone who took them seriously. It had ASK WHY on the cover of its annual report, and it hung up on a 31-year-old journalist who did.

The question is not whether your company has good values. The question is what happens to the person who uses them.

And me?

I sell AI. I believe the technology is real, that the gains are real, and that most companies not starting now will regret it in five years. I still believe that, having read this book.

But I have started asking myself an unpleasant question when I walk out of a meeting where everyone was thrilled: who in that room could have contradicted me without looking stupid?

If the answer is "nobody", it wasn't a good meeting. It was a sale. And I should not flatter myself that I'm the only one who can see the difference — because that is exactly what the smartest guys in the room believed about themselves.

Achilles got the glory. He just didn't get very long to enjoy it.

Related

• Brave New World — and the discomfort of Mustapha Mond being right
• The Picture of Dorian Gray — can a book be immoral?
• Why good explanations fool us

Bethany McLean and Peter Elkind, "The Smartest Guys in the Room: The Amazing Rise and Scandalous Fall of Enron", Portfolio, 2003. AI financing figures from company announcements, SEC filings and the BIS Annual Economic Report of June 2026. Current as of July 2026.