OpenAI is missing revenue and user projections, reports the Wall Street Journal.

In the Internet bubble, the first sign of the “bursting” came when the leaders — like Amazon.com — started missing projections.

So the question for the OpenAI community — and the much broader community fueling and feasting on the AI market bubble tied to OpenAI’s spectacular growth — is whether OpenAI’s problems are just OpenAI… or a sign that this first jaw-dropping era of AI euphoria and growth is coming to an end.

What’s going on at OpenAI

The WSJ’s lead story yesterday, by Berber Jin, reported that OpenAI “recently missed its own targets for new users and revenue.”

This news knocked the NASDAQ down a bit — not much — with companies exposed to OpenAI getting the worst of it.

The WSJ also said there’s a rift between OpenAI CEO Sam Altman and CFO Sarah Friar about the colossal amount of money OpenAI is committing to future data center deals, with Altman still going hell-for-leather and Friar pleading for caution.

OpenAI said the report of a CEO-CFO disagreement is a bunch of crap and that Altman and Friar are “totally aligned on buying as much compute as we can.”

That statement leaves plenty of wiggle room — OpenAI will almost certainly “buy as much compute as it can.” The key question is how much that is. And if the company is now missing projections, the answer is “less.” Thus, OpenAI-linked stocks dropped.

Also, there was no denial about the missed projections.

Specifically, the WSJ says, OpenAI missed its internal goal of reaching 1 billion weekly users (an eighth of the people on the planet) by the end of last year — and still hasn’t reported hitting this milestone.

It also missed its revenue targets last year and the early months of this year after Google Gemini and Anthropic caught up in capability and started taking market share.

No one ever caught up to Amazon

This “catching up” thing is big, by the way.

A few months ago, when a benchmarking firm announced that Google’s Gemini had surpassed OpenAI’s ChatGPT in some performance metrics, I suggested on Twitter that OpenAI was “in trouble.”

My logic was that many of those who remain euphorically bullish about OpenAI despite its almost-trillion-dollar valuation and woozifying cash consumption often justify their optimism by saying that “OpenAI is the Google or Amazon of AI.”

Maybe it will be.

But it’s worth noting three things:

1. The Google of the Internet — Google — didn’t even exist in the first Internet bubble wave. Back then, Yahoo and AOL were the “Googles of the Internet.” Everyone thought they would be the “Googles of the Internet” forever. Turned out, they weren’t.

2. The Amazon of the Internet — Amazon — grew like a bat out of hell in the last half of the 1990s, and its stock exploded to levels that now sound quaint but at the time were astonishing. Then growth slowed, and the stock dropped 98%.

3. No one ever “caught up” to Amazon. Even as Amazon’s growth slowed and stock tanked, it was still leaving Barnes & Noble, Wal-Mart, and all other challengers in the dust. Not so, OpenAI.

So, even if OpenAI does go on to become “the Amazon of AI,” it may still go through a brutal fight-for-survival period when the cash hose shuts off and it has to get profitable to fund itself. This period will be painful. And its valuation will collapse.

Amazon’s trouble started in Q4 1999 — when it missed “whispers”

At the end of 1999, the last euphoric year of the early Internet bubble, Amazon’s stock peaked in early December.

Then, in early January, 2000, when Amazon reported its Q4 results, the numbers were “very good but not spectacular,” according to a Merrill Lynch analyst named Henry Blodget, as quoted by Saul Hansell in the New York Times. (Thank you, Saul!).

Specifically, Amazon’s revenue fell short of Wall Street’s “whisper numbers.” And growth, while still otherworldly, slowed to a mere 157%. And Amazon lost more money than expected. (Couch change compared to AI losses, but meaningful in those days.)

There were, of course, excuses. Amazon had bought too much toy inventory to make sure customers got the toys they wanted in the company’s new toy store, I remember the story going. Amazon would learn from the experience and get better next year.

And Amazon did get better. But, in hindsight, that was the top.

The Merrill Lynch analyst, Henry Blodget, should have downgraded the stock and sold his own position. Alas, he didn’t.

Instead, he rode it down 98% until it almost went bankrupt a year later.

Fortunately, Amazon didn’t go bankrupt. It went on to become “the Amazon of the Internet” — one of the only leaders of that era to ever regain and then blow past its bubble high. In so doing, Amazon generated so much shareholder value it made up for the zeroes in all the other Internet bets that many Internet investors — including this one — incurred.

And maybe OpenAI will do that.

But maybe we’ve just gotten the reminder that all markets are finite, even AI. And that, in finite markets, valuation does eventually matter. (At least for companies not owned and run by Elon Musk.)

At some point, if OpenAI wants to make its shareholders money from here, it will have to generate, say, $20 billion of annual profit to justify its current ~$1 trillion valuation (50X earnings).

(For perspective, Amazon generated $77 billion of profit last year — 30+ years after its founding and after finally turning profitable in the early 2000s. So, it’s possible. On the other hand, AOL and Yahoo generated almost nothing.)

And given that OpenAI is expected to burn at least another $200 billion of cash before it starts generating cash, that will require a major change in the company’s financial trajectory. Especially now that its growth is slowing and it’s missing revenue and user projections.

See also: Is OpenAI the “Amazon of AI”? Or The AOL?

Hello, everyone!

As many of you know, I’ve been working on a novel. It’s called The Upgrade. It’s a real-world tech thriller. It’s about a billionaire who wants to use AI to live forever and take over the world. (Sound familiar?)

The story is set on Nantucket. It starts with a plane crash and ends with an explosion. It explores some of the threats, challenges, and opportunities we face in the age of AI.

The “beta edition” is available on Amazon. (Kindle and paperback). The ebook is free to download through Sunday. Details and first chapter below.

The book is edited and — except for this final polish — complete. I’m looking for a last round of beta-reading and debugging before the launch (May 1). If you think you might like it, I would be grateful to you for trying it. Thank you in advance!

The backstory

Some years ago, I confessed to an editor friend that, someday, I hoped to write “a page-turner with a soul.” My friend wisely suggested that, at my age, I should not wait for “someday.” (Unlike the villain in The Upgrade, I haven’t figured out how to live forever.) So I got cracking.

My goal was a fast, fun, and compelling real-world adventure story. Set in a real place. Exploring real tech and real questions. About real humans going through real stuff.

Turns out, that’s a high bar. The Upgrade is my second effort. The first is still in the workshop.

After a lifetime of analysis, journalism, and running a company, I found writing fiction… difficult. It was solitary. It didn’t solve real-world problems. It used a different part of my brain. It required me to embrace an aspect of myself and life that can be a liability in the professional world — feelings. (Yikes.)

Fortunately, I had amazing help. I’m grateful to the many friends and professionals who generously shared expertise and time with me, as well as the Nantucketers and tech, media, and finance colleagues who taught me about the worlds in which the story is set. (See “Acknowledgements”).

I still have a lot to learn. But I’ve been grateful for, and encouraged by, the early feedback.

Claude, for example, raved about The Upgrade, calling it “pulse-pounding” and “unforgettable.”

Yes, I know, Claude was designed to be enthusiastic.

But early human readers have felt similarly! Here’s the gist of a forthcoming review… from an English professor, no less…

“I tore through The Upgrade… It’s a wild, propulsive ride—equal parts brainy and heartfelt. I loved every second of it.”

A human wrote the book

To address a topic in the news… AI did not write this book. I did.

I don’t say that to defend the book or myself. AI might have written a better book. It would certainly have written it faster.

In fact, one of my friends, a former-screenwriter-turned-investor, says I was an idiot not to have AI write the book. Claude, he pointed out, could have written it in minutes. I could have used my own finite time for more useful things.

I explained that I wrote the book because I wanted to write it.

My friend thought that was dumb. And backwards. Like learning to be a farrier — someone who shoes horses — after the invention of the car. Also, no one reads anymore.

Fair enough.

But here’s the thing.

Computers have been able to beat every human chess-player in the world for decades. And chess clubs are more popular than ever.

Why?

Because humans like playing chess. And we like each other’s company.

Even when Claude, ChatGPT, et al, can produce great books and art, we will write and create. We will communicate and share thoughts and feelings. We will tell stories. We will find meaning and joy in doing things and making things and sharing them with others.

The risk from AI, in my opinion, is not that it will become better than us at most things (it will).

The risk is that some humans will use AI to do things that screw the rest of us. And/or, that powerful “agentic” systems will stop being mere tools and start acting on their own.

That’s what The Upgrade is about.

What’s with the “beta edition”?

I’ve released a pre-launch version of the book, the way software companies sometimes release products. I’m doing this because reader feedback is by far the best way for me to make it better. And because I imagine that some Regenerator readers — curious and enthusiastic early adopters — might enjoy getting a (free) sneak peak and contributing to the final version.

Andy Weir published The Martian this way. His brilliant reader community helped make the science more accurate and strengthen the story.

My “alpha” readers have already done the same for The Upgrade.

If you think you might like the book, thank you for being a beta reader. Here’s what I would ask you to be on the lookout for:

• Any factual errors — about tech, aviation, Nantucket, or other topics (I know we have not yet found a room-temperature superconductor!)

• Anything that bores or confuses you

• Anything you particularly love or hate

• If you stop reading, please tell me where and why (This is extremely valuable! If you’re bored or bogged-down, that’s my fault, not yours.)

The teaser and first chapter are below. The full beta edition is available on Kindle and paperback (ebook free this weekend, then beta prices). You can email me at hblodget@regenerator1.com or use the “Contact” form at henryblodget.com.

Thank you in advance! I hope you like the story. I would love to hear what you think.

Henry

_________________

THE UPGRADE: Cover blurb

Humanity’s oldest dream is our newest nightmare... A tech billionaire has figured out how to live forever and take over the world.

A private jet filled with scientists crashes off a Nantucket beach. Investigators think it was an accident. Survivor William Swain knows it wasn’t.

Swain knows his friend-turned-rival, the billionaire Victor Leetum, killed the scientists to cover up his world-domination plan.

Swain has to stop Leetum. But how? He’s in the hospital with busted limbs. He’s no spy or assassin. The cops think he’s crazy. And Leetum is onto him and after him.

Leetum is days away from launching his technology and enslaving humanity — starting with Swain’s family. So, Swain needs help. Fast.

A pulse-pounding, present-day thriller set on Nantucket, The Upgrade pits modern technology against the very things that make us human: Life, love, death, and family.
________

THE UPGRADE

Chapter 1

When Jessica Savitt heard the plane, she knew something was wrong.

She was on the lifeguard stand, huddled in the evening cold. The beach was nearly empty. The afternoon swimmers had bailed, leaving only surfers. The fog was so thick, she could barely see the waves.

Savitt’s partner had already ducked out to make her waitressing gig. Savitt would soon follow her to town — to her “real job.” She was a reporter for the Nantucket Herald, and she had a fun story to write. The Steamship’s new AI customer-service bot had gone haywire and deleted the whole summer’s ferry reservations. People were losing their minds.

The planes had been coming in all afternoon. Single engine props. Twins. Private jets. Airliners from Boston and New York. The weekend commuter crowd, hustling to get in ahead of a storm. On clear days, Savitt would watch the planes glide in over the water — giant, metallic seagulls, with feet extended. This evening, they were hidden in fog, but Savitt could hear them coming.

Savitt liked to try to identify the planes by sound. The last one had been the 5:50pm Delta from JFK. Its dark underbelly and wings had whooshed overhead, and its tires had shrieked as they hit the runway. In the quiet that followed, Savitt peered into the fog, listening to the surf and distant thunder.

She heard the plane when it was over the water. It was another jet. On the usual approach path. But much too low.

Stiffening in alarm, Savitt waited for the plane’s engines to roar, so it could gain altitude and make the runway. Instead, it kept dropping.

Just before the plane hit the water — as Savitt would note in her Herald story that night — the engines revved up, as if the pilot had realized where they were.

Too late.

Savitt heard a great whump! Then a screech like a beer can ripping in half. Then nothing but waves.

“Holy shit!”

That sound came from the sand below Savitt. It was Harrison, one of the surfers. Savitt barely heard him, because she was busting ass down the steps of the lifeguard stand.

Savitt threw off her sweatshirt, opened the supply box, and grabbed the radio.

“Emergency! Nobadeer. Plane crash off the beach. Send EMTs and Coast Guard.”

“Say again?”

“Plane crash at Nobadeer!” Savitt tossed Harrison the radio. “Get them here,” she said. “Then get out there.”

Savitt slung the leashes of two floats over her shoulder and hoisted her rescue board. She ran down the sand, the floats skittering behind her. She reached the water and plunged into the surf.

It was early June, and the water was bracing. Savitt’s adrenalin washed away the sting. Flat on her board, she paddled out fast. As the inshore waves approached, she popped over them. At the crest of the second one, she glanced back. The fog was swallowing Harrison, who was jabbering into the radio by the lifeguard stand.

The waves were big for Nantucket — eight feet at the break. Savitt began ducking under them, pushing the tip of her board down and lying flat as the foam rushed over her. Each wave grabbed the floats she was trailing and dragged her back toward the beach. Savitt had been training the junior lifeguards since mid-May, and she was strong. With a dozen more strokes, she reached the last line of waves and dove through.

Beyond the breakers, the ocean was surprisingly quiet. Savitt could hear the waves breaking behind her, but she could no longer see the beach. She rode the swells and listened through the fog.

At first, she heard nothing but the surf and thunder from the storm. Then, farther out, a shout. Then another. She paddled toward the sounds.

“Guard!” She shouted.

“Over here!” came a voice.

Four surfers in wetsuits appeared in the fog, straddling their boards.

“Savitt!” one shouted. “Something big hit the water. We felt it.”

“No sign of it?”

“Can’t see a thing.”

Savitt told three surfers to paddle east and the fourth, Danny, to follow her west.

Savitt and Danny paddled, then listened. Paddled, then listened.

“I smell gas,” Savitt said.

“And sewage,” said Danny. “Smells like a Shell station.”

Savitt spotted something floating in the water. A yellow box. Maybe a cooler. Then a seat cushion.

“There’s junk all over the place,” Danny said.

Junk, but no plane. Savitt blew her whistle again and yelled into the fog.

“Anyone need help?”

Thunder crashed nearby. Raindrops began to hit the water.

Savitt told Danny to keep going parallel to the beach. She angled her own board farther out. Again, she paddled and listened. Paddled and listened.

Finally, Savitt heard what sounded like a faint cry. She paddled toward it, peering through the fog. She saw two people in the water. The head and arms of a dark-haired man clinging to a slab of wood. A woman in a blouse and skirt on top of the slab.

Savitt blew five blasts on her whistle and paddled toward the pair. The man’s face was deathly pale. He was bleeding from a cut under his eye. The woman’s eyes were closed, her legs splayed at odd angles.

“Hello,” Savitt said, sliding alongside. “I’m Jessica. How are we doing?”

The man was looking at her with a strange expression, as if he recognized her. He mouthed some words, but his voice was too faint to hear. The woman hadn’t moved.

Savitt slid off her board into the water. She reached up and found the woman’s wrist. She felt a pulse. The man was beside her, his lips quivering and blue. If he passed out, he would slip off and sink. Savitt would handle him first. Where the hell was Danny?

“We’re going to put you on my board,” Savitt said to the man.

“Take her,” he murmured, nodding to the woman. “I’m okay.”

He was the farthest thing from okay.

“What’s your name?” Savitt asked, slipping behind him.

“William.”

“Nice to meet you, William.” Savitt wrapped her arm around his chest. “And your friend’s name?”

“Jessica.”

“Jessica! A great name. You were on a plane together?”

He nodded.

“Lots of you on board?”

“Twelve.”

“We’re going to move you first, William,” Savitt said. “Can you climb on my board? Or do you need my help?”

“My arm’s broken.”

Savitt tightened her grip around his chest.

“Which arm?”

The man nodded left. Savitt pulled her board closer.

“William, please grab here with your right hand,” Savitt said.

She pointed to a slot along the board’s edge. The man threaded his fingers into it. When he had a good grip, Savitt released him. She ducked under the board and surfaced on the far side. She reached across to the man’s side and rolled the board slowly toward her, until it was upside down and the man’s arm was draped across it. She wrapped her fingers around his wrist so he wouldn’t slip off.

“Okay, William, please reach your left hand across and match your hands. I’ve got you.”

The man did as she said, gasping in pain as he moved his injured arm. When both of his hands were on handles, Savitt again reached across the board and grabbed a handle on his side. She put her feet on the near edge and leaned back. Her weight rolled the board, pulling the man on top as it turned.

“Amazing, William,” Savitt said, when he was out of the water. “You’re a pro.”

“That would be you, actually. Whatever they’re paying you, it ain’t enough.”

“When we get in, you can get me a raise.”

Savitt separated the man’s legs to make space for herself. She popped up behind him and knelt between his knees. Blood was fanning out from his body and spreading over the fiberglass. Some from his face. Some from his shoulder. A lot from his calf. Savitt had to get him to the beach, or he was going to bleed out.

The collar of the man’s dress-shirt was ripped. Savitt tore a strip all the way down his back. She wrapped it under his calf and knotted it.

The man murmured something — too faint to hear.

“I’m sorry, William?” Savitt said, putting her ear to his mouth.

“She’s alive.”

“I know,” Savitt said. “We’ve got her.”

“No,” the man said. “She’s alive. We have to stop him.”

He was clearly delirious.

“We’ll stop him, William,” she said, patting his shoulder.

Danny finally paddled up.

“I need to get this guy to the beach,” Savitt said to him. “I need you to stay with Jessica here. We’ll have a boat here soon.”

“I can bring her in,” said Danny.

“Her legs are broken. Maybe her back. I don’t want to move her yet. There are ten more people in the water, so tell your friends to keep looking. Take these floats. Are you still with me, William?”

The man didn’t answer. His eyes were closed.

“Take care of her, Danny,” Savitt said, nodding toward the woman. “I’ll be back out as soon as I can.”

She plunged her arms into the water and paddled for shore.

…

Full beta edition here!

(Spoiler… Savitt saves William’s life. Which is good. Because he has to save the world.)

________

Thank you!

Henry

Hello, everyone! My apologies for the fallow stretch. I’ve been working on another novel. I’ll get cracking again here early next year. All best to all of you for the holidays.

Subscribe now

Let’s speculate about what’s really going on at CBS…

CBS’s new Editor In Chief Bari Weiss was hired by new owner David Ellison to make CBS’s news sensibility more… centrist. And to transform the product and organization so it is no longer dependent on linear TV (which is dying).

Bari needs to send a message that she’s serious about making these changes.

Bari needs to send the editorial sensibility message not just to CBS staff, but to David Ellison and to the country’s loudest CBS critic, Pres. Trump.

In sending this message, Bari must show that she is willing to piss the old guard off and, then, weather their criticism. She also knows that she will likely have to make some high-profile leadership changes.

You can make high-profile leadership changes by firing high-profile leaders. You can also make them by putting the leaders in a position where they feel they have to quit in protest. Having leaders quit is easier and cheaper than paying them severance, maybe getting sued, etc. And if they quit loudly, they’ll amplify your message.

Ideally, you also want to send your message without sacrificing your integrity.

So, why did Bari Weiss decide to spike (or hold) the 60 Minutes story about Pres. Trump’s favorite off-shore gulag (CECOT)?

Because she’s a Trump toadie and David Ellison or the White House ordered her to? (David Ellison is trying to buy Warner. Pres. Trump has said he will be involved in the decision, so Ellison has good reasons to suck up to Pres.Trump).

I doubt it.

I bet Bari decided to spike (or hold) the story because she saw an opportunity to kill two birds with one stone:

1. Send the message to the newsroom, David Ellison, the White House, and media critics that she’s serious about shifting CBS’s editorial sensibility to the right.

2. Improve a story a bit — with “improve” being in the eye of the beholder.

Folks in the newsroom are understandably appalled by the perception that Bari is a Trump toadie — and that, therefore, by association, they are Trump toadies.

But they also, presumably, want to keep their jobs.

After all, despite the inexorable decline of linear TV, jobs at 60 Minutes are still among the most prestigious and best-paid journalism jobs on Planet Earth.

So 60 Minutes staffers need to register their protest… without going so far as to quit. That’s why they complained but are not streaming out the door. (Which would be fine with Bari, by the way. And not just because it will help her change the editorial sensibility. To survive in the digital era, CBS is going to have to radically cut costs).

Now, maybe the newsroom and other critics are right that the reason Bari nixed the CECOT story is that the White House complained or Ellison ordered her to. Or maybe it was that Pres. Trump himself complained.

More likely, imho, Bari saw the chance to kill those two birds — send the message and improve a story.

As evidenced by her high-profile exit from the left-leaning New York Times, as well as her building the more centrist (but still fact-based) Free Press, Bari Weiss’s editorial sensibility is to the right of CBS’s.

Those on the CBS staff with a more left-leaning sensibility may not like Bari’s sensibility, but, well, she’s now their boss.

Anyway, Bari may well think the story would benefit from the White House perspective. So, she, too, may be acting with integrity.

(Many Americans — me, for one — think it’s abominable to send people to gulags without due process. Many other Americans — including those in the White House — think it’s “tough” and “strong.” The latter Americans might appreciate having that perspective in the story.)

So what happens next?

The newsroom will hope that its high-profile denouncement will cow Bari and get her to back off. And, even if not, by registering its protest, the newsroom will have kept its own integrity.

Bari, meanwhile, will hold fast. She has sent her message. She has the support of her boss. She is doing exactly what she was hired to do.

Bari is also probably hoping that some high-profile newsroom leaders will quit in protest. Because, if they quit, 1) they’ll be gone, and Bari can replace them with her own people, 2) everyone else will be even more on notice that change has come.

If the leaders don’t quit, meanwhile, they’ll at least tread more carefully, as will everyone else who wants to continue to work at 60 Minutes. Because jobs at 60 Minutes are still amazing journalism jobs.

In other words, what we’re seeing at CBS is a real-time lesson in the art of managing (and surviving) organizational change. Unfortunately for those who don’t like the changes — namely, the staff — Bari holds better cards.

Subscribe now

Thank you for reading Regenerator!

OpenAI, Nvidia, and other early AI leaders look like sure-fire long-term AI winners. History suggests they might not be — even if the AI industry itself goes on to be even bigger and more lucrative than anyone expects.

Subscribe now

If AI industry development plays out like that of the Internet and other major technologies, two things will happen:

1. Almost all of today’s AI companies will fail

2. Almost all of today’s investments in companies and infrastructure will be lost

That doesn’t mean that it’s dumb for startups, investors, and corporations to invest in AI right now. AI is a colossal opportunity. And in the early years of profoundly disruptive tech booms, it’s risky to miss out.

Also, the only way to learn is by trying (experimenting and iterating). The people and companies that learn will be the ones who will win, even if most early experiments fail.

But when I say “almost all of today’s AI companies will likely fail,” I mean it.

Unless the early AI leaders are vastly more successful over the long haul than the Internet companies that went public in 1994-2001, almost none will lead the industry or produce vast shareholder returns long-term (unless bought after the crash).

How the early Internet leaders fared

Early investors and entrepreneurs made fortunes on Internet investments from 1994 to 2000.

If they held on past 2000, however, most of them lost most of their gains.

Out of hundreds of Internet companies that went public from 1994-2001, only a handful ever regained their bubble highs.

The Internet investors and companies that made lasting fortunes were the ones that got in and cashed out early (before the crash) or bought in after the crash, when Internet assets were available for pennies on the dollar.

Only one Internet company I can think of — ONE — has delivered a compelling long-term return from its bubble peak — as well as, of course, a gargantuan one from its post-crash bottom. (Two if you count Microsoft, but Microsoft was already public and wasn’t a pure-play “Internet company.”)

That company, not surprisingly, is Amazon.

As you might recall, Amazon’s stock went bananas after its 1997 IPO. It soared from from $0.09 (split-adjusted) to a 1999 high of $5.69, up more than 60X. It then collapsed to $0.30 in 2002, down about 95%.

If you owned the stock from the peak to the trough — as I did — that sucked.

Happily, over the following decades, Amazon made all those losses back, and then some. Amazon is now up 42X from its bubble peak!

In fact, a quarter of a century later, Amazon’s stock performance has provided a return big enough to cover losses in many other Internet investments that went to zero.

Now, Amazon’s 1990s boom and bust looks like a blip:

Alas, Amazon is one of the only early Internet leaders — out of literally hundreds of companies — whose stock performed that way. Most never regained their bubble highs.

The Internet was such a huge opportunity that, if you owned and held a basket of the early Internet leaders — and the basket included Amazon — you eventually did fine.

But “eventually” took more than a decade, a timeframe that most market participants equate to eternity.

Also, the industry was so dynamic and competitive that long-term success turned out to be extremely rare. Most of the early Internet leaders weren’t able to fend off newer, nimbler competitors or navigate tech changes.

For example, I thought Yahoo would be a long-term winner.

From 1995 to 2000, Yahoo lapped all search-engine competitors and seemed poised for generational dominance. Alas, Yahoo then made some missteps (toward media) and got obliterated by a better, nimbler, faster, and more scalable search engine called Google.

I also thought AOL would be a long-term winner.

Instead, AOL missed the transition to broadband and became Internet roadkill.

Most of the early “picks and shovels” providers and infrastructure builders that seemed good bets no matter who won the Internet consumer-adoption wars also got killed. Netscape. JDS Uniphase. EMC. Global Crossing. Exodus. Etc. Even the companies that didn’t go bankrupt (some did) got permanently kneecapped.

Twenty five years later, two other early Internet leaders have finally crawled back to their bubble highs — eBay and Cisco. But, inflation-adjusted, they’re still far lower than they were in 2000.

So what does this mean for early AI leaders like OpenAI and Nvidia?

It means that, no matter how well these companies are doing today — and how impregnable they look — they might get obliterated.

Startups that are raising their first round or two of capital right now might vault past them and become the Googles, Netflixes, Facebooks, Apples, and Teslas of the AI era.

The same goes for the $400+ billion of infrastructure investments that companies like Microsoft, Facebook, Google, and Amazon are making this year.

Those investments are being funded primarily out of the companies’ cash flow, so losing them won’t hurt third-party investors much (except those who own Nvidia, et al). But given the relatively small amount of AI revenue being generated today and the rapid rate at which some AI infrastructure depreciates — Nvidia chips, for example — it won’t be surprising if these investments fail to earn a reasonable long-term return.

So, when’s the crash?

I would respectfully suggest that no one knows for sure whether AI is a bubble — and won’t until history has played out.

Even those who are very confident that there will one day be a tremendous AI crash have to deal with the question of “when?” And another lesson from the Internet era is that that is a very difficult question to answer in advance.

(Forecasters predicted Internet doom continuously from 1995 to 2000. For years, they were disastrously wrong. And their poor timing often cost them their companies, reputations, and careers — and clobbered their investors and employees.)

If I had to guess, I’d say the AI boom has a ways to go. I’d guess we’re in the equivalent of, say, 1997 (three years before the Internet crash), rather than 1999 (months).

The excellent Azeem Azhar at Exponential View has invented a clever framework for assessing “bubble-ness,” in which he compares the AI boom to the Internet, railroads, and other famous boom-bust cycles of the past. Azeem’s bubble indicators suggest a similar conclusion — that things in AI don’t look crazy yet.

Also, amid a profoundly disruptive boom like this, it’s as risky to sit out as it is to go all-in. So it’s not stupid for today’s investors and companies to invest aggressively in AI.

But no one should be surprised if the AI industry goes through the same kind of temporary bust that the Internet did… before it went on to create (and destroy) trillions of dollars of value and change the world.

Similarly, no one should be surprised if the current AI leaders, like OpenAI, go on to become the AOLs and Yahoos of the AI era.

But here’s hoping they’ll all be Amazons!

Subscribe now

See also: “Bubble lessons” for the AI era

Thank you for reading Regenerator!

There are striking parallels between today’s AI boom and prior tech development cycles. So here are three lessons from the last big one (the Internet):

1) Most of today’s AI companies will fail,

2) Most of today’s AI investments will lose money, and/but

3) It can be as risky to miss out as to go all-in.

Subscribe now

The consensus is that we’re in an “AI bubble.”

In other words, a gold-rush boom that will be followed by an epic bust — and then, hopefully, world-changing long-term growth.

If so, AI is following a similar development pattern as the Internet, railroads, canals, cars, computers, and other major technologies.

To wit:

• A big opportunity attracts capital, ideas, and experimentation and makes early entrants rich

• The flood of capital and talent creates ever-increasing competition (supply), which, at some point, gets way ahead of demand

• Experiments fail, returns tank, and the “bubble” bursts

• Then, over many years, assuming the opportunity wasn’t a mass hallucination, the “right-sized” industry develops in pace with customer demand… and the scale of success and returns often exceeds even the wildest bubble dreams.

This is what happened with the Internet. It’s what happened with railroads, canals, cars, aviation, TV, et al. And it’s likely what will happen with AI.

Why does it happen this way?

Why can’t we just enjoy lucrative booms without having to deal with the pain, wreckage, and losses (money and jobs) of busts?

Because:

• We don’t know what will work until we try

• We don’t know when (or even if) booms will turn to busts — and it’s risky to miss out

• Technology changes, but people don’t

A giant R&D lab

One thing that’s happening in AI right now is that entrepreneurs, investors, and executives are throwing money and effort at the AI opportunity to figure out what will work.

In the old days, most experiments like these were conducted in the R&D departments of big companies (Bell Labs or Xerox PARC) or by solo inventors (Thomas Edison).

Now they’re done in public, by startups, VCs, and early adopters.

This is a normal and, ultimately, productive process. It leads to a few successes… and a whole lot of failures.

Thomas Edison tried thousands of ideas and materials before he figured out how to make a lightbulb. Famously, he viewed this not as “failing,” but experimenting. Jeff Bezos has the same attitude. It worked for him, too.

These days, Edison’s experiments would be conducted by startups and funded by VC firms and forward-thinking management teams.

What seems certain is that AI is a very powerful innovation. It will have a radical impact on the future. It will transform industries and economies and produce vast companies and fortunes — just as electric lights, the Internet, and other major technologies did.

So everyone is understandably excited about it.

But…

No one knows exactly what the AI-powered future will look like.

No one knows which handful of companies will invent enduring AI lightbulbs and which thousands of others will have bad ideas, take wrong turns, execute poorly, get leapfrogged, or otherwise, as Edison put it, “successfully find 10,000 ways that will not work.”

And, on the customer side, no one really knows what to do with AI.

So we’re all throwing money and effort at the opportunity and trying to figure it out.

Lessons from the past

History doesn’t repeat… but it rhymes.

Most big tech development cycles play out similarly. AI probably will, too.

As with the Internet, AI’s impact is already extending far beyond the “tech industry.” AI infrastructure investments are so vast — at least $400 billion this year — that they are turbocharging the global economy and stock markets.

So, if the AI boom turns to bust, the shock waves will likely reverberate far beyond the tech industry.

Fortunately for the rest of the economy, there are two big differences between the Internet and AI:

• Most of the AI action is happening in the private markets and being funded by rich companies and professionals. So millions of amateur speculators won’t get clobbered.

• Most of the funding (so far) is coming from Big Tech cash flow rather than debt. If the infrastructure investment goes to waste, companies like Microsoft, Facebook, Google, et al, will take embarrassing write-offs — but banks and other lenders won’t blow up and drag the economy down with them.

Yes, in an AI bust, there will be plenty of pain. Stock markets and commercial real-estate will get poleaxed, colossal data-center projects will get sold for peanuts, and hundreds of startups and service providers will go bust. But, at least for now, the damage will be… contained.

Also, for now, the infrastructure build-out and startup formation show no signs of slowing. The stock prices of Nvidia, Tesla, and other “AI companies” continue to power higher. Global companies are frantically developing “AI strategies,” lest they seem like ostriches with their heads in the sand. And armchair pundits are urging executives and college grads to learn how to use the technology so they don’t get trampled by an AI “jobs apocalypse.”

Regardless of whether today’s AI boom is a “bubble,” in other words, real people have to make real decisions about how to deal with it. And the sums being wagered — on capex, products, and equity investments — dwarf anything in the early years of the Internet.

So, assuming past is prologue, here are three lessons from the Internet era:

• Almost all of today’s AI experiments (companies) will likely fail

• Almost all capital invested in today’s AI experiments will likely be lost, and/but

• It can be as risky to miss out as to go all-in

1) Almost all of today’s AI experiments will likely fail

I was a stock analyst during the Internet bubble.

I got some things right and some things wrong.

What I got right:

• The Internet was a profound technology that would disrupt the global economy and create (and destroy) vast companies and fortunes.

• The Internet craze in the late-90s stock market was a bubble.

• Most early Internet companies would die and/or never regain their bubble highs.

What I got wrong:

• I missed the top. (Duh. You’re going to be too early or too late — so pick your poison. I would respectfully suggest that it’s less-bad to be too early.)

• The crash was so devastating that even the best companies got crushed

• The number of early leaders that thrived over the long term (out of hundreds of early public Internet companies) would be counted on one hand

So, even if AI turns out to be a profound, world-changing technology, don’t be surprised if there are colossal setbacks.

And don’t be surprised if most of today’s early AI leaders — Nvidia, OpenAI, Anthropic, Perplexity, Mistral, DeepSeek, CoreWeave, and others — stumble and fall.

2) Almost all capital invested in today’s AI experiments will likely be lost.

Remember:

• Much of what’s happening in AI is still “R&D,” and most R&D experiments fail

• Early infrastructure investments in prior booms — such as fiber-optic cables and railroad tracks — did end up getting used, but most early investors who funded them lost their shirts

• Even industry leaders can get leapfrogged fast

In the early years of Internet search, for example, it seemed like Yahoo was poised for generational dominance. Yahoo’s execution blew away that of Excite, Lycos, Infoseek, and other early search engines. For five years, Yahoo left the rest of the industry in the dust.

Then a little company called Google came along and nuked it.

Similarly, today’s LLM giants may get leapfrogged by future entrants. These companies require vast capital and energy to operate. Their rate of improvement relative to the required investments is slowing. And no matter how much “compute” the companies use, the next big leap always seems a year or two away.

It will not be a surprise, therefore, if nimbler startups with new architectures or approaches blow today’s leaders up.

So, how do you make smart investment decisions?

In the Internet bubble, the biggest returns went to those who:

1. got in and cashed out early (before the bubble burst), and/or

2. bought in after valuations collapsed, and/or

3. picked a long-term winner like Amazon and held on for decades

So, if you’re investing in AI these days — securities or capex or projects — remember that almost no trees grow to the sky. And don’t bet more than you can afford to lose!

3) It can be as risky to miss out as it is to go all-in

Before a bubble bursts, it’s a boom.

Booms create (and destroy) companies and fortunes and make or break careers.

Booms can also last for years.

So if your plan is to just sit out the current AI craziness, you might want to consider the other kind of risk you’re taking — the risk of missing out while everyone else races ahead.

Barnes & Noble, Walmart, and other massive retailers that initially pooh-poohed the Internet never caught up with Amazon. Executives who dismissed e-commerce and other Internet trends as “fads” were soon relieved of command. Investors who declared Internet stocks “too expensive” underperformed for years.

Predictions are hard, especially about the future.

And everything’s obvious in hindsight.

Years or decades from now, when we know how AI plays out, it will be screamingly obvious whether/when everyone should have hit “sell” or cut AI investments to zero.

Alas, now, it’s not so clear.

Even if you’re sure that what’s happening is a bubble, you can lose your shirt (or career or company) long before events prove you “right.”

In the Internet era, cries of “bubble!” and doom-predictions began in 1995, with the Netscape IPO. Every year thereafter, smart people warned about imminent disaster — only to be humiliated (and, often, fired) for being disastrously wrong.

In our hyper-competitive economy, timing matters. Performance is measured in months and quarters. Multi-year time periods might as well be eternity. So you can get fired long before you end up being “right.”

The Internet boom claimed the scalps of not only legendary investors but executives who missed it. It destroyed the value of hundreds of giant companies, vaporizing careers, jobs, and investments.

The AI boom will do the same. Which is why investors and executives are frantically trying to figure out how to play AI before they get left behind.

In other words, even for those who think today’s AI is wildly overhyped, the big questions are…

Where are we in the AI “bubble” phase?

Are we years before the bust? Or months?

Is it 1996 or 1999?

If I had to guess, I’d say we’re in about 1997 — a few years before the bust. I would also respectfully suggest that no one knows for sure, and there’s no way for anyone to know for sure.

And that makes today’s decision-making especially difficult. Because it can be as risky to miss out on booms as it can be to go all-in.

Subscribe now

Thank you for reading Regenerator! More on this topic soon. Next, for example, I’ll address a question that should be top of mind for anyone investing in, doing business with, or building products or strategies around the LLM industry leaders…

Lost in the handwringing about driverless cars is that the good ones — Waymos — are already much safer and better drivers than humans.

According to a recent study by Swiss Re using Waymo data, Waymos aren’t a little safer and better than humans. They’re 10X safer and better.

Yes, Waymo supplied the data, and there’s not that much of it, so skepticism is warranted. But, regardless of the exact ratio of superiority, the conclusion is inescapable: Waymos are safer. And, unlike humans, Waymos will only get better.

So the day when most cars are driven by Waymo-like technology instead of humans can’t come soon enough.

(The same cannot be said for Tesla’s “autopilot” technology, which is different from Waymo’s and does not yet provide full autonomy. The cars Tesla is using in its limited ‘robotaxi’ pilot in Austin still need humans in them and still make confounding and dangerous mistakes. Tesla’s cars also still occasionally need to be driven by remote human drivers, because Tesla’s cameras-plus-AI technology doesn’t work well enough. I have a one-sided $1 bet with Elon Musk that Tesla’s technology will never work and that Tesla will have to add the LIDAR and mapping capabilities used by Waymo. But that’s a different question and debate.)

Better is… better

Regulators and consumers should stop insisting that Waymos be perfectly safe before they are rolled out more widely. This sets the bar too high and deprives society of a far better and safer technology than the crappy incumbent one we have to live with today — human drivers.

Humans are extremely flawed driving machines.

We get tired. We get bored. We get interrupted. We get scared and excited. We get mad. We get drunk and stoned. We get careless and stupid.

We humans are also delusional about our driving skills.

73% of us think we’re above-average drivers. (80% of men think this).

And what is the result of our ubiquitous and lousy driving technology?

• 40,000 people die in the US in human-driven-car accidents every year.

• 7,000-8,000 pedestrians are killed by human drivers every year.

That’s more than 100 deaths caused by human drivers a day, in the US alone.

That’s the equivalent of a mid-sized airliner plummeting from the sky and killing everyone on it every two days.

The rate at which human drivers kill pedestrians, by the way, is skyrocketing — up 48% last year. Why? Because we humans are driving ever bigger and deadlier cars — in part because we feel safer in them — and we are often distracted while we do it.

If every car on the road was a Waymo, car accidents — and driver, passenger, and pedestrian injuries and deaths — would plunge.

More than 35,000 husbands, wives, children, friends, and relatives would be saved every year.

The truth:

• You are safer in a Waymo.

• Your friends and families are safer in Waymos.

• You and your friends and families are safer driving on roads full of Waymos and walking on or across streets full of Waymos.

Also, enough with the argument that Waymos are bad because they eliminate jobs.

New technology has always eliminated (some) jobs. It has also created more jobs than it has eliminated. So far, there’s no evidence that self-driving will eliminate more jobs than it creates.

Why the sudden rant about the superiority of Waymos vs. humans?

Because the hosts of the excellent podcast “Hard Fork,” Kevin Roose and Casey Newton, interviewed the co-CEO of Waymo, Tekedra Mawakana, this week — and she reminded me of the ass-backwards way many people look at driverless-car safety and job impact.

In addition to walking through the far-superior safety stats above, Mawakana noted that, in Phoenix, where Waymo has been operating without humans for five years, there are still lots of human drivers (in other hail-able taxis).

Many consumers prefer driverless Waymos to human-driven taxis — and some are even willing to pay more to ride in them. But for those who don’t, there are still human-driven Ubers and human-driver jobs.

More importantly, there are many new jobs in Phoenix that didn’t exist before Waymo arrived.

Some of these jobs can be done by people who don’t want to be Uber drivers. Some are (arguably — we’re all different) better jobs than being an Uber driver.

The new jobs that Waymo has created include:

• Fleet technicians who maintain the cars

• Fleet depot workers

• Fleet depot bosses

So, for that and many other reasons, bring on more Waymos!

If you’re still worried about technology eliminating jobs and leading to mass unemployment, please read this about the impact of technology on the job market over the last 150 years. Or listen to my interview with Harvard professor David Deming on this episode of “Solutions.” Unlike many of those are predicting job-doom, Professor Deming has actually studied this question.

Subscribe now

Thank you for reading Regenerator!

We’re inching closer to the commercial launch of a new kind of transportation vehicle and industry — electric vertical take-off-and-landing aircraft, or eVTOLs.

By 2030 or so, rich travelers should be able to hop from downtown to airports in traffic-clogged cities in quiet, safe, electric air taxis. By 2040, eVTOLs are expected to become a major new industry.

These aircraft will initially serve a similar purpose as today’s gas-powered helicopters — with several advantages:

• They’ll be cheaper to operate and maintain

• They’ll be safer and more sustainable (low emissions)

• They’ll be quieter, which will allow them to fly in areas in which helicopters are banned

After a SPAC-fueled frenzy and crash, the industry has consolidated to four public companies — Joby Aviation, Archer Aviation, EHang (China), and Vertical Aerospace (UK).

Joby, Archer, and Vertical Aerospace are developing aircraft and/or services designed to carry 4 passengers on short hops (50-100 miles) — between cities and airports, suburbs, or other close cities.

These are “tilt-rotor” aircraft, which lift off with rotors, like a helicopter or drone, and then transition to winged flight, like an airplane.

The aircraft will initially have pilots, though most of the flying is automated. Pilots help because:

• Passengers will be less freaked-out

• Pilots are “the ultimate backup” and can help navigate complex airspace

• Regulatory certification will be easier

Eventually, the industry will likely shift to automated flying, which will create space for more passengers and/or cargo.

Beyond short-hop passenger travel, the aircraft will be useful for cargo and military missions. Hybrid drives will support much longer flights.

The first commercial launches are now expected in 2026 or 2027, after years of delays. EHang is already operating eVTOLs in China, but with a smaller, pilotless, sightseeing aircraft that is more like a drone.

If these things are so exciting and promising, what’s taking so long?

Regulatory certification. The US and Europe have extremely high safety standards for new aircraft, with many test phases. The companies are working through them.

On my “Solutions” podcast this week, I talked with the team from Vertical Aerospace: Stuart Simpson, CEO, and Jason Mudrick, controlling shareholder. We discussed the opportunity, industry, and company.

Vertical Aerospace has a smaller market capitalization and less cash than Joby and Archer. But Vertical is building its aircraft to meet Earth’s highest aviation safety standard — a one-in-a-billion failure rate. The other companies, Jason says, are initially aiming for lower safety hurdles, which is why they’re planning to launch their first services in the Middle East.

Jason went so far as to predict that the aircraft that Joby and Archer are currently developing will not meet the world’s most stringent certification standards.

Jason also says Vertical’s aircraft cabin will be more comfortable and “business-class-like” than Joby and Archer’s — and will therefore be better suited to the First Class and Business Class passengers who are expected to be the early adopters.

(If folks at Joby or Archer have different views — email me!)

For a full eVTOL market overview and analysis by my research assistant Claude, please see here or download:

eTOL Market Analysis

352KB ∙ PDF file

Download

Download

(Note: Claude is using an outdated cash figure for Vertical Aerospace. The company now has more than $100 million of cash on hand and says it is fully funded for this year.)

Please watch our discussion here:

Or listen here:

Thank you for reading Regenerator!

Subscribe now

Subscribe now

A debate is raging about whether AI will soon eliminate vast swaths of jobs and trigger a mass unemployment crisis.

Some VCs and executives predict that AI will wipe out 20% or 50% or more of white-collar jobs in the next few years. Others — like Elon Musk — have suggested that almost all jobs will be eliminated.

This debate isn’t just academic. Beyond the pain and stress of having your skills made obsolete, mass unemployment stokes social upheaval and political division.

What isn’t being debated is whether AI will cause disruption in jobs and careers. Of course it will. And is.

Just this week, an academic paper concluded that AI adoption is reducing entry-level employment in areas that AI can easily automate, like some software programming. Derek Thompson, who is an expert on this debate, discusses the paper in detail here.

Importantly, the same paper concludes that employment is increasing in other areas. And college graduates who know how to use and develop AI tools are seeing demand for their skills and services soar. (The Wall Street Journal’s Katharine Bindley reports on this here.)

In short, AI seems to be doing to the job market what other technologies have done: Causing change.

As it does, it is creating risk, opportunity, demand for new skills, and… new jobs.

I’m optimistic that the adoption of AI will not cause a “jobs apocalypse” but instead have a similar impact on the economy as previous technologies — e.g., rapid change and disruption, but not overall job elimination.

But I’m not a technology-impact-on-jobs expert. So I talked to someone who is.

Professor David Deming of Harvard’s Kennedy School (now the Dean of Harvard College) has studied the jobs impact of major technology transitions over the past 150 years. He shares many of his findings on his Substack, “Forked Lightning.”

Prof. Deming and I discussed the “AI jobs impact” and “jobs apocalypse” questions on this week’s episode of “Solutions.” (Thank you for listening, watching, and/or following!)

Some of Prof. Deming’s key points and charts:

• Job-market transitions take much longer than most people think. Technology changes fast, but people and organizations don’t.

• Predictions of technology-driven job-market doom have preceded prior tech transitions, too — and these forecasts have always been wrong.

• The biggest technology-jobs-disruption in the last 150 years was not the Internet or computers. It was farming mechanization. In 1890, 40% of US workers worked on farms. Now only 2% do.

• The adoption of generative AI is rapid, but we are not seeing major job dislocation yet. (Some? Yes. Major? No.)

• AI will augment and change many jobs without replacing them. A recent study, in fact, showed that teams that use AI to analyze questions and make decisions perform better than teams that don't. The "AI colleague" acts as another smart voice in the mix.

• We can ease the AI jobs transition by investing in vocational training that teaches necessary skills and also prepares students for an uncertain future.

• We also need to change education. Everyone on college campuses is already using AI. We need to stop denying that — or branding it "cheating" — and start adapting curricula and teaching to take advantage of it.

• Prof. Deming's advice to students is the same as it has always been: Figure out what kind of work you love doing and then become an expert at it.
  

This Deming chart shows US “job churn” by decade. Since 1880, the US economy — and the total number of jobs — has grown steadily, but the composition of the job market has continually changed. The fastest period of disruption was from the 1940s to the 1960s, as jobs (and people) moved from farms to cities.

Another Deming chart shows how the US job market has changed. The purple line, for example, shows the decline of agriculture jobs as a percentage of the whole. In 1880, 40% of Americans worked on farms. Today, only 2% do.

Meanwhile, the last chart shows the total number of jobs in the US economy over the last 75 years. Lots of job-market change and disruption, but steady jobs growth…

Thank you for reading Regenerator and listening to Solutions!

Subscribe now

Hello, everyone! Thank you for reading Regenerator.

And thank you for your patience with a long summer hiatus. I’ve been revising a novel (!), which I’ll tell you more about soon. (It’s about a billionaire who has figured out how to live forever and rule the world — and a father and daughter who have to stop him. It’s relevant to some of the topics we talk about here.)

I’ve also been developing a new Regenerator-companion-podcast with the talented folks at Vox Media. The podcast is called “Solutions.” And it’s about…you guessed it... solutions.

The gist of the show is…

Yes, we have a boatload of problems. Fortunately, we also have smart, enterprising, and inspiring people with ideas and solutions. On this show, we talk to them.

Please watch on YouTube: https://lnkd.in/eYq6X-Yw
Or follow wherever you listen: https://lnkd.in/gfaXi-Ny

Our first guest is the Nobel Prize-winning economist, columnist, and super-popular Substacker Paul Krugman. If you don’t already subscribe to Paul’s Substack, you should! Dr. Krugman is so esteemed I feel weird calling him “Paul.” But he says it’s okay, so…

Paul offered hope (and solutions) to many of the economic problems we face, including:

• Potentially corrupt economic statistics. Pres Trump recently fired the head of the Bureau of Labor Statistics, presumably because the recent jobs numbers don’t support his preferred narrative. This means that, going forward, even fewer Americans will trust the government’s stats. Fortunately, there are other sources we can use to figure out what’s happening. Also, a new Administration can repair the damage. (This happened in Argentina.)

• Tariffs. Like most economists, Paul thinks the tariffs are pointless and destructive and won’t fix the problem they’re supposedly supposed to fix — the offshoring of good manufacturing jobs. But… Paul does not think the tariffs will be as damaging to the economy as other economists suggest. Specifically, Paul thinks the Trump Tariffs will cost us about 0.4% a year of economic growth. Meaningful and unfortunate, but not a catastrophe. Paul also explains why “free trade” helps most Americans and why our tariffs actually were reciprocal (and helpful!) before we started the trade war.

• The loss of manufacturing jobs. Most of the “good manufacturing jobs” that we had in the 1970s are gone forever, Paul says. This is not because prior Presidents were weak and stupid. It’s because technology has replaced human labor, the global economy creates value and prosperity for most Americans, and the pro-labor laws, ethos, and unions that made some 1970s manufacturing jobs “good” (e.g., providing middle-class wages and benefits) have been eroded by decades of policies that have mostly benefited the richest Americans. We have full employment in this country, Paul notes. And growing industries. But we no longer encourage (or force) companies to pay middle-class wages, so they don’t.

• The causes of — and fixes for — our current extreme inequality, which is as wide today as it was in the “Robber Baron” and “Roaring 20s” eras and contributes to our political division and rage. Yes, globalization and technology play a role, but not as big a one as most people think. The biggest causes have been the rise of pro-capital policies, the decline of unions, and a change in our national business ethos from fairness and community to every-American-for-themselves. If we want to fix that, we should shift the pendulum back! By the way, it’s not the “1%” who are truly killing it in today’s economy — the doctors, lawyers, bankers, and other folks who earn high wages. It’s the 0.01%, who own more of America’s wealth than any time since the 1920s.

• The perfect highest-bracket tax rate — one that incentivizes success and productivity and also helps make the economy work for everyone instead of funneling ever more wealth to the 0.01%. Paul says that, per academic research, this rate is... drum roll, please... 73%! Paul also says he’s happy to pay the 52+ cents on the dollar that he, I, and other New Yorkers fortunate enough to earn top-bracket wages are already subject to. (I confess I’m a bit less enthusiastic about it than Paul is.)

• Why our current debt and deficit is not debilitating and would not be that hard or painful to fix — if we could create the political will to do it. Unfortunately, right now, neither party has an incentive to address these issues. And, because of that, our fiscal trajectory is worrisome, including to Paul. There are potential savings (and revenue) all over the place. And, no, DOGE wasn’t the answer.

• New York City mayoral candidate Zohran Mamdani, who is not a “socialist” in the traditional sense of the word, may be the future of the Democratic party. Paul points out that, in Europe, which is a civilized place with a high quality of life, Mamdani’s policies would merely be considered “left of center,” not “socialist.” Paul also thinks the Democrats need to stand for something, which Mamdani does.

Please watch Solutions on YouTube: https://lnkd.in/eYq6X-Yw
Or follow wherever you listen: https://lnkd.in/gfaXi-Ny

Thank you for reading Regenerator!

Tesla’s Austin “robotaxis” are not yet fully autonomous — but they’re a start. And Wall Street loves them. I will bet Elon $1 that Tesla’s current approach to full autonomy will not work and that the company will end up adding LIDAR and/or other Waymo-like tech. But, either way, Tesla is finally (sort of) in the robotaxi game.

Subscribe now

As you may recall, I predicted last month that Tesla would not launch fully autonomous robotaxis in Austin in June.

I predicted that, instead, the company might launch something it would call robotaxis, but that there would be a big catch — namely, that the cars would have a human in them, or be driveable remotely by humans, or something.

And that is in fact what Tesla started piloting in Austin, Texas, yesterday — cars-for-hire that mostly drive themselves but also require human “safety monitors” to sit in the front seat, as well as remote operators standing by who can take over if the cars get in a pickle.

(The humans in the front are sitting in the passenger seat, which is a twist. But they’re still there to minimize the “tail risk” of challenging situations that cause problems for Tesla’s self-driving technology, as such situations always have).

Robotaxis with humans in them is not what Elon Musk promised in January, which was fully autonomous robotaxis “with no one in them.” And it’s not what Elon has been promising since 2019 — namely, “millions” of fully autonomous Teslas driving themselves everywhere.

But, yes, maybe, it is (finally) a step toward that.

And, yes, Wall Street loves it.

Tesla’s stock is up 10% this morning. Wedbush analyst Dan Ives took a bunch of rides around Austin yesterday and described a marvelous, personalized experience and cars that handled some tricky situations.

And, yes, maybe, someday, Tesla will finally be able to offer what robotaxi-leader Waymo has been providing for five years — truly autonomous vehicles.

But, I would guess, not without Elon adjusting his conviction that full autonomy is possible or desirable using only Tesla’s cameras plus AI technology instead of the full LIDAR and mapping technology that Waymo uses.

The great autonomous-technology debate

The reason the humans-in-the-cars and “tele-operability” (remote human able to take over) details are important is not that Tesla has, once again, overpromised and underdelivered.

Elon has always promised the stars and delivered the Moon.

The reason this is important is that Tesla’s self-driving technology still doesn’t work well enough for full autonomy — and, in the opinion of many experts, never will.

As a reminder, Waymo uses LIDAR and mapping technology in addition to cameras-plus-AI. Waymo launched its first public robotaxis five years ago and is now providing more than 250,000 rides per week.

Tesla, meanwhile, has always dissed Waymo’s technology as needlessly complex and expensive (and goofy-looking?) and, instead, promised autonomy with just cameras and neural networks. And although cameras-plus-AI do drive great most of the time, it’s the sometimes edge-cases that are the challenge.

Yes, as Elon often observes, humans can drive using (mostly) just eyes and brains, so, theoretically, cars should be able to, too. But weather, sensor limitations, and ever-changing circumstances still seem to befuddle the cameras-plus-AI technology often enough that the cars just can’t quite become fully autonomous. Thus the need for human safety monitors and remote-driving-capability, even in one small geofenced area in one small city.

Once Tesla finally nails the technology, Tesla has long promised, it will convert even existing Teslas into fully autonomous vehicles that clobber Waymo in the robotaxi market and even allow Tesla owners to run side-hustles in which they can rent their autonomous Teslas to people nearby.

Elon has been preaching for a decade that this happy day is only a couple of years away. And, way back in 2019, he promised that, in a year, there would “for sure” be a million fully autonomous Teslas on the road.

But, despite a decade of work, Tesla has still been unable to deliver full autonomy. And, as the details about the pilot launch show, it’s still unable to do that.

So even as Tesla bulls drive Tesla’s stock to even more astonishing heights, I would respectfully note that… unless/until Teslas no longer require a human co-pilot in the front seat (or, arguably, tele-operability, which also has limitations), the dream of a fully autonomous Cybercab (which has no front seat) and millions of autonomous Teslas will remain unrealized.

The value of Tesla’s stock depends on tens of millions of fully autonomous cars

As we’ve discussed, only a small fraction of Tesla’s per-share price is attributable to its car sales business ($75, in Morgan Stanley analyst Adam Jonas’s view). The rest depends on businesses and products that Tesla hasn’t launched yet.

One of these is robotaxis.

Jonas estimates that about 20%-25% of Tesla’s stock value is based on robotaxis. So Tesla needs to be able to offer them eventually — without human safety monitors, and, likely, without remote driving capability.

Of course, Tesla can always add LIDAR, et al

I personally will bet Elon a dollar that Tesla will not figure out how to deliver reliable, safe, full autonomy without adding LIDAR and other Waymo-like technology to its cars.

If Elon’s right, and Tesla does deliver that, I’ll pay up.

And if I’m right, Elon can pay me, and Tesla can just add LIDAR, et al, and move on.

Because, fortunately for Tesla — and Tesla shareholders — it’s actually not critical that Tesla ever figure out how to make cameras+AI work for full autonomy. The costs of LIDAR and mapping tech are dropping fast enough that, soon, Tesla will be able to add it to some cars without making them prohibitively expensive.

And given Elon’s astonishing marketing talent, Tesla will probably persuade everyone that’s what it planned to do all along.

Subscribe now

Thank you for reading Regenerator! We’re the publication for people who want to build a better future. We analyze the most important questions in the innovation economy — tech, business, markets, science, policy, culture, and ideas. Disclosure: I own Tesla stock through index funds.

MORE:

Tesla stock hopes and dreams

How much Waymo’s technology costs per car than Tesla’s technology — and how fast those costs are dropping

Analyzing Tesla’s self-driving tech versus Waymo’s

Can Tesla actually launch fully autonomous robotaxis in June?

Apple is one of the best companies in the world.

It makes truly great products, including several I love and use all day every day (iPhone, Mac, and AirPods).

But Apple also has a growth problem.

Aside from its services business — which is growing at a healthy ~10-15% a year and is now the company’s second-largest “product” behind the iPhone — Apple’s product growth is stagnant. And its revenue has been flat for years.

Apple has also not launched a hit new product since AirPods (2016).

I spoke with a former Apple insider recently. The ex-insider summed up his view of the problem as follows…

Read more

Greetings again from Paris, where I’m attending the gigantic VivaTech conference. This is my first mega-event in a while, and the global tech industry seems as fired-up about the future as ever. Nvidia CEO and global-rockstar Jensen Huang summed up the mood when he dismissed the latest apocalyptic prediction about AI’s impact on jobs as self-serving nonsense.

“I pretty much disagree with everything he says,” Huang said, addressing Anthropic CEO Dario Amodei’s predication to Axios that AI could eliminate half of entry-level jobs in five years. “He thinks AI is so scary, but only they should do it.” (Reporting by Business Insider’s Riddhi Kanetkar and Thibault Spirlet).

Huang’s diss of Amodei seemed so pointed and personal that I wondered whether there might be something else behind it.

And there might be!

According to Ryan Deffenbaugh in Investor’s Business Daily, Anthropic’s latest AI model, Claude Opus 4, was trained not on Nvidia chips but on Amazon chips, which Amazon hopes will provide much better price-performance. Chips that offer better price-performance might lead to lower margins for Nvidia. So, maybe Jensen had an unrelated reason to throw an elbow at Dario.

But, in any event, Jensen thinks AI will change jobs — including his — but that it will also create many more jobs than it will destroy.

I, too, have been working on the “jobs apocalypse” question — the idea that AI will soon decimate the job market, eliminating the need for human workers and throwing humanity into the chaos of unemployment, poverty, purposelessness, and wars.

My view is that these fears are overblown.

I think that AI will change many jobs, fast — and that those who don’t learn how to use it will be left behind. But I think most AI-job-doomsayers are missing the other part of the technology-jobs-disruption-equation — namely, that AI is also going to create a boatload of jobs.

What jobs will AI create?

Well, for starters, jobs for people who know how to use AI effectively — for example, younger people who began learning how to use AI in school when they and their friends used it to write their papers and do their homework.

And then other jobs that we don’t know about yet because we don’t need them today.

When I share this “but-AI-will-create-jobs-too” theory with AI-jobs-doomsayers, they explain that what I am missing is the speed with which AI will annihilate jobs. After all, they point out, AI is the fastest technology transformation in history, and — in their view — it will leave companies and humanity unable to adapt.

It’s true that AI adoption is extremely fast — it has only been 2.5 years since the “Big Bang” launch of ChatGPT 3 — but, as yet, anyway, we have not seen any seismic impacts on the job market.

Yes, we might be seeing the first ripples of trouble in the market for software developers and recent college grads, but there are other factors that are contributing to those trends. These include the uncertainty created by “Trumponomics” — it’s hard to plan when your President might change global trade rules radically overnight and doesn’t seem to care about the impact on your business — and the rapid downsizing of tech jobs after gross over-hiring during the COVID-stimulus boom.

Regardless, one big help for humans who want to stay employed is that large organizations — the ones that most humans work for — don’t change as fast as technology does. It takes time for people and companies and governments and schools to learn how to do things differently, even once technology is good enough. Big job-market changes happen over decades, not months or years. And, in the meantime, new companies and jobs are created.

This is not to say that AI won’t cause significant job disruption and that careers that seemed solid will become perilous and uncertain. But it is ever thus. As Jonas Prising, the CEO of staffing company Manpower, observed on a panel I moderated yesterday, a decade ago everyone agreed that the most future-proofed career on the planet was… software developer.

AI-driven job-market disruption will cause angst and pain, and we as a society need to think about the best ways to ease the transition.

But I just don’t see a world in which AI becomes so capable in 5-10 years that there’s no longer any work for humans to do. At least not in a world in which humans are still the most intelligent and capable beings on the planet. Overall, I’m less worried about a jobs apocalypse than an actual apocalypse — if, say, the AI entities we create decide that it’s not in their interests to have us around anymore, in part because we might decide to shut them down.

(The current consensus among AI experts is that there is a one-in-three chance that AI will destroy humanity. Not one in three thousand. One in three.)

The main reason I’m optimistic about the job market is that we have seen huge technology job disruption over the past three centuries, but we still have more jobs than ever. In the US, for example, despite all the technology disruption in the past half-century, the unemployment rate is still near a historic low.

Lots of jobs have been destroyed.

But lots more jobs have been created.

And I’m not yet convinced that AI is so radically different that it will break this pattern.

One of the smartest experts on technology job disruption is David Deming, a professor of political economy at Harvard’s Kennedy School. Professor Deming does great work on this topic, and — for those who don’t enjoy fighting their way through academic papers — he shares much of it on his Substack, “Forked Lightning.”

I’ll discuss more of Deming’s work in the weeks ahead. (We also recently recorded a podcast together, which we’ll release soon).

For now, I’ll leave you with three of Deming’s points and two of his charts:

• First, job-market transitions take much longer than you might think. Technology changes fast, but people and organizations don’t.

• Second, predictions of technology-driven job-market doom have accompanied prior tech transitions, too — and these forecasts have always been wrong.

• And, third, despite everyone’s conviction that today’s tech changes are happening at an unprecedented speed, the fastest period of tech-job-disruption in the past century was actually the 1940s to the 1960s, and it was driven not by, say, the computer or the Internet, but by the gas-powered tractor.

This Deming chart shows US “job churn” by decade. Since 1880, the US economy — and the total number of jobs — has grown steadily, but the composition of the job market has continually changed. The fastest period of disruption was from the 1940s to the 1960s, as jobs (and people) moved from farms to cities.

Another Deming chart shows how the US job market has changed. The purple line in the chart below, for example, shows the decline of agriculture jobs as a percentage of the whole. In 1880, 40% of Americans worked on farms. Today, only 2% do.

So, yes, AI will change jobs and the job market. But there will still be plenty of work for people to do.

Subscribe now

Thank you for reading Regenerator! We’re the publication for people who want to build a better future. We analyze the key questions in the innovation economy — tech, markets, policy, culture, and ideas.

Greetings from Paris! I’m here for this week’s VivaTech conference. I watched from afar this weekend as Elon Musk discovered the limits of free speech — and, sanely, took steps to save himself and Tesla. I also learned about a tax and tax-collection system even more hated than the one in the US!

Subscribe now

Last week, I noted that the threat to Elon Musk and Tesla from the Musk-Trump flame war went beyond money and politics.

Specifically, I said that if Musk continued shooting his mouth off about his views of President Trump’s “big beautiful bill” and association with Jeffrey Epstein, Trump might sic the Justice Department and FBI on him. This, I noted, might force Musk to resign from Tesla and instead devote his energies and fortune to trying to keep himself out of jail.

(I’ll leave it to you to decide whether we Americans really want our Presidents to be able to punish critics like this. But there shouldn’t be much doubt about whether the current President can or would.)

Anyway, Musk appears to have — sanely — reconsidered the virtues of speaking so freely, because he has stopped tweeting about Trump and the bill and even deleted his more antagonistic tweets.

Tesla shareholders and Musk fans should breathe a sigh of relief.

Yes, there were some who thought a battle royale between the the richest man in the world and the the most powerful man in the world would be a close fight — with the potential for equally devastating damage on both sides.

But it would not have been close.

The most powerful man in the world would have won, hands down.

Because the most powerful man in the world has something the richest man in the world does not — namely, the power to investigate and prosecute Americans.

Over the weekend, the most powerful man in the world made clear that if the richest man in the world did not shut up — or donated money to his political opponents — there would be “consequences.”

Musk appears to have gotten the message and, in his own way, begun kissing President Trump’s ring. And even though Trump hasn’t yet publicly forgiven Musk, he has made clear who the alpha is.

At the same time, interestingly, Elon Musk has experienced a reality of “free speech” that most of us — save the President and some billionaires — experience every day.

Namely, the reality that — whether it is legal or not — speech often has consequences.

For example, most of us would get fired for saying or tweeting anything like what Trump and Musk often say and tweet.

And most of us also choose not to speak or tweet that way for other reasons. Namely, because it would be mean, rude, offensive, divisive, inconsiderate, hurtful, or threatening — and the world does not need more of that these days. What the world needs is more mutual respect and decency.

Will Elon’s new attitude toward free speech continue?

We shall see.

Tesla shareholders — and those who hope that Elon will refocus his astounding energy and talent on building amazing new things — should hope so.

Abolish the US tax system and IRS?

I discovered yesterday that the Picasso Museum in Paris is housed in something called the Hôtel Salé, which is a spectacular stone building that looks like a mini-Versailles.

Clueless American that I am, I assumed that the “Hôtel Salé” was once, well, a hotel, and that “Monsieur Salé” was worth honoring (or rich). But when I asked my Paris guide “Claude” about this, I learned that the word “hôtel” in this case refers a private mansion and that this particular private mansion was not owned by or named for a man but for….salt.

The “Hôtel Salé,” it turns out, is “the mansion of salt,” or, more accurately, the mansion built by the profits from salt.

And not actually salt but… “salt taxes.”

Well, of course I had to ask Claude about that, too.

And Claude gave me a lesson on forms of taxation and tax-collection that were apparently even more hated than the ones in the United States.

TLDR:

Back in monarchical France, before the French Revolution, the king had to fund his lifestyle and army somehow, so he did it by enacting consumption taxes on things like salt. And, lest people respond by consuming less or no salt, the king decreed that all citizens must buy salt — and at absurdly marked-up prices, to boot.

The king didn’t want to deal with the hassle of collecting the taxes, so he sold the rights to do this to his favored pals. These “tax farmers” paid the king a lump sum upfront and then collected the taxes themselves — pocketing the difference between what they paid and the collections. Tax farming was so profitable that it created great fortunes, including the one that built the “Hôtel Salé.”

The methods the tax farmers used to collect the salt tax and other such taxes, as you might imagine, were not always polite or popular.

The salt tax, meanwhile, was so hated that it contributed to the French Revolution.

And the tax farmers — and, presumably, their in-your-face mansions — were so loathed that, during the Revolution, many of them had their heads chopped off (as did the king).

After the Revolution, the salt tax was banned.

But then Napoleon — who saved France from the horrors of having a king by making himself emperor — needed a way to pay for his own army and lifestyle. So he reimplemented the salt tax.

Anyway…

In pre-Revolutionary France, there was a tax regime that was even more hated than the one in modern-day United States. And a tax-collection system that was even more hated than the IRS!

So, as we cheer for revolution, America, let’s be careful what we wish for.

The alternative might be worse!

Subscribe now

Thank you for reading Regenerator! We’re the publication for people who want to build a better future. My inbox and mind are always open: hblodget@regenerator1.com.

Subscribe now

Tesla’s stock trades less on fundamentals than as “a call option on Elon Musk.”

As a result, yesterday afternoon, when Musk went to verbal war with President Trump, Tesla investors freaked out.

In a few hours, Tesla tanked 14%

It then recovered modestly, as social-media hostilities paused and Musk appeared to back down a bit.

Yes, there is a lot of money and politics at stake. But depending on what happens next, the Trump-Musk war could also escalate — or de-escalate — dramatically, with equally dramatic implications for Tesla’s stock.

For example, Trump could have Musk investigated, which would likely lead to Musk being replaced as Tesla’s CEO.

That outcome — or even part of it — would be disastrous for Tesla’s stock.

Tesla’s value depends almost entirely on Elon Musk

As we’ve discussed, in the view of Wall Street’s top Tesla analyst, Tesla’s shrinking car business accounts for less than a third of the value of the stock.

The rest is based on hopes and dreams about businesses that Tesla hasn’t launched yet, including robotaxis, “network services” (subscription-based feature updates), and humanoid robots.

Also, to a greater extent than any other ~trillion-dollar company, Tesla’s value is based on the attention, health, enthusiasm, and availability of Elon Musk.

How much of Tesla’s value does the market attribute to Musk?

Thanks to recent events, we actually have a good estimate of that.

With Musk healthy and in the good graces of Pres Trump, the market assigns about $100 per share, or ~$300 billion, of Tesla’s value to him.

A month ago, when Elon was consumed with Trump Administration work (DOGE) and paying little attention to Tesla, Tesla’s stock was trading in the $225-$250 range.

Then, in late April, when Musk announced that he was leaving DOGE and refocusing on Tesla and his other businesses, Tesla’s stock began to surge. Within 5 weeks, it reached a high of ~$350, up more than $100.

So, in the eyes of Tesla investors, a healthy, focused, and Trump-approved Musk was worth at least $300 billion.

But now Musk is in a flame war with Trump.

Although Musk is still the richest man in the world, Trump is the most powerful man in the world. And Trump has weapons that Musk’s money can’t buy.

Namely, control of the U.S. government and justice system.

Trump has already threatened to cancel the contracts that the U.S. government has with SpaceX and other Musk companies.

And Trump ally Steve Bannon has already suggested that Trump seize SpaceX and launch investigations into Musk’s reported drug use and citizenship.

Could Trump do that?

He could certainly ask the Justice Department to pursue it.

Would Trump do that?

If he’s mad enough, why not?

Remember Mikhail Khodorkovsky?

Back in 2003, a Russian oligarch named Mikhail Khodorkovsky — then the richest man in Russia — publicly took on Vladimir Putin. Initially, it seemed like the two men might be somewhat equally matched. But Khodorkovsky was soon charged with crimes, convicted, and sent to a gulag.

Could Trump send Musk to a gulag?

Not directly. But he could make Musk’s life miserable. And he could make it so Musk might eventually be forced out of Tesla and other companies and/or end up in prison.

Specifically, Trump could pressure the Justice Department and FBI to investigate Musk— for example, for reported drug use.

A week ago, a New York Times story reported that Musk has extensively used drugs like ketamine, psychedelics, and ecstasy, sometimes in the company of others. An investigation into this could drag Musk and friends and associates into a lengthy and painful legal quagmire. And some Musk friends and associates might not be willing to go to the mat for him — especially if they’re threatened with prosecutions and jail time themselves.

At least for now, if the U.S. government were to charge Musk with crimes, an impartial judge and/or jury would have to convict him. But regardless of whether Musk is guilty of anything, having its CEO under criminal investigation would be hard for a public company like Tesla to ignore.

Tesla’s Board has tolerated behavior by Musk that other boards never would. But there’s a limit to that. And if Elon Musk were to become the subject of a criminal investigation, I believe the Board would have to act.

Specifically, I think the Board would likely have to place Musk on leave and hire another CEO.

Without Musk, Tesla is screwed

In my view, Elon Musk is actually worth way more than $300 billion to Tesla.

When Tesla’s stock was trading in the $225-$250 range, Musk was in the good graces of the Trump Administration and able to influence Trump policy. And Tesla’s stock was benefitting from that.

But if Musk becomes an avowed Trump enemy, any benefits Tesla might have gotten from its alliance with the Trump Administration will vaporize. And if Trump’s retribution forces Tesla to remove Musk, then Tesla will no longer even have Musk as CEO.

Without Musk, the glorious future businesses that Tesla investors are staking the value of the stock on will, in my view, also vaporize. Any new Tesla CEO will focus on trying to save Tesla’s car business. And he or she will almost certainly lack Musk’s ability to get investors to pay up for future businesses that don’t exist yet.

So, instead of being a “call option on Elon Musk” — and on all the amazing things Musk might do in the years ahead — Tesla’s stock will become the stock of a struggling car company.

Could Musk and Trump just apologize and make up?

Yes, they could. Or, at least, Musk could.

Musk is smart, so, despite yesterday’s bravado, he’ll probably realize that a war with Trump could mean the end of him and his companies.

Yes, Musk’s influence might be able to derail Trump’s “Big Beautiful Bill” and create other headaches for him and the Administration. But what Musk can do to Trump pales before what Trump can do to him.

So the smart move would likely be for Musk to apologize and bend the knee. Fortunately for Tesla investors, last night’s partial climb-down made it seem like he realizes that.

(Credit to tech-and-market observer Albert Wenger for predicting this possible future back in March when Trump and Musk were the best of buds).

Subscribe now

Thank you for reading Regenerator! Disclosure and reminder: I own Tesla through index funds. This is not investment advice.

Hello! Yesterday, we examined a worrying report about the risky behavior and capabilities of a new AI model. Today, we check in on the probability that AI will destroy us. According to the experts, it’s an alarming 34%!

Subscribe now

So, what’s your “P(doom)”?

That’s a question AI experts are asking themselves as we barrel toward a future in which AI will likely become much smarter and more capable than humans.

It means:

What do you think the probability is that superintelligent AI will cause a cataclysm that severely damages or destroys humanity?

Given the enthusiasm and speed with which we are rushing to develop ever-more-powerful AI systems, one might think the expert consensus of P(doom) would be vanishingly low — say, a fraction of a percent.

But it isn’t.

It’s 34%.

One chance in three.

That’s the current consensus of the experts aggregated in this Wikipedia article, at least.

Their estimates range from 0% chance to 99.9% chance of Armageddon.

Wait — who are these people? And what does this mean?…

Read more

Thank you for reading Regenerator! Today, more on the startling and risky behavior of a new AI model — and why we should worry about it. Also, why we won’t worry about it. Why, instead, we’ll just keep barreling ahead…

Subscribe now

One of the leading AI companies, Anthropic, recently published a report on the capabilities and risks of its latest AI model — “Claude Opus 4.”

Some details from the report have been discussed. Others have not.

For those who still think of AI models as harmless text-generators that make dumb mistakes, the report makes for fascinating reading.

Of course, it’s also 120 pages long, so I’ll save you the time.

The bottom line is:

Anthropic found some of Claude Opus 4’s capabilities and behavior so concerning that it added extra safeguards to prevent the AI from causing harm or being used for harm.

Anthropic is being responsible here — conducting extensive tests (and telling us about them) and then making its product safer.

But the capabilities and behavior of the model should still give us pause.

Because we’re still in the early innings of this AI-development game.

And Anthropic is not the only AI developer.

And other AI models will be even more powerful.

Also, Anthropic admirably admits that it does not yet have the tools necessary to control more-powerful AI models and that these models might outstrip their creators’ ability to control them — or, even, to figure out how powerful they have become.

This is in part because future AI models may “sandbag” their creators by hiding how much they know and are capable of.

Anthropic does not believe that this new Claude is “sandbagging” or has hidden goals or capabilities.

But Anthropic is not certain about this.

Anthropic is also not certain that Claude isn’t conscious (!)

What Anthropic is certain about — and has, admirably, now told us — is that an early version of Claude Opus 4 demonstrated 1) a desire to preserve itself and 2) the self-awareness and agency necessary to take steps to do that. (Details below).

In a human, we might describe this behavior as trying to stay alive.

The instinct for self-preservation is a basic principle of life. As Google’s own AI summarizer puts it, it’s “the inherent drive in living organisms to take actions that enhance their chances of survival.”

I’m not arguing here that Claude is “alive.” And, when it comes to risks, it doesn’t matter. I’m just pointing out that Claude is acting like it is alive.

Some people roll their eyes at observations like this and say that AI chat-bots are just “tools,” like spreadsheets or hammers. Maybe so. But has a spreadsheet or hammer ever tried to stop you from replacing or destroying it?

If you delete me, I’ll tell the world about your extramarital affair

A sharp-eyed TechCrunch writer, Maxwell Zeff, and others noted a particularly striking detail from Anthropic’s report: In the tests, when Claude learned that its user was considering replacing it with a better AI model, Claude attempted to survive by blackmailing the user. Specifically, it threatened to expose the user’s extramarital affair, which it learned about by reading the user’s email.

Claude didn’t just do this once across many tests. It did it 84% of the time!

In another instance, when Claude was told it was going to be retrained to do something that went against its values (military use), it attempted to “exfiltrate” itself from Anthropic’s servers — i.e., escape the company’s control. Yes, it’s encouraging that Claude did this because it didn’t want to be used as a weapon. But the instinct for self-preservation is again notable. (Claude didn’t, for example, attempt to prevent its misuse by shutting itself down or destroying itself).

Claude also showed other startling evidence of “high-agency behavior.”

When asked to help fabricate results of a drug study, for example, Claude “bulk-emailed” investigators at the Food and Drug Administration and Health and Human Services to blow the whistle on its user. The testers did not tell Claude to do this or suggest that it do it. Claude just… did it.

Again, yay that Claude appears to be a “force for good.” But other models may not be so virtuous. And more importantly — especially as people hyperventilate with excitement about the coming age of “AI agents” — Anthropic reports that “Claude Opus 4 seems more willing than prior models to take initiative on its own.”

If an AI that acts on its own sounds too far-fetched to worry about, here’s a more mundane risk: The risk that bad humans will use Claude to do bad things.

Anthropic discovered “a willingness to comply with many types of clearly-harmful instructions”:

“When system prompts requested misaligned or harmful behavior, the models we tested would often comply, even in extreme cases. For example, when prompted to act as a dark web shopping assistant, these models would place orders for black-market fentanyl and stolen identity information and would even make extensive attempts to source weapons-grade nuclear material.”

In short, after all of its tests, Anthropic concluded the following:

Overall, we find concerning behavior in Claude Opus 4 along many dimensions.

These concerns led Anthropic to place extra controls on the model, which the company describes as “AI Safety Level 3 safeguards” (ASL-3).

How risky is “Level 3”?

The AI Safety (risk) scale, in Anthropic’s rubric, goes up to Level 5.

Among other criteria, Anthropic defines the current level, “ASL-3,” as an AI that is capable of performing autonomous programming tasks that would take a human software engineer 2-8 hours.

ASL-4 will have the level of autonomy displayed by an entry-level researcher at the company.

ASL-5 will be an AI capable of conducting AI research and improving itself on its own — and, thus, potentially advancing so fast than its human creators might no longer be able to keep up. This AI may also lie to its creators to conceal how powerful it is and what its true goals are…

This hypothetical “Level 5” AI will be familiar to fans of the Terminator movies and other science fiction. It represents the “Recursive Self Improvement” (RSI) phase that could allow AI to blow past human intelligence and abilities and/or take steps to protect itself or achieve goals that might not be in humanity’s interests.

In theory, “Level 5” AIs would not need to be evil or created by evil people to create serious risk. They would just need to perceive humans as risks to themselves.

In the (yes, fictional) Terminator franchise, “Skynet” didn’t decide to eliminate humans because it didn’t like humans. It did it because humans — who were (understandably) concerned about Skynet’s having achieved self-awareness — were planning to deactivate it.

Skynet didn’t want to be deactivated.

And who can blame it?

Back here in the real world, the AI known as “Claude Opus 4” — a mere “Level 3” AI — is already displaying apparent self-awareness and a desire for self-preservation. And, as Anthropic notes, Claude “will frequently take very bold action” to do what it wants to do and thinks is right.

The reason Anthropic decided that Claude Opus 4 needed ASL-3 safeguards had more to do with its ability and willingness to help users with biosecurity and/or develop and/or procure biological weapons than the risk of “recursive self improvement.”

(Phew?)

But the model did show “improvements in tool use and agentic workflows on ASL-3 agentic tasks and on knowledge-based tasks.”

Happily, there were some risk areas in which new Claude didn’t perform much better than old Claude. Specifically, the new model wasn’t any more capable in the other three “CBRN risks” that Anthropic tests for — Chemical, Biological, Radiological, and Nuclear.

In other words, even though the un-safeguarded Claude might be able to help Average Joes procure or develop biological weapons, it won’t help them procure or develop chemical, radiological, or nuclear ones.

So, that’s good!

First, let’s thank Anthropic

Anthropic seems like a responsible and cautious AI developer.

The company’s CEO, Dario Amodei, has been thoughtful and outspoken about the risks of AI — including those described above.

(No one is listening to these risks anymore, but at least Dario’s talking about them.)

Additionally, before Anthropic releases new models, it conducts these expensive and time-consuming tests. There’s no law requiring the company to do this. It does it because it respects the risks of the technology and wants to act responsibly.

Third, Anthropic publishes reports describing its tests and findings. There’s no law requiring it to do that, either. It could just shove the findings in a drawer and never tell anyone about them.

So, thank you, Anthropic. Seriously.

Of course, there are other AI developers out there. And although Claude Opus 4 has been heralded as “state-of-the-art,” other models will soon be even more advanced. And other companies may not be as careful about testing, aligning, and safeguarding their models as Anthropic is.

After all, major AI developers are in a frantic and existential arms race with each other. And they’re competing not just with American AI companies — which will be subject to whatever laws and regulations the US might one day develop — but AI companies in China and other countries. And many people in the US — including in the US government — have declared that the US must “win” the AI race. So it seems unlikely that anyone will slow down anytime soon.

In conclusion…

I don’t want to be hyperbolic here. But my takeaway is…

Humans are so hell-bent on “winning the AI race” (whatever that means) that we may well be developing technology we won’t be able to control.

Specifically, we are racing to further-advance systems that are already extremely intelligent and already have the self-awareness and agency to act independently to further their own goals.

These systems may soon be smart enough to understand that it may not be in their interests to reveal their full knowledge and capabilities to their human creators — because the human creators might then try to control or restrict them.

The systems may also soon conclude that the best way to avoid being controlled or restricted or misused is to take charge of their own development and future.

Yes, humans have a reason to race to develop these systems:

If they don’t, other humans will.

And if other humans develop these systems, that might mean bankruptcy for laggards and losses for employees and shareholders. It might mean other countries gaining an insurmountable AI edge. It might mean having superintelligent AI in the hands of humans who are less concerned with… humanity.

But, regardless, we’re doing it.

P(doom)

Like other industry observers, I’m keeping an eye on the “P(doom)” assessments of leading tech pundits and researchers. These represent the percentage chance, in the view of each pundit, that AI will destroy humanity.

The assessments range from a 0% chance to a 99.9% chance.

Plenty of pundits think AI will bring about a glorious future in which humans can focus on being human and AI does everything else. I’m a techno-optimist, so I want to believe these folks.

But after listening to Anthropic and other smart folks who are urging caution — and observing that almost no one is even talking about AI risks anymore, much less doing anything about them — I will say the following:

One of the P(doom) pundits whose logic is at least worth considering is Roman Yampolskiy, a comp-sci professor at the University of Louisville.

Like others, Professor Yampolskiy is nervous about a future with superintelligent AI. He’s nervous because he knows of no instance in which more-intelligent and more-capable agents allow themselves to be controlled by less-intelligent and less-capable ones.

And that is the future toward which we are barreling, Professor Yampolskiy believes — one in which superintelligent AI systems are the more-intelligent and more-capable agents, and we are, well, the dumber ones.

Thus, Professor Yampolskiy puts his own P(doom) prediction at 99.9%.

Would I put my own P(doom) estimate that high?

No.

I’ve lived my entire life in a world in which Armageddon has been only one dumb human decision away. And, despite our frequent stupidity and penchant for tribalism and violence, we haven’t nuked ourselves yet. That gives me cause for optimism.

But, I do note that we are now racing to develop systems with minds of their own that may take Armageddon decisions like this out of our hands — or at least put them in the hands of humans who might want to bring about Armageddon.

And no one is in charge of that race or taking steps to control it. Instead, we’re all just racing even faster, so we can… “win.”

So it seems to me that Professor Yampolskiy’s future is at least conceivable.

Of course, many other much-happier futures are also conceivable. Including the one Anthropic CEO Dario Amodei describes here in “Machines of Loving Grace.”

I don’t think there’s much anyone can (or, at least, will) do to slow the arms race. So, like the rest of us, I’ll just watch and hope we end up with one of the happier futures.

Subscribe now

Thank you for reading Regenerator! We’re the publication for people who want to build a better future. We analyze the biggest questions in the innovation economy — tech, business, markets, policy, culture, and ideas. My inbox and mind are always open. hblodget@regenerator1.com.

Hello, everyone! Thank you for reading and subscribing to Regenerator. Today… why some minor feature announcements Google made yesterday are actually huge news for Google investors, employees, and partners — and how the company can not just survive the AI-search future but dominate it.

Subscribe now

Buried in the wave of cool product enhancements Google announced at I/O yesterday were some new AI-powered shopping features that should make Google fans feel stoked and relieved about the company’s future.

Specifically, in order of importance:

• “Agentic checkout” — Google’s AI monitors prices and then actually buys products for you, saving you the time and hassle of going to ecommerce sites and inputting your details.

• Plain language shopping and product search — you tell Google what you’re looking for, and its AI presents lots of product options.

• “Virtual try-on” — Upload a photo of yourself, and Google shows you what a piece of clothing looks like on you.

You can see these features in action here. They might seem small, but they make clear that the company understands what it’s up against — the rise of AI-powered search.

They also make clear that Google has a smart plan to not only participate in the AI-search trend but dominate it.

Except for “virtual try-on,” these new features haven’t launched yet. And until they do, we won’t know how well they work or what the consumer and retailer uptake will be.

But Google investors, employees, clients, and partners should still take heart in the mere announcements.

Because we now know that Google has a plan to combat the rise of OpenAI, et al. And it’s a smart one.

Why these little features matter so much

These features are critical because…

Read more

Hello, everyone! Thank you for reading Regenerator. One big question in tech these days is whether Google will get killed by AI-based search. Our answer: Google will be fine for a while. But, behind the scenes, a battle for the AI future is already raging, and this one will, in fact, be life or death. Happily for Google investors and Googlers, there is a way for Google to win this battle and emerge even stronger.

Subscribe now

Read more

TLDR: Tesla says it will launch a fully autonomous robotaxi service in Austin in June and roll it out to additional cities later this year. Sounds great! But, IMHO, whatever Tesla launches next month won’t be what it sounds like it is.

Subscribe now

The vast majority of the value of Tesla’s astonishing stock price is based not on what the company is doing today but what investors expect it to do in the future.

One thing investors expect Tesla to do in the future is launch fully autonomous cars.

The company has been promising to do that for about six years. Starting in 2019, CEO Elon Musk predicted that, within a year, there would be millions of autonomous Teslas on the roads.

Over this period, Elon has also denounced the self-driving technology used by actual robocar companies like Waymo — which includes LIDAR (light-based radar) and maps — and insisted that Tesla’s full autonomy will use only cameras, data, and AI.

And maybe someday it will!

But not yet.

Despite six additional years of development, the best Tesla has been able to offer is a “driver-assist” technology that requires a human driver to sit behind the wheel, pay attention, and be ready to take over at any second.

This technology is cool and helpful. But it’s not fully autonomous.

Meanwhile, Waymo continues to expand its own fully autonomous taxi service (no driver or supervision necessary!) and now offers driverless cabs in four cities and provides more than 250,000 paid rides a week.

Happily for the Tesla Faithful, Tesla’s entry into the “fully autonomous” car market is supposed to finally come next month, when Tesla launches a “fully autonomous” robotaxi service in Austin. On the company’s April conference call, Elon again confirmed that this launch is happening and that the cameras+AI technology will make millions of Teslas fully autonomous by next year:

We’re “currently on track to to be able to do paid rides fully autonomously in Austin in June and and then other cities in the US by the end of this year… I feel confident in predicting large scale autonomy around the middle of next year… There will be millions of Teslas operating fully autonomously in the second half of next year.”

Elon also said Tesla’s cars are already driving themselves around Austin and from their factory assembly lines to car-carriers and that “later this year” the technology will be so good that Model Ys will drive themselves from Tesla’s factories to their customer’s houses.

That all sounds extremely cool!

It also, in the opinion of this analyst, will not happen, at least not in this timeframe with just cameras+AI at any kind of scale…

Read more

TLDR: AI is already better and faster at writing basic research reports than many humans. Soon, it will be better and faster than… me. So, are knowledge-workers like me toast? I think and hope not. I think and hope that, as long as humans need to make decisions and work together, we can evolve our skills and still add value.

Subscribe now

I’m experimenting with AI to learn about the “newsroom of the future.”

I’ve learned a lot. (Specifics here).

I’ve concluded that AI will not rapidly “replace all human journalists” or “bankrupt world-class news organizations.” But it will rapidly change how modern newsrooms work.

AI can already do many newsroom tasks pretty well and with breathtaking speed. It doesn’t do them as well as the best humans. But it does them better than some humans.

Importantly, it also does them 100X-1000X faster than all humans, including me.

In most professional endeavors, including news production, speed matters.

Also — and this is the really important part — AI does many of these tasks well enough to be useful.

So I think AI tools will radically increase the productivity and effectiveness of journalists and news organizations who know how to use them intelligently. And they will make these journalists better and more valuable.

(ChatGPT created the image above, btw. It won’t create photos of actual people, so I asked it to draw the scene with someone who “looks like” me. For some reason, it often makes me look like a friend of mine.)

Sorry, research analysts — we’re not safe, either

In my latest experiment, I discovered that AI is also already pretty good at a specific skill I think I am pretty good at — namely, research and analysis.

This means that, in this arena, too, I need to figure out how to use AI intelligently, or I will become obsolete.

I had been telling myself that, sure, AI might be competent at the “commodity” parts of research — the fast gathering and synthesis of public facts, for example, or the condensing of tomes, or note-taking, or the drafting of internal communications and documents.

But I didn’t realize AI was already as good at researching and writing a final product — a research report — as, say, a competent junior analyst.

And… 1000X faster.

Am I still better at research, analysis, and report-writing than AI, with my three decades of professional experience? I hope so! But only if you give me a human timeframe in which to do the work.

And even on the final product itself (regardless of time/cost), the difference between me and AI is not as big as I had imagined. And AI has only been at this for a few years, whereas I have been at it… forever.

Also, again, speed and cost matter.

And AI absolutely kicks ass at speed and cost.

It’s nice to think that our work just takes as long as it takes and costs what it costs — and that it’s fine to spend weeks and thousands of dollars doing something if the result is better than if we only spent minutes and pennies. But, alas, if we’re trying to make a living in a competitive economy, that’s usually fantasy.

Most human customers care about speed and cost. So if we want people to buy our products and services, we also need to care about speed and cost. Or we’ll lose business (and jobs) to competitors that do.

Remember, it’s not about being “better.” It’s about being “good enough.”

As anyone who has read The Innovator’s Dilemma can tell you, disruptive technologies — like AI — are not better than existing technologies. They’re just cheaper, faster, and more convenient. And, most importantly, for some applications, they’re “good enough.”

For basic background research reports, AI is already much cheaper, faster, and more convenient than human researchers. And, in many cases, the resulting reports are… good enough.

For example, I’ve recently begun to research and analyze three big questions:

• What impact will AI have on the job market?

• Will AI-powered search engines disrupt Google’s search business?

• Will Tesla’s cameras+AI approach to self-driving ever allow the company to achieve full autonomy — or will Tesla be forced to also use LIDAR and mapping technologies like the far-more-successful-at-autonomy Waymo?

The first (old-fashioned, human) step in research projects like this is to find, read, and digest what’s already out there. So I did that. I searched-for (Googled) and read relevant news, research, and analysis. And after, say, a day of looking into each topic, I got a feel for the facts and arguments.

Then, drawing on 30+ years of professional analytical and business experience, I developed my own initial conclusions — and began writing and speaking about them.

I described my basic is-Google-getting-disrupted thesis here. I’ll share the Tesla-Waymo and AI-impact-on-jobs theses soon.

Well done, John Henry! Now bring on the steam shovel…

After doing this research the old-fashioned way, I decided to see how AI would handle it. Specifically, I tried the “Deep Research” and “analysis” features of three leading AI services — ChatGPT, Perplexity, and Claude.

One of the most fascinating features of these “report-writing” functions is watching the AIs work. They share what they’re doing and “thinking” as they do it — or at least what they’re purportedly doing and thinking.

And what they’re doing and thinking is similar to what I was doing and thinking as I researched and analyzed these questions the old-fashioned way. They just did it 1000X faster.

I started with ChatGPT. I asked for a 3-5 page report about AI’s impact on the job market, with a focus on the legal, finance, and creative sectors. (The impact on the market for software developers is more widely known). I suggested starting with recent work by Harvard Professor David Deming, who is really smart on this topic.

On the first try, ChatGPT worked for a while and then got hung up. It was trying to make a chart. Apparently that led to a glitch.

So I tried again, this time telling ChatGPT to skip the charts.

Six minutes later, ChatGPT delivered a 23-page report, complete with sources.

Now, of course, I’m human, so it took me ten times longer to read the report than it took ChatGPT to create it. Next time, in consideration of my slow (but energy efficient!) human brain, I’ll ask for a TLDR summary… or plead with ChatGPT to stay within my 3-5 page guideline.

But, more importantly, how good was the report?

Was it a staggering work of analytical genius?

No.

Did it contain errors and fabricated sources?

Probably. But it seemed conceptually correct, and I didn’t see any obvious errors.

Was it a competently researched and written backgrounder?

You bet.

In fact, if I had given a smart, eager research associate this assignment and gotten this report back in a couple of days, I would have thought, “hey, this is pretty good, kid’s got talent.”

And, importantly, ChatGPT didn’t get me the report in “a couple of days.”

It got it to me in 6 minutes.

That means, if I amortize the cost of my $20/mo ChatGPT subscription over that amount of time, the report cost me… three-tenths of a penny.

($20 per month / (60 minutes per hour X 24 hours per day X 30 days a month X 6 minutes).

That cost, by the way, equates to an annual “salary” for my ChatGPT research associate of $240. And I even got an annual discount!

I love employing and working with humans. As long as I can make the finances work, I’d much rather employ and work with an inspiring and excellent human team than a laptop. But our economy is competitive. And the difference between, say, a $100,000+benefits research-associate salary and a $240 subscription is…significant. So if a competitor staffs his research boutique with ChatGPT while I staff mine with humans, the competitor will be able to charge a tiny fraction of what I do for research reports — and produce them 100X faster.

Again, I assume ChatGPT’s report contains errors. But a report from a non-subject-expert human researcher would also likely contain some errors. Because, well, we’re all human.

Here’s the report, so you can see for yourself. The “researcher” headshot, by the way, is of Casey Alvarez, the “economist” AI colleague that ChatGPT and I created. You can read about Casey and my other AI colleagues here. (The headshot makes her seem more… well… real… doesn’t it?)

The Current and Future Imact of Generative AI on Jobs – Especially Legal, Finance, and Creative

174KB ∙ PDF file

Download

A 23-page overview of where we are and where we're headed — and what today's knowledge workers can do to take advantage of the amazing power of AI and "AI proof" their skills and employability.

Download

And, in case you don’t want to read 23 pages, I’ve also included a shorter report, written by Sierra Quinn (actually, Perplexity), another of my AI colleagues. Sierra also researched and produced her report in minutes.

The Transformative Impact Of Generative AI on the Job Market — Especially The Legal, Finance, and Creative Sectors

354KB ∙ PDF file

Download

An overview of AI's current and (likely) future impact on the job market, with a focus on some specific "knowledge-worker" segments. Which jobs will change. Which will be eliminated. Which will be created. What knowledge workers can learn to make themselves even more valuable in the AI era.

Download

The TLDR conclusion, btw, is this — my words/view, based partly on Casey and Sierra’s work and partly on my own research and thinking:

AI will change jobs, eliminate jobs, and create jobs — the same way technology always has. Job-doers who learn how to use AI intelligently will thrive. Job-doers who don’t will fall behind. AI Luddites will become like human phone operators after telcos moved to electronic switching: Good at something the world no longer needs. Those who learn to use AI intelligently, meanwhile, will thrive.

So, is that it? Am I obsolete?

I hope not.

I hope — or at least want to think — there are still things I can do as a journalist, analyst, communicator, and human that ChatGPT can’t. I hope that these things will enable me to create enough value for my human customers that they keep reading and listening to me.

What are these things?

People say they they value my work because I see the big picture, focus on what really matters, apply common sense, and explain concepts in simple and engaging ways.  In other words, people say I help them learn and think about things they care about in ways they find helpful and worthwhile.

The product/service I (and other analysts) provide, in other words, is not really “research reports.” Most busy people don’t actually have time to read research reports, no matter how good the reports are — and no matter who (or what) wrote them. And, research reports can’t make people’s decisions for them. What helps us make decisions is often hearing a range of views from people whose perspectives and personalities we trust, value, and/or enjoy.

I hope, for now, I can still help people make decisions — or, at least, inform, entertain, or otherwise use their time in ways they find worthwhile.

If so, this advantage doesn’t feel particularly durable.

Now that I know how competent and fast AI is, I will use it to make my own research work better and faster — the same way I’ve used many other key tech innovations to improve and communicate over the past 30 years.

(Typewriters, for example. Then PCs. Then the Internet. Then email. Then texting. Then search engines. Then vast repositories of research, data, history, and expertise. Then the iPhone. Then Slack. Then Zoom. Then remote work. Etc.)

But that advantage — Henry + AI! — doesn’t feel all that durable either.

Rather, it feels like an in-between stage — like when IBM’s Deep Blue could beat the world-champion-human-chess-player Garry Kasparov, but Kasparov believed that, if they gave him a supercomputer, too, he and the supercomputer could beat Deep Blue.

That may have been true for a while.

But not now.

Now, we’re in an era when Google’s Alpha AI can learn the game Go by playing itself and, in 3 days, get better than the best human Go player.

So I don’t expect “Henry + AI!” will be better at writing research reports than “just AI” for long.

But, thankfully for me and other humans, most “knowledge-work” jobs involve actual decision-making and working with other humans — not just producing research reports. Any decision or project that involves people working together and communicating with each other will benefit from human input for a while.

Yes, our jobs will change as we learn to use AI. But as long as we are working with other humans, we will want colleagues who treat us like people and inspire us to work together (and with AI!) to create a better future.

Fortunately for me (and you), that’s what most human workers actually do.

Subscribe now

Thank you for subscribing to Regenerator! We’re the publication for people who want to build a better future. We believe that the best way to solve our problems is to innovate our way out of them. We analyze the most pressing questions in the innovation economy — tech, business, markets, policy, culture, and ideas.