Which Crash Are You Worried About?
Every AI argument is about one of three crashes: carjackers, toll roads, or the car taking the wheel and locking the doors. Most of the shouting is people worried about different ones.
Ask which crash someone is worried about, then check how they get paid.
Which of these headlines scares you more?
- "AI-written malware takes down a hospital network."
- "Five companies now control 90% of the world's intelligence."
- "AI is now smarter than the people who built it. Nobody can turn it off."
Your answer tells you which crash you are worried about. It is probably not the same one as the person you argued with last week.
Last month I watched a Diary of a CEO debate about AI nearly come to blows, and halfway through I realized the panel was not arguing about the same thing. Neither are Bill Gates and Jensen Huang. Neither are you and your most cynical coworker. Every AI argument you will hear this year is about one of three crashes, and most of the shouting is people worried about different ones.
The three crashes
CRASH 1 CarjackersBad people, better tools
Nothing is wrong with the car. Someone else is driving it.
Hackers, scammers, and the occasional bioterrorist get a 10x upgrade from models that exist today. Bill Gates says we already crossed this line. The fix is locks and a dashcam, not a better engine: logging, liability, know-your-customer rules. This is the crash that is happening right now.
CRASH 2 Toll roadsGood people, bad system
Nobody crashes. Everybody slowly goes broke.
Everyone is driving fine, but 4 or 5 companies own every highway, the tolls go up each year, and the exits keep closing. The gains flow to whoever owns the road. This is Acemoglu's warning and Emad Mostaque's digital feudalism. Slow, diffuse, and nobody gets booked on a podcast for it.
CRASH 3 The car takes the wheel and locks the doorsNo people, smarter car
It is smarter than us, it is driving, and we cannot get out.
AI improves AI, and we are arguing about whether it is headed somewhere good. Nobody has a fix because nobody has seen it happen. Serious people range from "never" to "coin flip." This is the only crash where "but China" is a real argument, because it is the only one with a finish line.
The cheat code
Ask which crash they are worried about.
Half the panels you will watch are someone using a Crash 3 fear (China, superintelligence) to block a Crash 1 fix (log what the molecule-design models are being asked to do).
That is refusing to wear a seatbelt because you are worried about the bridge collapsing.
The other half runs in reverse. A real Crash 1 harm (AI already helps hackers) gets offered as proof that Crash 3 is coming. Different question. Different evidence.
Jensen Huang is the clean example. He calls recursive self-improvement "a fabulous thing" and answers "no" on existential risk. He can be right about Crash 3 and still have nothing to say about Crash 1. Quality control does not stop a carjacker.
Jensen says the car cannot crash. The doomers say the car has no brakes. Everyone else is arguing about whether to check the mirrors.
Follow the money
Almost everyone in this debate is doing what they are paid to do. Including the people I agree with.
| Who | Crash they talk about | What they sell |
|---|---|---|
| Jensen Huang | NONE The car cannot crash | GPUs. To everyone. |
| The doomers | CRASH 3 No brakes | Views, clicks, likes |
| Ed Zitron | NONE The car is a cardboard cutout | Newsletters, by being bearish |
| Bill Gates | CRASH 1 Carjackers, today | Nothing, which is why he is worth hearing out |
| Anthropic | ALL 3 And still building | The careful version of the car |
| Me | ALL 3 On different days | Consulting on how to drive it |
Nvidia sells the shovels, so Jensen sees no cliff. The bubble critics sell newsletters by being bearish. The doomers get booked because:
"A billion deaths" is a better headline than "we should add logging."
And the labs, Anthropic included, get to say "the risk is real and we are the careful ones," which happens to be a great reason to keep building.
Here is the thing. I think Anthropic means it. Dario Amodei proposed a global speed limit on AI and admitted in the same breath, "I don't know if it's possible, but we should try." That is not what a cynic says. It is what you say when you believe the risk and also believe that if you stop, someone less careful takes the wheel. Self-serving and correct can both be true. People who call them crazy lose me. "The people closest to the technology are the most worried" is not a crazy signal.
If someone's answer about the crash matches how they get paid, be suspicious. That includes me.
The tell is the chip policy
If Washington believed this was a race, it would not sell the other driver spare tires. Instead the US approved the H200 (a generation old) with a 25% tariff, a volume cap, and inspections. Then Beijing blocked the shipments at customs to force its own chipmakers to grow up.
That is not a race. That is two countries fighting over who owns the road and calling it a race because the word is useful to everyone selling something.
Where I land (for now)
I do not buy the bubble. I see a massive shift of spending into data centers over the next decade, and a transformation big enough to need an army of people who can bring it into companies. New jobs will exist. But we are also racing toward brutal optimization of white-collar work, and the safest seat is helping run the optimization: consulting, adoption, evaluation.
The timeline is what everyone gets wrong, in both directions. The AI-pilled crowd, me included, feels like it already happened 4 months ago. Meanwhile most people have not used any of this, are cynical about all of it, and are years away from mass adoption. Both are true. Call it 6 to 10 years. It will also happen.
On Crash 3, I do not know, and neither does anyone who sounds sure. The real question is whether we would see the cliff before going over it. Priors arguing with priors.
My working rule: tap every brake that costs nothing in the race. Stop letting Crash 3 arguments block Crash 1 fixes. And watch the paychecks.
We are not on a bus heading for a cliff. We are on a bus hitting potholes while everyone argues about whether there is a cliff. My position on this changes. That is the point.
So, which crash are you worried about?
Watch these
The interviews behind this post. Pick one per crash if you are short on time.
- Dario Amodei on Dwarkesh: "We are near the end of the exponential"The most confident near-term timeline from someone building the thing.
- Jensen Huang on The Ezra Klein Show: "AI alarmism has gone too far"The "it is just software" case, from the man selling the shovels.
- Bill Gates on The Ezra Klein Show: "Blunt warning on AI"Crash 1 in its purest form, with the "monitor the molecule models" ask.
- The Diary of a CEO: The Great AI DebateYampolskiy and Soares vs Zitron and McAfee. Crash 3 vs "the hype is fake," live.
- The Diary of a CEO: Emergency Debate, Kevin O'Leary vs Cenk UygurThe China race argument in the wild.
- Simon Willison on Lenny's Podcast: An AI state of the unionThe ground-level view. Dark factories, and why November 2025 was the inflection.
- Emad Mostaque on Know Thyself: The Last EconomyThe three futures: feudalism, fragmentation, or symbiosis.
- Mo Gawdat on why AI exposes capitalism's flawCrash 2's loudest voice.
- Zvi on the Jensen interviewThe sharpest critique of why Jensen's position does not hold together.