Why I Train AI for People and Companies That Do Good

The 20%

Yesterday I read Scott Alexander's collected opinions on AI. Two numbers stuck with me. He puts the odds that the first truly powerful AIs will want to wipe us out at about 20%. And he gives roughly a 20% chance that the year 2100 is a future we would call good — not good to whoever's living in it, good to us, by the values we hold right now.

I read that and decided to spend my working life on the second 20%. To explain why that isn't naive, I need a cat.

The cat in the trunk

You know Schrödinger's setup: a cat sealed in a box with a radioactive atom and a vial of poison. Until you open the box, quantum physics says the cat is both alive and dead at the same time. Schrödinger meant it as a joke at physics' expense — a way of showing how silly it is to stretch quantum weirdness up to the size of a living animal. But notice what the setup quietly does to you: it writes you out of the story. You get two moves — build the trap, then wait. The universe rolls the dice; your only job is to lift the lid and find out what happened. It's a tidy little tale of fate. The future is sealed in a box, and we're all standing around it, waiting for 2100, when someone finally opens it and we learn which world we got.

Now drill an air hole in the box.

The hole changes everything — because the box's real job was never to keep the cat in. It was to keep information from getting out. The alive-and-dead trick only holds as long as nothing outside the box can tell which one the cat is. But the instant a single air molecule bounces off a warm, breathing cat instead of a cold, still one and slips out through the hole, the outside world has "felt" the cat. Physicists call this decoherence: the both-at-once weirdness snaps into a plain fact — not when a person looks, but the moment the inside touches the outside. With a hole in the box, the cat's fate isn't on hold until you open the lid. It's being decided the whole time, molecule by molecule, by contact with the real world.

The hole is a kind of free will

And the hole isn't only where information leaks out — it's where attention goes in. There's even a real result in quantum mechanics, the quantum Zeno effect, where watching an unstable system closely enough can hold it in place — and a whole field of engineering built on top of it, where you measure a system constantly and nudge it back on course in real time. Watching stopped being passive the moment we got fast enough to act on what we saw. Drill the hole, watch closely, act on what comes out, and you're no longer the bystander Schrödinger made you. The free choice in the story was never inside the box. The atom doesn't choose; the math doesn't choose. The free choice is drilling the hole — noticing that the setup casts you as a spectator, and refusing the part.

That's how I think about AI and the year 2100. The sealed-box version of the future — sit back, argue about the odds, open the box in 75 years — is a choice, and a bad one. The other option is to connect AI to the real world early and often, in small steps you can watch and undo, and act on what you see. Don't wait to find out which future we got. Help decide which future we get.

A 2,500-year-old wager

But betting your life on a 20% needs more than physics — it needs an answer to "what if you're wrong?" The best one I know is 2,500 years old. In the Kalama Sutra, a town full of people tell the Buddha they've been visited by too many fake gurus, each glorifying his own ideas and tearing down the rest, and they no longer know whom to believe. His answer is the most modern thing in ancient literature: don't believe something because it's repeated, because it's in a respected book, or because the person saying it seems impressive — even if that teacher is me. Test it against what you can see for yourself: does acting on it lead toward harm, or away from it? (I published a plain-English translation at plaindharma.com.)

Then he closes with a wager. A person who lives well holds a winning ticket in every branch: if there's a next life, they're in good standing for it; if there isn't, they still lived in peace here. Either way, they come out fine.

If that sounds familiar, it should — it's Pascal's wager, about 2,100 years early. But notice what the Buddha's version fixes. Pascal asks you to bet on belief — believe in God, because if you're wrong the cost is infinite — and people have poked holes in it ever since: which God? can you even make yourself believe on command? is a belief you only hold as a hedge worth anything? The Kalama wager skips all of that, because you bet your actions, not your beliefs. You don't have to talk yourself into anything about the afterlife, or about 2100. You just act — and no matter how it turns out, you've already collected the one prize you keep for sure: a life well lived, right now. Pascal's bet only pays off if you guessed right about God. The Buddha's pays off either way.

To be fair to Pascal, he half-knew this. His advice to someone who couldn't believe wasn't "force yourself" — it was act as if: go to church, go through the motions, and belief will follow. Read kindly, he was betting on actions too. And faith in this sense — not being sure you'll win, but staying loyal to a path even when you can't know how it ends — isn't a flaw in the bet. It's the key to it, for a reason I'll get to.

Run my bet through the same machine. Three branches:

• The good future happens — I did right, and I helped cause it.
• The world muddles through — neither paradise nor disaster — and I did right, and I spent those decades building clean energy and useful tools the muddling world badly needs. (Honestly, this is the one where the business pays off best.)
• The worst happens anyway — I did right, and lost nothing extra by it.

There's no version of the future where I regret the work — a bet you can't lose. And it doesn't depend on Scott's 20% being right. His number could be way off in either direction and my choice wouldn't change, which is exactly the kind of belief the Kalama test says you're allowed to act on: the part that matters is something you can check for yourself. Does the work lead toward harm, or away from it?

And don't take "the middle branch pays best" on faith either — run the Kalama test on the economics. Pick any market where clean and dirty compete head to head, and look at what the numbers already say:

The odds aren't fixed

There's one place the physics comparison breaks down — and that break is the whole point. A radioactive atom doesn't care what anyone believes; it decays at its own fixed rate no matter what. But the odds of a good future by 2100 aren't fixed like that. They're a guess about what billions of people will do — and a guess like that can be self-fulfilling. The odds rise or fall depending on how many people treat the good future as reachable and act like it. The bad future needs no help from anyone. The good one is built entirely by volunteers.

So believing the good future is possible is itself one of the things that makes it possible, in a plain, slow-adding-up way: a person with a clear aim makes thousands of small choices differently, for years, and all those choices add up. The watched future gets built.

So that's why

This is why AimHuge trains the people it trains. When I teach a team to use AI, I'm handing them a power tool, and I'd rather hand it to people building things I want more of in the world. It's why Zeph — clean energy in the markets that need it most — gets the most ambitious AI setup I can build: AI that reads, drafts, and acts while a human keeps watch, with a checkpoint before anything that can't be undone. Watch fast, act early, never let trouble build up while no one's looking. That's not just a way to build software. It's the air hole, drilled on purpose.

Every team that learns to use AI well — and learns to keep a human hand on the approval gate — is another hole in the box. Another channel coupling this technology to real work, real consequences, real judgment, instead of leaving it by default to the incumbents with the deepest pockets and the least reason to change.

Scott Alexander says you can't know which future we'll get until we get there. He's right. My answer is the cat's answer: then never stop opening the box.

If you're building something good and want your team to be dangerous with AI, that's what the workshops and advisory are for.

If you're already AI-powered and doing good, apply for funding at zeph.vc.

And if you're an entrepreneur still in the trunk — you know you want out, you just don't know which way to leap — browse the Zeph Idea Bank, pick one, and start a business that makes money and does good. The good branch is made of volunteers. Consider this the recruiting pitch.