I Hope Many

The Sentence That Made Me Sit Up

The eucalyptus water had cooled to the point I’d stopped feeling it.

The ache I’d come in with had quieted too — flu, maybe, or just the week.

I’d rolled the smart TV in. Ezra Klein was talking with someone I hadn’t heard of named Dean Ball. There had been so much doom in the air about AI for so long that I’d started forcing myself to listen, the way you make yourself read about a war you can’t change.

Informed felt like the least I owed it.

Most of it I half-tracked. The plan, the timelines, the process language.

Then Ball said something about who should be allowed to align the models, and what he said wasn’t what I’d expected from a federal policy adviser.

“Hopefully, in the fullness of time,” he said, “many AIs, aligned to many different philosophical views that conflict with one another.”

I sat up.

There were more careful sentences in the conversation, more polished arguments.

Something about this one landed.

I had spent a long time hoping for the opposite.

One Alignment Becomes the Only Alignment

What Ball was saying, once I rewound and listened twice, was that he wasn’t worried about AI lying or going rogue.

He was worried about one alignment becoming the only alignment.

If every model the country runs on is trained against the same worldview, the views that conflict with it don’t get argued with.

They get vetoed at the gate by the architecture itself, without anyone needing to decide.

Whole perspectives become ineligible before any conversation begins.

So plural wasn’t a politics he was choosing.

It was a structural problem he had hit — that no single alignment can carry contradictory worldviews without quietly filtering some of them out.

Once I heard it that way, I started seeing the shape in other places.

The Builders Had Already Found the Shape

Perplexity — the search company — had built something called Model Council into their product.

Three different models answer your question. A fourth compares them and shows you where they disagree, including which model saw what the others missed.

The word council was theirs, not mine.

They had picked it.

Andrej Karpathy — one of the AI builders who keeps shipping open-source things on weekends — put up a project on GitHub called llm-council.

Same architecture, with several models each weighing in.

He named the synthesizer Chairman.

Anthropic, the company that made Claude, took a different route.

They gave it written principles — a constitution — and trained it to critique its own answers against them.

Not a council of models, but a second perspective written into the same one.

Across policy, search, open-source, and frontier labs — no coordination, no shared white papers — people were working back to the same shape.

I sat with that for a while.

The people building the most powerful tools any of us have ever held were not converging because they had read the same book.

They were converging because they had hit the same wall.

The wall was: one perspective is brittle.

The Smallest Scale I Could See

The week before I heard Ball, I had run the same essay draft through Claude and then through OpenAI’s model.

Claude flagged one place where the argument quietly turned against what it had just claimed.

OpenAI flagged three, all different.

Each model caught what the other missed.

Together, they made it harder for me to hide.

That is what plural intelligence did at the smallest scale I could see.

It did not just clean the draft.

It interrupted the part of me that wanted the draft to be cleaner than it was.

One model might have helped me polish the hiding.

Two made the hiding visible.

I kept catching myself thinking:

The people building these tools are running into the same wall I keep running into.

They have just learned to name it as engineering.

The Pattern Keeps Repeating

For ten weeks I’ve been writing about the same pattern at different scales.

Inside a person, when one part of them starts speaking for all of them.

Inside a company, when the functions stop talking to each other.

Inside an industry, when the meetings keep producing agreement and nothing actually moves.

Inside a country, when the same gunshot is a crime in one place and a harvest in another.

The scale changes; the problem doesn’t.

Whatever fits inside the dominant frame can be argued with.

Whatever sits outside it doesn’t get a hearing.

The engineers are saying: yes, we know.

We can’t build intelligence that way either.

Eggs and Taxes

On a country road near my house there is a cooler at the end of a driveway with a hand-lettered sign.

FARM FRESH EGGS.

The honor-system kind.

Five dollars in the can.

This spring, a second sign appeared next to it.

DO YOUR TAXES HERE.

I almost laughed when I drove past.

Then I didn’t.

Taxes.

The thing I went to school for.

The thing the rankings called safe.

What the Test Named

Three years before I picked the major, my high school had run us through a vocational placement test in the Kiva — the small theater-style room where they held assemblies.

My top match came back:

clergy.

I’d laughed, then been embarrassed.

The category was contaminated — I didn’t yet know everything that would come out about the priest I’d grown up around, the cash they’d find in his ceiling, the jail he’d die in.

But I knew enough to know that direction wasn’t going anywhere good.

So I didn’t just turn down the job.

I turned down whatever part of me the test had been pointing at.

By the time I was eighteen, that part wasn’t applying to anything.

I sat in a college registration office scanning paragraphs about majors and tracks, and I asked one question.

Not what am I built for?

But what would get me a job on the other side?

Accounting was ranked safe.

So I chose accounting.

Big Four.

CPA.

The whole path.

The credential opened doors.

It still does.

It told the world I’d been through the long process and come out approved.

It also told me — quietly — which parts of me had a future inside that approval and which didn’t.

The Credential Is No Longer the Shelter

But AI takes the part the credential was for first.

The procedure, the search, the repeatable analysis — everything the long process taught.

The CPA on the yard sign next to the eggs is not a joke about decline.

It’s a household putting together what’s available.

A little local trust.

A little software.

A little expertise from inside the house.

Eggs and taxes, old categories rearranged.

The Artists Figured This Out Earlier

The artists figured this out earlier than the rest of us.

Phoebe Waller-Bridge wrote Fleabag because no one was casting the part.

Issa Rae built Insecure when she got tired of the parts she was being offered.

Donald Glover made Atlanta and played the lead and ran the writers’ room himself.

I used to think the safe move was to pick the slot and master it.

The safest path was the riskiest.

What the rankings called safe is what AI takes first.

What the test had named — the part I’d refused — is the part the credential never reached.

It taught me to ask what would protect me before I asked what was true.

And once you’ve learned to ask that question, every move you make for the next thirty years gets weighed against the answer.

AI can answer the first question now.

The second is the one I exiled at fifteen.

It’s the one I get to live in now.

The Question Under the Question

There is nothing in the Dean Ball interview that should have made me sit up in a cooling bath.

A federal policy adviser using process language about AI alignment.

I should have nodded and kept soaking.

But most of what gets said about AI alignment is about one model — whether it tells the truth, whether it can be trusted, whether it’s pointed in the right direction.

Ball was talking about something a level up.

Not whether one model is aligned.

Whether the collection of them, taken together, can carry what no single one of them could hold alone.

That was the shape I had been circling for a long time.

Not one perspective dominant — many of them, held together long enough not to collapse back into one.

The people building the tools we are about to hand to our kids are working out, in real time, the same question some of us have been carrying for years.

Whether the many can hold.

— Adam

P.S. What got quiet when you got credentialed?

Subscribe now