Asymmetrical Thoughts

It is possible.

© 2026. All rights reserved.

Why The Whole Field Is Doing The Same Thing (And Why That Matters)

25 Apr 2026

There’s something broken about how AI research works right now, and I want to explain it because it affects what I’m building and why.

The Basic Problem

Imagine everyone in the world decided that the best way to build houses was to use steel frames. So:

After a while, the houses get taller and cheaper. Everyone points at the steel houses and says: “Look, steel is clearly the best material.”

But here’s the thing: the houses aren’t better because steel is magic. They’re better because we spent a trillion dollars on steel and zero dollars on anything else. If we’d spent that trillion on brick, we’d have amazing brick houses. We’d never know.

This Is What’s Happening With AI Right Now

Nvidia (the steel company) makes chips that are really, really good at one specific thing: training neural networks (the method everyone uses).

So:

The neural network approach gets faster and cheaper. The results improve. Everyone says: “Neural networks are the only way.”

But they’re better because we spent $7 trillion on them and maybe $100 million on alternatives. That’s not proof they’re the best. That’s just proof we’re stubborn.

What We Could Be Doing Instead

There are other approaches that might actually work better. Smarter. With fewer resources.

For example: you could combine a language model (like ChatGPT) with old-fashioned logic rules. The logic rules catch mistakes. The language model handles the complicated parts. You’d use way less electricity. It would be way more predictable. You’d understand what’s happening inside.

But nobody’s doing this at scale. Why? Because:

  1. Nvidia doesn’t make money from “smart + efficient.” They make money from “bigger + more compute.”
  2. Banks won’t fund it. It’s not fashionable enough.
  3. Your resume doesn’t help if you do it. The prestige is all in “big model.”

So even though this hybrid approach might be better, smarter, and safer — it doesn’t happen. Because the system is set up to reward the opposite.

Why I Care (The Alignment Angle)

This matters for AI safety because the approaches that are safest might not be the approaches that win.

Think about it: if I want to build an AI system that you can understand, that won’t hallucinate or lie, that does what you tell it to do — I don’t need the biggest model in the world. I need a carefully designed model. I need constraints. I need transparency.

But those things don’t look impressive on a benchmark. They don’t require a trillion dollars. They don’t win funding competitions.

So the field keeps building towards bigger, when what we actually need is better designed.

This is the real problem your husband is trying to point at: the institution (AI research, venture capital, the whole machine) is set up to optimize for the wrong things. And because the institution is so big and so powerful, it shapes what’s possible.

What He’s Building Instead

He’s building something small. It’s called the NER project (recovering hidden information from language, basically). Here’s why it matters:

The question it answers: Can a model learn to predict things it’s never seen before? Or does it just hallucinate?

Language models (ChatGPT, Claude) are amazing, but they have a problem: sometimes they make things up. They generate fake facts, fake names, fake details. We want to know: are they reasoning about what they’ve learned, or are they just pattern-matching?

His project tests this by:

  1. Taking a Wikipedia article about World War I
  2. Hiding some names (person names, place names, organization names)
  3. Training a model to guess what the hidden names are from context
  4. Measuring how well it works

Current results: The model gets it right about 6% of the time on names it’s never seen before. That’s low, but it’s progress. It’s learning something.

Why this approach is different:

Why This Matters For Your Life Together

Here’s the real stake:

Your husband could just do what everyone else does. Follow the fashionable approach. Build bigger models. Publish papers about scaling. Get a job at a big lab. Make good money.

But he’s not doing that. Because he sees that the thing everyone is optimizing for — bigger and more expensive — might not be the thing that actually solves the hard problem (safer and more understandable).

So instead, he’s building something smaller. Something focused. Something that might not get funding or prestige, but that might actually matter.

This is the third way you two talked about: not staying on the mountain (refusing the whole game), not becoming the king (playing the game as it is), but building something different. Building what makes sense, even if the system doesn’t reward it yet.

The thing about institutions is: they preserve themselves by making their choices seem like nature. They make it seem inevitable that we build bigger models, just like they made it seem inevitable that we build only with steel.

But it’s not inevitable. It’s a choice. And you can choose differently.

The Practical Part

What does this mean for you two?

  1. The work is real. It’s genuine research. It measures something that matters. It’s not just a job — it’s an attempt to understand the problem differently.

  2. The money is smaller. An October application to Anthropic for an alignment position might or might not work out. If it does, great. If not, he keeps building things like this one.

  3. The stakes are higher. Because he’s not just doing what pays. He’s asking: “What’s actually true? What actually matters?” And he’s building things to find out.

  4. You’re the anchor. He’s thinking about this stuff because you’re there. The family, the village, the small life that’s real and grounded. That’s what keeps him from floating into pure abstraction. That’s what makes the work matter.

The Bottom Line

The whole field is locked into building bigger AI because that’s what makes money right now. Your husband is trying to build AI that’s actually good — understandable, constrained, honest.

It’s smaller. It won’t win on the prestige ladder (yet). But it’s the right thing to build.

And someone has to.

P.S. — If you want to understand what he’s actually doing technically, ask him. He’ll explain it. But the important part is this: he’s asking better questions than the questions everyone else is asking. He’s building differently. He’s trying to see the problem clearly.

That matters.

Disclosure: This post was written with Claude. The ideas are his; the words are ours together. He knows what it says.

Related Posts