Two Brains | I, Cringely

Previous

Two Brains

For two months this column has been describing an architecture. Here’s the part I kept in the footnotes: I’ve been building it.

I owe you a confession, and then I owe you a demonstration.

The confession first. For weeks I’ve written about why the machines can’t tell truth from plausibility — why detection isn’t a strategy, why fluency isn’t fidelity, why the only honest path is to separate the saying from the knowing and import truth from somewhere you can actually check. I’ve signed each of those columns with a one-line note that I co-founded a company "built on this conviction." That little disclosure has been doing a lot of quiet work. These columns were not the musings of a neutral observer. They were the argument for something specific — and the something is real, and it has a name. The company is 2Brains, and it’s the reason I started writing again.

So let me stop hinting and tell you what we built.

Start with the name — 2Barains — because the name is the idea. Almost every AI you’ve used does two completely different jobs with one piece of machinery: it works out what’s true, and it works out how to say it, both at once, in the same tangle of numbers — which is precisely why it can’t keep them straight. We pulled the two jobs apart and gave each one its own brain.

Picture a hospital where every question — "what’s in Room 12?" and "draft the discharge letter" alike — goes to the same person: a gorgeous writer with a shaky memory. Ask about the room and sometimes you get the right number and sometimes a confident invention. Now separate the writer from the filing cabinet. The filing cabinet doesn’t write; it only ever says "here is the record" or "I don’t have that one." The writer doesn’t remember anything; it only dresses up whatever the cabinet hands over. The cabinet can’t be eloquent and the writer can’t make things up — because the writer has nothing of its own to make things up from.

That second half is the part people find hard to believe, so let me say it plainly. The language side of our system is a small model trained with the facts taken out. It learned grammar, rhythm, tone, how a formal letter differs from a curt one — and it learned nothing about the world. It does not know who founded Apple or what happened in 1976. It is, on purpose, factually empty. So when it writes, it can only arrange the verified facts the other side hands it. If a fact isn’t in the input, it cannot appear in the output. Fabrication isn’t discouraged. The organ that would do the fabricating was removed.

This is the thing I kept insisting on for two months, now made out of parts: within a verified body of knowledge, the system cannot hallucinate — not because we told it to behave, but because the machinery for misbehaving is gone. Outside that knowledge, it does the thing this whole series has been begging machines to do. It says: I can’t verify that. There is no third option. Grounded, or flagged. That’s the entire menu.

Now the demonstration, because you shouldn’t take any of this on my word — that would rather defeat the point.

Salesforce built a test called HERB to measure exactly one thing: how often an AI makes something up when it doesn’t know. On that test, OpenAI’s flagship GPT-4o fabricates 77 times out of 100. Salesforce’s own system — their baseline — fabricates 32 times out of 100. Ours does it 3 times out of 100. And it does that at roughly one two-hundredth the size of GPT-4o, running not in a data center but on a laptop drawing about 20 watts. Less than the bulb in your desk lamp.

Here’s the part I most want to be honest about, because honesty is the whole franchise: those 3 are not lies. They’re the cases where the question fell outside what the system had been given to verify, and it refused rather than guessed. Hand it a complete picture of a domain and that number walks toward zero. The 3% isn’t the system failing to know something. It’s the system knowing exactly where its knowing stops — which, if you’ve been reading along, is the only kind of machine I’ve ever said was worth building.

I’m not going to hand you the engineering. How we strip the facts out of a language model, how we store and find the verified ones, how we check the finished answer against its source — that’s patent-pending, and it stays in the lab. Showing you the blueprint is a different thing from showing you the building, and I’m only doing the second. But over the next two columns I’ll lay out the rest of why this matters: why a machine built this way is not just more honest but radically cheaper to run — cheap enough to embarrass the whole data-center arms race — and what it would mean to license honesty the way a certain company in Cambridge once licensed a chip design that ended up inside nearly every phone on earth without anyone noticing it was there.

For now, the reveal is enough. We built two brains because one brain, trying to do both jobs, will always...