You’re sitting in a standup meeting when someone on the ML team says the word “architecture.” They mean transformer blocks. They mean attention heads and residual streams. They mean a thing that was designed, once, by people, on purpose, with a diagram.
Meanwhile, a paper landed in Nature Neuroscience in September 2026 that says the actual brain — the wet one, the one you’re using to read this — didn’t get built that way at all. Jokhai, Dundes, Ahsan and colleagues report that two parallel neural ectoderm progenitors contribute to the developing brain. Not one population that fans out into everything. Two. One set heads toward forebrain and midbrain. The other set heads toward hindbrain. Separate lineages, running at the same time, committed early.
I review AI tools for a living. I read a lot of copy about systems “modeled on the brain.” So let me be blunt about why this particular finding stuck with me.
Parallel, not hierarchical
The intuitive story about brain development — the one most of us absorbed somewhere and never questioned — is a tree. Start with a trunk of undifferentiated cells, branch, branch again, and eventually you get cortex over here and brainstem over there. Clean. Hierarchical. The kind of diagram that fits on a slide.
What this research describes is closer to two pipelines running concurrently, each already pointed at its destination before the interesting part begins. The companion work by Dundes and colleagues in 2025 framed the question directly: are early neural ectoderm cells already fated to produce either forebrain/midbrain or hindbrain while still in the gastrulating embryo? They tested it with two complementary approaches. The answer, per the 2026 paper, is yes — parallel and lineage-committed.
That’s a different kind of system than the tree. It’s less “one design that specializes” and more “two designs that ended up in the same skull.”
Why an AI reviewer cares
Because the neuroscience metaphor in this industry is doing an enormous amount of unearned work.
Every other product page I read tells me something is “brain-inspired” or “neurally architected” or built on principles borrowed from human cognition. Almost none of them can tell you which principles. The metaphor functions as a vibe, not a claim. And the vibe has always assumed the brain is one unified thing that emerged from one unified process, which is convenient if you’re selling one unified model.
The findings here cut against that in a way I find genuinely useful for calibrating claims:
- Biological intelligence didn’t come from a single origin story. At least two parallel progenitor populations built the thing.
- The regional split isn’t a late-stage specialization. It’s baked in early, at the lineage level.
- “The brain” as a single design object may be a category we invented for convenience rather than something evolution actually produced.
If you’ve ever wondered why multi-model systems, routers, and mixture-of-experts setups keep outperforming the one-big-model dream in practice, this is at least an interesting coincidence. Biology apparently didn’t bet on monolithic either.
What I’m not going to claim
I’m not going to tell you this validates any particular AI architecture. It doesn’t. Developmental biology is not an engineering blueprint, and people who treat it as one produce bad products and worse blog posts. Two progenitor populations in an embryo tell you nothing about optimal routing in a language model. The honest version of the takeaway is narrower and more about epistemics than engineering.
I’ll also flag something that annoys me: the original research is closed access. A finding this foundational about how brains get made, and the paper sits behind a paywall. The abstract circulates, the headlines circulate, Hacker News argues about it, and the actual methods and figures stay locked up for anyone without institutional access. If you want the public to reason carefully about intelligence — biological or artificial — locking the primary source is a strange way to go about it.
The recalibration
The paper was received on 7 November 2025 and published in 2026. Fast turnaround for something that reframes a default assumption.
Here’s what I’m taking into the next product demo. When a founder tells me their agent framework mirrors human cognition, I’m going to ask which part. Forebrain? Hindbrain? Because as of this year, those aren’t two neighborhoods in the same city. They’re two construction projects that happened to finish adjacent to each other, run by crews working in parallel from the very start.
Most people making the brain comparison won’t have an answer. That’s fine. But I’d rather they say “we built a good system” than borrow authority from a biology they haven’t read. The actual story of how brains assemble themselves is stranger and more interesting than the metaphor, and it’s not sitting still. Neither should the claims we make about it.
🕒 Published: