A Dish Human Brain Cells Just Outplayed My GPU

Ninety percent. A dish of living human brain cells — grown in a lab, fed a sequence of electrical pulses — got ninety percent of the next tokens right.

That is not a typo. The tissue that folds your laundry and forgets your anniversary, plugged into a task the silicon world burns megawatts on, guessed the next word nine times out of ten with no GPU in sight.

Meet Sean Cole. He told the GPU industry to go to hell with a napkin sketch and a box of neurons. He is the CEO of Parasma, a two-person YC Summer 2026 startup parked in San Francisco, and his entire thesis is four words that should make every chip company on earth nervous:

Brain cells are compute.

I studied this from a hundred miles away and I still felt my teeth rattle. Because here is the number that matters: the human brain runs on the power of a light bulb. And it learns forever — no re-training, no catastrophic forgetting, no $40 million fine-tune because a customer changed a comma. It adapts. It lives.

The whole AI world is a junkie shooting GPUs into its arm to fake what a slab of meat between your ears does by default.

The kid who made neurons play Doom

Cole didn't start with a PowerPoint. He skipped the bachelor's, took an MSc in Artificial Intelligence with Distinction at Sussex, then did something every AI lab on the planet said was a parlor trick: he trained a culture of human neurons to play Doom.

Not a simulation. Real neurons, grown in a dish, wired to a screen. They moved. They hid. They shredded demons with the reflexes of a caffeine addict. And hundreds of millions of people watched the clip go berzerk across social media.

That was the party trick. This is the business.

"In March we trained them to play Doom; now we got them to do next-token prediction," Cole wrote in Parasma's YC launch. The dice went from a party game to the actual currency of the AI economy.

The test that broke the linear silicon ceiling

Parasma ran the cells through token prediction with full sentence context: ninety percent accuracy, corrections fed back in as electrical stimuli. Randomized control runs scored at pure chance — so it's real learning, not a heater in a box.

Then the hard test. A non-linear task where no single token could give away the answer. An isolated run hit 78.1 percent — over the 75 percent ceiling a linear silicon decoder can ever reach. Translation, shoved through the Thompson meat grinder:

There are problems a GPU cannot solve that this soup of neurons just solved.

"Crossing this boundary that a linear silicon decoder cannot represent is exactly the kind of result we built Parasma to pursue," Cole said. That is the quiet sound of a semiconductor epoch beginning to itch.

"The engineers who hate vibe coding and AI the most are the people who would benefit the most from embracing it."

— Garry Tan, via TheFoundersPack

Same shape as Parasma. The people richest in the old paradigm hate the new one the hardest — and the guy who benefits most is the ex-pro gamer with a dish of neurons and a two-person company who said the silicon ceiling doesn't exist, there's a different substrate underneath it.

Two people. A dish of meat. A bet on biology.

The team size on Parasma's YC page says two. Two. A co-founder the chips giants ignored, an open job for a founding scientist, and a thesis that scales by learning, not by buying more H100s.

Cole's own path is the weirdest resume in the batch: ex-professional gamer. Built a digital-item marketplace that listed north of $1.7 million in virtual goods. Then biopunk. The arc reads like a Manic Panic career-change fever dream — and it works, because the connective thread is raw pattern-recognition horsepower applied with no respect for convention.

The pack that isn't playing the loop

Every other AI startup is moaning about GPU supply, GPU price, GPU power. Parasma looked at the problem sideways and asked the only question that matters: why is this the only substrate? That's the same shape as any pack a poor, weird, hungry group adopts — don't fight the giants on their field, build the field they can't see.

Read Form Your Own Pack and you'll meet the thesis in a petri dish. A small team, a shared bet on a vision nobody wants to underwrite, a channel that starts as a cult following and becomes a revolution. The pack doesn't out-resource the incumbent; it out-thinks the frame.

The same shape as what we're building at Genflick — a small crew and one wild bet on AI filmmaking, treating every creator as a member of the pack rather than a user to be harvested. If a two-person lab can make neurons predict tokens, a small team can make an entire film. Try it free.

The Founder

Sean Cole

Founder & CEO, Parasma (YC S26)

@vytalow  |  LinkedIn

Sean Cole is a self-described AI obsessive, ex-pro gamer, and the man who taught living neurons to play Doom. He skipped the bachelor's, took an MSc in Artificial Intelligence with Distinction at the University of Sussex, and now runs Parasma on the thesis that the human brain is the most energy-efficient learning machine ever built — and it's sitting right there.

Links: YC Profile | Website | @ParasmaAI | Partner: Tyler Bosmeny

Parasma is YC Summer 2026. Header art: AI-generated cyberpunk, magenta and cyan, CRT scan lines. Human brain cells were harmed? No — they were promoted.