DARPA's New Secret: Computers Made of Human Brains
TL;DR
Cortical Labs has 200,000 human neurons learning Doom on a chip — In Melbourne, the Australian startup’s CL1 system keeps living brain cells alive for 500+ days and translates game state into electrical signals the neurons can respond to in real time.
The real breakthrough isn’t Doom — it’s the interface layer — Dylan’s big takeaway is that once Cortical Labs exposed its “biological operating system” via a Python API, outsider Shawn Cole got a Doom demo working in about a week, versus 18 months of research to get neurons playing Pong in 2022.
These neurons are trained through predictability, not classic rewards — Using Karl Friston’s free energy principle, correct behavior gets stable electrical stimulation while mistakes trigger chaotic noise, pushing the network to reorganize itself to avoid surprise.
DARPA, In-Q-Tel, and data centers are already in the picture — Dylan says DARPA launched a 42-month biomputer program, the CIA’s venture arm is invested, and Cortical Labs opened a 120-unit biological data center in Melbourne with a larger Singapore facility reported by Bloomberg for 2026.
The energy argument is enormous: 20 watts versus supercomputer-scale power — He contrasts the human brain’s roughly 20-watt power draw with the 21 million watts used by Frontier in Kentucky to match brain-scale computation, framing biomputing as a possible way around AI’s energy crisis.
The science is real, but the claims are contested and the ethics are brutal — The 2022 Pong paper was peer-reviewed, but the Doom demo is mostly shown through company media, while critics from Harvard, Johns Hopkins, and the Allen Institute challenge the learning claims and philosophers like Thomas Metzinger warn the systems could be primitive sufferers.
The Breakdown
A petri dish playing Doom, and why that’s not even the wildest part
Dylan opens on the image that hooks the whole video: 200,000 living human brain cells on a silicon chip in Melbourne learning to play Doom. They’re bad at it — spinning, crashing into walls, dying constantly — but they’re measurably better than random, which is enough to make the whole thing feel less like a stunt and more like a new category of machine.
The Australian team behind “synthetic biological intelligence”
He grounds the story in Cortical Labs, founded in 2019 by Hong Weng Chong, alongside Andy Kitchen and neuroscientist Brett Kagan. Dylan lingers on how weirdly cross-disciplinary they are — doctor, software engineer, stem-cell scientist — because this only exists at the overlap of biology and computing.
How blood becomes neurons, and neurons become hardware
The setup is as unsettling as it is elegant: take human blood cells, rewind them into induced pluripotent stem cells, then coax them into cortical neurons and grow them on a 59-electrode chip. Dylan calls the CL1 a “tiny little hospital for brain cells,” with pumps, filtration, and temperature control keeping some cultures alive for more than 500 days.
Translating Doom into electricity the neurons can understand
The neurons never “see” a screen; a software layer converts things like enemy position, direction, and health into patterns of electrical stimulation. Dylan loves Alan Lafler’s framing here: the job is to translate “the digital world of Doom into the biological language of neurons, which is electricity,” then decode the neurons’ spikes back into actions like turning or shooting.
Why random noise teaches the cells to play
The most memorable part for Dylan is the learning mechanism: when “Stan,” his joking name for Doomguy, does something right, the neurons get stable, predictable signals; when he messes up, they get chaotic noise. He compares it to walking through a building where the lights flicker wildly every time you make the wrong move — you’d adapt fast, not because anyone explained the rules, but because your nervous system wants the chaos to stop.
From Pong in 18 months to Doom in a week
Dylan treats this as the true headline. Cortical Labs’ 2022 Neuron paper used about 800,000 neurons to learn Pong after 18 months of careful work, but in 2025 independent developer Shawn Cole used the new Python API at a Stanford hackathon to get Doom running in roughly a week, which suggests the bottleneck was the interface, not the biology.
Wetware-as-a-service, military interest, and the 20-watt pitch
Once the platform became programmable, the story stopped feeling academic and started feeling industrial: a CL1 costs around $35,000 locally or about $300 a week via the cloud, which Dylan notes they literally call “wetware as a service.” He connects that to Melbourne’s 120-unit biological data center, a planned Singapore expansion, and DARPA’s 42-month biological processing unit program — all driven by the promise that brains run on 20 watts while Frontier needed 21 million watts.
The scientific pushback and the consciousness problem nobody can shake
Dylan is careful here: the Pong work was peer-reviewed, but the Doom demo hasn’t appeared in the same kind of paper, and critics argue the system may show short-term plasticity more than anything like true learning. Then he lands on the part that really haunts him: if these systems are useful precisely because they’re biologically human-like, bioethicist Hank Greely asks whether they might also be human-like enough to suffer — making the Doom demo feel less like a meme and more like “some new kind of interface being born.”