Живые нейроны запускают Doom и учатся | Владимир Алипов
Vladimir Alipov discusses Cortical Lab's CL1 processor, which uses living human brain cell cultures to run computations, including the game Doom. He explains how biological neurons learn through reinforcement, where predictable structured signals serve as positive reinforcement, and covers both the advantages and limitations of biological neural networks compared to artificial ones.
Summary
The video features Vladimir Alipov discussing Cortical Lab, an Australian company that grows living neuron cultures on specialized chips and uses them for computation and learning research. Their latest achievement is running the video game Doom on a chip called CL1, which is built on living human brain cells — continuing the internet meme tradition of running Doom on unconventional platforms.
Alipov outlines the key advantages of biological neural networks over artificial ones: dramatically lower energy consumption, faster training, the ability to retrain on new tasks (something artificial neural networks struggle with), and greater versatility across problem types. He also humorously notes they won't require massive water cooling systems like data centers do.
However, he also covers significant disadvantages: you cannot copy or scale a trained biological network the way you can with software — a new one must be trained from scratch. Additionally, the cultures are fragile and can die if nutrient solutions, temperature, or acidity are disrupted.
Alipov walks through Cortical Lab's scientific journey — they first trained neuron cultures to play Pong in 2022, with that research published in the prestigious journal Neuron. Doom represents a much harder challenge due to its 3D environment and far greater number of inputs. Game state information is encoded into electrical signals fed to the neuron plate, and the neurons' response patterns are decoded into in-game actions like shooting or turning.
A central scientific question Alipov highlights is how feedback is delivered to the cell culture — i.e., how neurons are told whether they did something right or wrong. Surprisingly, the 2022 article found that a predictable, structured signal itself acts as positive reinforcement for neurons, with no need for chemical rewards like glucose or dopamine. Alipov finds this deeply significant, drawing a connection to Karl Friston's 'prediction machine' theory of the brain — suggesting that for biological systems, living in a predictable world is inherently rewarding. He extends this to human psychology, noting that anxiety often stems from excessive uncertainty, while learning is essentially the process of making an unpredictable world more predictable.
The video also briefly touches on what neuron cultures are — cells grown in Petri dishes, fed nutrient solutions, and naturally inclined to form synaptic connections with each other. Alipov shows images of neurons in culture and explains that the chip used by Cortical Lab has electrodes that both stimulate and read signals from specific anatomical regions of the culture.
Key Insights
- Alipov explains that a predictable, structured signal alone acts as positive reinforcement for neurons in culture — no chemical reward like glucose or dopamine is needed — suggesting that predictability itself is intrinsically rewarding to biological neural systems.
- Alipov notes that unlike artificial neural networks, biological neuron cultures cannot be copied or scaled — a new culture must be trained from scratch each time — which he identifies as a critical practical disadvantage.
- Alipov connects the neuron culture findings to Karl Friston's 'prediction machine' theory, arguing that learning for both cultured neurons and human brains is fundamentally the process of making an unpredictable world more predictable, and that anxiety arises from living with too much uncertainty.
- Alipov describes how Doom is encoded into electrical signals fed to the neuron plate, and the neurons' response activity patterns are decoded into in-game actions — but notes that current performance is roughly equivalent to a complete beginner who has never played a shooter before.
- Alipov points out that Cortical Lab's CL1 device is sold commercially for around $5,000–$10,000 and was originally positioned as a scientific instrument for studying learning processes in neuron cultures, but has since shifted toward demonstrating actual computational use cases.
Topics
Full transcript available for MurmurCast members
Sign Up to Access