Amazing: a chip integrating 200,000 human neurons capable of playing Doom

In the midst of the artificial intelligence era, an Australian project challenges our perception of computing and cognition: a chip integrating approximately 200,000 human neurons cultivated in the laboratory plays Doom, the legendary iconic video game. This is not a classic AI coded for the occasion, but a true biological network – a novel form of advanced technology combining the fields of neurosciences, neural computing, and brain-machine interfaces. Cortical Labs, the responsible startup, highlights a radically different path for the computer of tomorrow: reinvesting the living into computing power.

While traditional artificial intelligence solutions require massive energy and hardware resources, this biological prototype opens a window on forms of adaptive learning capable of combining energy efficiency and intrinsic plasticity. The CL1, this system integrating neurons and silicon, is not a mere curiosity but a major milestone towards a future where machines are inspired not only by the brain but by its living components.

This fascinating journey into the convergence of life and technology questions our certainties about the boundary between natural and artificial, with at stake perspectives that are as exciting as they are worrying regarding ethics and the status of hybrid entities. Here is a detailed exploration of what a neuronal chip capable of playing Doom involves, and what it heralds for upcoming innovations.

The neural chip: a revolution in neural computing with human neurons

The innovation proposed by Cortical Labs disrupts the landscape of neural computing by directly integrating living human neurons on a silicon chip, thereby fusing biology and engineering into a potentially autonomous system. These brain cells, cultivated from stem cells, form a network capable of emitting electrical signals that are received and interpreted by an electronic system, thus creating a unique communication loop in the computing world.

This approach is radically different from classical artificial intelligences, which only simulate neural networks through code and purely digital architectures. Here, computing is no longer imitation: it is an effective bioelectrical process, prioritizing the natural plasticity of neurons to learn and adapt to complex tasks. This brain-machine interface system offers a novel mode of interaction between the living and the virtual.

Thanks to a system called biOS, the neurons “live” in a generated virtual environment simulating their natural interactions. This neural chip also stands out by its ability to maintain these neurons for several months thanks to an autonomous support system, a major technical feat.

A concrete example of this success is the adaptation of gameplay in Doom, where the neural network learns to perform actions such as moving or shooting simply from electrical stimulations. This is not predefined learning but an adaptive capability linked to the given “goal.” This is no longer a simple code-based AI but a biological computer capable of a form of real-time intelligence.

The following table illustrates the key differences between a classic computer and this biological neural chip system:

Criterion	Classic Computer (Digital AI)	Biological Neural Chip (CL1)
Computing support	Silicon + algorithms	Living human neurons + silicon
Type of processing	Algebraic, simulation	Bioelectrical, adaptive plasticity
Autonomy	Active cooling required	Autonomous, biological, passive cooling
Duration of maintenance	Indefinite, as long as hardware functions	About 6 months with integrated life support
Energy consumption	Very high (megawatts for advanced AI)	Estimated at a few watts, bioenergetic operation
Learning	Supervised or not, data-based	Goal-oriented adaptive learning in closed loop
Interaction with the real world	Via classic peripherals	By direct biological electrical stimulation

Thus, this neural chip materializes an unprecedented innovation in neural computing, revealing new potentials to combine biology and technology. This direction could revolutionize not only data processing but also the way we conceive artificial intelligence.

How living neurons learn to play Doom: decoding the brain-machine interface

Asking how a set of living neurons, deprived of eyes, hands, or traditional sensory system, can learn to play Doom leads to a fascinating discovery about the functioning of brain-machine interfaces. The trick lies in how game data are transformed into electrical stimuli directly injected into the neural network.

The CTO of Cortical Labs, David Hogan, explains that the game’s video stream is converted into complex patterns of electrical impulses that act like a “brain language.” These signals are perceived by the group of neurons which react by modulating their electrical activity. The system then detects certain configurations of this activity as commands: move forward, turn, shoot.

The major technological challenge was to establish a feedback loop where neurons receive stimulation, produce a response, and this response is in turn interpreted to adjust the stimuli. This is how the network “learns” not through the quietness of a rigid program, but through continuous adaptation to a defined goal, here progressing in the video game.

A somewhat figurative parallel: a player without a controller who learns by touch and feeling in a virtual universe, guided only by tactile perception of obstacles and action feedback. This “sensory” learning mode is highly rich as it exploits the natural plasticity of the human brain, even through a reduced and inanimate model.

The video presented by Cortical Labs illustrates this feat: the neural network manages to rudimentarily play Doom, moving the virtual player and shooting at targets. This achievement does not aim for a score or competition but the tangible demonstration that such a system can adapt in real time to a complex task. Intelligence is thus seen as a dynamic process, not a fixed performance.

This approach offers a new perspective on machine learning, symbolizing a bridge between biology and software, but also a potential cognitive revolution in research on brain-machine interfaces.

Doom: the ultimate test for synthetic biological intelligence

Going from the simple game Pong to Doom represents a prodigious escalation of complexity for a living neural network on a chip. When Cortical Labs presented DishBrain in 2022, neurons had already demonstrated mastery of Pong, a simple game with limited control and response elements.

Doom, on the other hand, offers a 3D environment with many more unforeseen elements: multiple movements, enemies, variable objectives, maze exploration, complex action management. This technical and cognitive progression symbolizes a genuine qualitative leap in the capability of a biological network to master complex tasks.

To train for this complexity, Cortical Labs designed an infrastructure called “Cortical Cloud” allowing the management of multiple neural networks for multiple tasks, thus showing a clear will to extend this technology beyond simple games.

This stress test of biological neural computing is fundamental: it shows that the technology is not a one-off demonstration but is moving towards practical applications in adaptability and real-time learning, offering different formats of intelligence than purely “digital” AIs.

This shift from playful to applicative opens the way to hybrid systems capable of various tasks, in fields such as robotics, personalized medicine, or autonomous management in complex environments.

Energy and technological advantages of the neural bio-computer

While classic AIs such as large language and artificial vision models consume colossal amounts of energy and heavily strain IT infrastructures, the biological neural chip appears as a green and efficient innovation. A living human network naturally consumes much less energy per computing unit than purely digital solutions and knows how to optimize its processing mechanisms.

The human brain consumes on average about 20 watts to manage a broad spectrum of complex functions, combining perception, memory, movement, and adaptation. This reference strongly inspires engineering efforts around neural chips and neural computing.

The technology of the CL1 does not aim to produce “more powerful” intelligence than a top-tier GPU but to explore a new form of computer capable of continuous learning at very low energy cost, thanks to natural biological plasticity.

Here are some key points highlighting the strengths of the bio-computer:

• Reduced consumption: little electrical energy required, no active cooling.
• Natural adaptability: neurons adjust their connectivity and dynamics based on stimulations.
• Extended autonomy: maintenance of neurons alive up to 6 months without major intervention.
• Closed-loop operation: dynamic interaction between software and biological network.
• Potential applications: real-time learning, adaptive control in changing environments.

In a context where data centers and AI infrastructures compete in energy power, this approach opens a door to a more environmentally respectful future and potentially more efficient for certain types of computation.

The other paths of biological and neuromorphic intelligence: a global overview

Cortical Labs’ project fits into a broader movement aiming to bring biology and computing closer together to transcend the limits of classical silicon. Several approaches coexist:

1. “Wetware” systems: these platforms exploit brain organoids or living neural networks cultivated for complex bioelectronic processing. For example, Brainoware, cited in Nature Electronics, highlights their potential in temporal and spatiotemporal data processing.
2. Remote neural interfaces: like FinalSpark’s Neuroplatform, allowing access to and control of biological networks for research and development.
3. All-silicon neuromorphic: chips inspired by the brain but without living cells, using for example spiking neural networks such as Intel Loihi 2 or Hala Point, the latter integrating more than a billion artificial neurons.

Each of these avenues aims to achieve greater efficiency, adaptability, and better data management while limiting energy consumption. Cortical Labs’ project, with its neural chip integrating human neurons, thus brings a unique dimension, with a true biological network at the heart of a computing platform.

Ethical issues surrounding biological intelligence integrated on chip

Manipulating human brain tissues and cultivating them in networks naturally raises fundamental ethical questions, especially regarding moral status and governance. To what extent do these networks possess a minimal form of consciousness or sensitivity?

Recent academic debates recommend a cautious approach, with strict governance for these technologies, notably with clear frameworks regarding cellular origin and donor consent.

Currently, the culture of neurons present on the CL1 shows no indicators of conscious experiences or emotions. But as these systems gain complexity, understanding their moral impact becomes an urgent necessity.

The emergence of technologies combining biological and digital also invites reconsideration of our definitions of intelligence, life, and associated rights. This debate also crosses the development of brain-machine interfaces, calling for transparency and multidisciplinary reflection.

Future applications and disciplinary convergence for new forms of intelligence

Beyond the spectacular feat of playing Doom, this prototype opens many avenues towards concrete applications:

• Adaptive robotics: machines controlled by living neural networks capable of real-time adaptation to complex environments.
• Personalized medicine: modeling patient neural networks to test effects of neurotherapeutic treatments in the lab.
• Advanced human-machine interfaces: improving prosthetics and neural stimulation through direct interaction with cultured neurons.
• Fundamental research: exploring cognition, plasticity, and biological learning mechanisms through a hybrid model.
• Ecological neural computing: towards less energy-hungry machines performing well in specific tasks.

This convergence of neurosciences, computer engineering, biotechnology, and robotics perfectly illustrates the complexity and richness of upcoming projects, where the boundary between humans and machines becomes more porous than ever.

What is a neural chip integrating human neurons?

It is a hybrid computing platform where human neurons cultivated in the laboratory are combined with a silicon chip, allowing bioelectrical processing capable of autonomous learning and adaptation.

How can these neurons play Doom without a classic sensory system?

The game’s video stream is converted into electrical impulses sent to the neural network, which responds with activity configurations interpreted as game commands, creating an adaptive learning loop.

What are the energy advantages of this technology?

It consumes much less energy than classic AIs thanks to the natural plasticity of living neurons, eliminating the need for active cooling and allowing prolonged autonomy.

What are the ethical issues related to this technology?

Cultivating human brain tissues raises questions about the moral status of neural networks and the need for strict frameworks concerning consent, governance, and minimizing risks of artificial consciousness.

What are the future applications for this neural chip?

It could revolutionize adaptive robotics, personalized medicine, human-machine interfaces, neuroscience research, and pave the way for less energy-consuming and smarter machines in specific domains.