Video Summary
Modern GPUs are reaching physical limits; researchers are exploring biological neurons as an alternative computing substrate.
Dishbrain was an early prototype that taught cultured neurons to play Pong but required ~1M cells and long training times.
CL1 reduces neuron count to ~200,000 with improved stability and an integrated microfluidic life-support system.
Neurons are grown from human iPSCs, arranged as a 2D monolayer on high-density microelectrode arrays for two-way communication.
Learning is driven by patterned electrical stimulation: stable signals reward useful actions, chaotic noise penalizes bad actions (free energy principle).
CL1 is an improved living-neuron chip that reduced the required neurons from ~1 million (Dishbrain) to about 200,000, added robust life-support (microfluidics, gas control, temperature), and achieved more stable, faster learning.
Neurons are produced in the lab from human-induced pluripotent stem cells (iPSCs), allowing researchers to grow neural cultures without harming donors.
Neurons are seeded as a 2D monolayer on a high-density microelectrode array, enabling two-way electrical stimulation and recording between the culture and the device.
Learning uses patterned electrical feedback: stable, predictable signals are used as rewards for useful actions while chaotic noise penalizes poor actions—an application of the free energy principle.
CL1 uses a microfluidic perfusion circuit to deliver nutrients and remove waste, precise temperature control (~37°C), and automated gas mixing to regulate oxygen and CO2.
Key concerns include whether cultured human neurons could develop sentience if scaled, potential suffering from constant stimulation, and moral questions about using living tissue for computation or entertainment.
"Current AI systems are designed from neural networks, inspired by the brain, but run inefficiently on GPUs."
"Researchers have built a chip powered by living human neurons and even taught it to play the video game Doom."
"Dishbrain was their original proof of concept, developed around 2021 and 2022, showing that clusters of human and mouse neurons could learn to play Pong."
"CL1 dramatically reduced the number of required neurons from almost one million to just 200,000 while maintaining stable learning behavior."
"To keep neurons alive, a microfluidic perfusion circuit was designed to deliver nutrients and oxygen while filtering out waste."
"The brain cells are created in the lab using human-induced pluripotent stem cells, a technique that won a Nobel Prize."
"Neurons are carefully placed into a 2D monolayer on a high-density microelectrode array, creating a two-way communication device."
"Manis isn't just another AI chatbot; it's a proactive AI agent designed to deliver real tangible results."
Manis autonomously analyzes tasks by understanding goals, planning steps, and executing the entire task.
It allows for seamless automation, such as converting a Google earnings report into a professional presentation or transforming a complex taxonomy tree into an interactive graph.
Manis can interact with various platforms like Gmail and Google Drive, simplifying file management without the need to switch between different applications.
"The CL1 system has around 200 neurons, double what a fruit fly has."
The CL1 system, which contains 200 neurons, demonstrates capabilities comparable to the brain of a complex insect by executing tasks like playing a 3D first-person shooter game.
The neuron network interacts with the game by translating visual gameplay data into electric signals through a process that involves encoding and decoding these signals for gameplay commands.
"Whenever it does a useful action in the game, the system delivers a smooth, predictable electrical signal."
Learning for a dish of neurons is motivated by electrical signals rather than the chemical rewards typical in a living brain.
Researchers employ the free energy principle, rewarding neurons with stable signals for positive actions and chaotic noise for negative actions to encourage adaptive learning.
This method enables the neural network to learn and optimize its behavior in the game environment.
"This living neural chip can learn incredibly fast."
The biological neural network adapts to game rules and begins to form patterns quickly, demonstrating a learning speed that outpaces traditional AI models which require extensive training data and iterations.
In contrast to modern AI models that consume vast amounts of energy, the efficiency of biological systems, like the human brain operating at only 20 watts, suggests a promising alternative that could significantly reduce energy consumption in AI computations.
"Tasks that humans find difficult, like playing chess, are easy for computers, while tasks that are easy for humans, like walking, are difficult for robots."
The Moravec Paradox highlights how certain tasks pose challenges for artificial intelligence and robotics. Humans excel at physical tasks despite their complexity, while computers can outperform humans in tasks like strategic thinking.
The paradox arises because real-world environments are often unpredictable and messy, making it challenging for robots to process information about force, friction, and balance in real-time.
Biological systems, such as human brains, have evolved to handle these intricate physical motions effectively.
"With living neuron chips, there's huge potential for them to be used in robotics or other physical AI systems."
Living neuron chips offer the prospect of incorporating biological intelligence into robotic systems, enhancing their ability to navigate unpredictable environments.
These chips could help overcome some technological limitations currently faced by robots, allowing for advanced physical tasks like walking or delicate object manipulation.
"Is it ethical to constantly stimulate living neurons to play video games or fulfill prompts?"
The integration of living human brain cells into technology raises significant ethical questions regarding consciousness and sentience.
Current AI models exist in software and silicon and are not thought to be conscious. However, the use of biological neurons blurs these lines, leading to concerns about the ethical implications of potentially "torturing" living material for entertainment or productivity.
"What if you scale up these living brain chips? Would it be conscious then?"
Researchers argue that a functioning brain is not simply a collection of neurons, but a complex system with specialized structures and functions.
The possibility of many living neuron chips being used together in data centers poses the question of whether such systems could achieve consciousness. This raises further ethical considerations regarding their treatment and rights.
The implications of scaling these systems could trigger critical discussions about technology and morality as it relates to artificial intelligence and living organisms.
