As generative AI systems progress, so does their appetite for energy. Training and running large-scale language models consumes a huge amount of power. AI energy demand is projected to double over the next five years, increasing 3% of global total electricity consumption. But what if the AI chip functions like a human brain and can handle complex tasks with minimal energy? The growing chorus of scientists and engineers believe that the key may lie in organoid intelligence.
AI enthusiasts were introduced to the brain-inspired concept of chips in July at the UN AI AI in Geneva. There, David Gracias, professor of chemistry and biomolecular engineering at Johns Hopkins University, gave a speech discussing the latest research into biochips and AI applications. A focus on Nanotech, Intelligent Systems and BioEngineering, Gracias research team can build functional biochips that combine neural organoids with advanced hardware, allowing the chip to run and interact with living tissues.
Organoid Intelligence is an emerging field that blends lab-grown neurons with machine learning to create new forms of computing. (The term “organoid intelligence” was coined by Johns Hopkins researchers, including Thomas Harton.) Neurons called organoids are, more specifically, three-dimensional clusters of brain cells grown in laboratories that mimic neural structure and function. Some researchers believe that so-called biochips (organoid systems that integrate live brain cells into hardware) could outperform silicon-based processors such as CPUs and GPUs in both efficiency and adaptability. When commercialized, experts say biochips can potentially reduce the incredible energy demand of today's AI systems, whilst increasing learning capabilities.
“This is an investigation into another way of forming a computer,” Gracias said.
How does a biochip mimic the brain?
Traditional chips have long been limited to two-dimensional layouts. This allows you to limit the way signals flow through the system. This paradigm is beginning to change as chip makers are currently developing 3D chip architectures to increase the processing power of their devices.
Similarly, biochips are designed to emulate the three-dimensional structure of the brain itself. The human brain can support neurons with up to 200,000 connections. Level of Interconnectivity Gracias says flat silicon chips cannot be achieved. This spatial complexity allows biochips to transmit signals across multiple axes, allowing for more efficient information processing.
The Gracias team has developed a 3D EEG (EEG) shell that is wrapped around an organoid and allows for more stimulation and recording than traditional flat electrodes. This cap fits the curved surface of the organoid and creates a better interface for stimulating and recording electrical activity.
To train organoids, the team uses reinforcement learning. The electrical pulse is applied to the target area. When the resulting neural activity matches the desired pattern, it is enhanced with dopamine, a natural reward chemical in the brain. Over time, organoids learn to associate specific stimuli with outcomes.
Once you have learned the patterns, they can be used to control physical actions, such as steering miniature robotic vehicles through strategically placed electrodes. This demonstrates neuromodulation, the ability to generate predictable responses from organoids. These consistent responses lay the foundation for more advanced features, such as facial recognition, decision-making, and stimulus identification, which are essential for applications such as generalized AI inference.
The Gracias team is in the early stages of developing a miniature self-driving vehicle controlled by Biochips. This is proof of the concept that a system can act as a controller. This experimental study suggests future roles in robotics, prosthetics, and biointegrated implants communicating with human tissue.
These systems also offer promising disease modeling and drug testing. The Gracias group is developing organoids that mimic neurological diseases such as Parkinson's disease. By observing how tissues of these diseases respond to various drugs, researchers can test new treatments on dishes rather than relying solely on animal models. It can also elucidate the potential mechanisms of cognitive impairment, which current AI systems cannot simulate.
These chips are alive and require constant care: temperature control, nutrient supply and waste removal. The Gracias team continuously monitors integrated biochips to maintain integrated capabilities for up to a month.
Fred Jordan (left) Martin Kitter is the founder of Final Park, and the Swiss startup is developing a biochip that the company claims can store data in living neurons.Final Park
Scaling biochip technology challenges
However, important issues remain. Biochips are fragile, highly maintained, and current systems rely on bulky lab equipment. Reducing them for practical use requires biocompatible materials and techniques that can autonomously manage life support functions. Neural latency, signal noise, and scalability of neuron training also present hurdles for real-time AI inference.
“There are a lot of biological and hardware questions,” Gracias says.
Meanwhile, some companies are testing their bodies of water. Swiss Startup Finalspark claims that biochips can store data in living neurons. This is a milestone known as “Biobit.” This step suggests that biological systems can one day perform core computing functions traditionally processed by silicon hardware.
FinalsPark aims to develop remote-accessible biocell bars for general computing in about 10 years. The goal is to match performance digital processors while being exponentially energy efficient. “The biggest challenge is programming neurons because we need to know a whole new way to do this,” says Kurtys.
Still, moving from the lab to the industry requires more than just a technical breakthrough. “We have enough funds to keep the lab up and running,” Gracias says. “But we need more funding from Silicon Valley to begin research.”
It is still unknown whether the biochip will enhance or replace the silicon. But as AI systems demand more and more power, the idea of brain-like thinking chips and SIP energy is becoming more and more attractive.
In the case of Gracias, the technology can be shipped to the market faster than we think. “There are no major show stoppers on the way to implement this,” he says.
From the article on the site
Related articles on the web
