A living computer? This AI made from interconnected "mini-brains" requires a million times less power than ChatGPT!

Every time you have a discussion with ChatGPT about why your tummy hurts or when the next season of your favourite show will arrive, the application guzzles down an entire half-litre bottle of freshwater. This helps the artificial intelligence keep itself cool and composed while it answers your “hard-hitting” questions. Despite this scary bit of information being common knowledge for a while now, the AI boom has only gathered more *ahem* steam in recent years.

As AI technology advances, the energy consumption of the IT industry is also projected to spike significantly, potentially reaching 3.5% of global electricity by 2030. This is undoubtedly a lot of energy and a mountain of toxic carbon dioxide — levels of emissions we certainly cannot afford at the moment.

However, our brains, despite still being far more complex and efficient than current computers, clearly don’t require nearly the same amount of energy to function. To emulate the supreme model organ, scientists have pulled a move straight out of science fiction: they grew brains of their own to hook up to a computer!

In a massive development, a Swiss start-up named FinalSpark has launched a 'biocomputer' that connects to living, pulsing brain cells. Their online platform “taps” into spherical clusters of lab-grown human brain cells about half a millimetre in diameter, which are kept alive via an electrode and a microfluidics system that supplies the necessary water and nutrients. And yes, this system consumes far less energy than traditional, bit-based computers.

This innovative approach, referred to as wetware computing, utilises researchers' ability to culture organoids in the lab. Organoids, essentially mini replicas of individual organs, have gained popularity as a research technique alongside the rise of energy-intensive artificial neural networks such as ChatGPT.

The electrodes can supply electric pulses to the brain organoids and measure how the cells respond. These organoids are exceptional at determining patterns, which is the kind of processing very helpful for computer engineers trying to solve their many problems. During the training process, the mini-brains can also be “rewarded” with dopamine rushes and “punished” using electrical stimulation.
Far less energy than traditional digital processors
FinalSpark claims that their bioprocessors "consume a million times less power than traditional digital processors." To put this into perspective, training a single large language model like GPT-3, a precursor to the latest GPT-4, required 10 gigawatt hours of energy — roughly 10,000 times the per capita energy consumption of an average Indian citizen in 2023. In contrast, the human brain operates its 86 billion neurons with just 0.3 kilowatt hours per day, less than 30 times as much energy as a light bulb!

Currently, FinalSpark's system is being used to run lengthy experiments on brain organoids. Researchers can connect to this system remotely, and the mini-brains can be sustained for up to 100 days with their electrical activity continuously monitored. The system is freely available for research purposes, and numerous research groups have already begun using it.

However, it may be noteworthy that FinalSpark is not the first to attempt connecting probes to biological systems or programming neural networks to perform specific tasks. In 2023, US researchers built a bioprocessor that connected computer hardware to brain organoids, enabling the system to learn to recognize speech patterns. Another Australian organisation created the DishBrain device that learned to play Pong within five minutes.

Looking ahead, FinalSpark plans to expand the capabilities of their platform to manage a broader range of experimental protocols relevant to wetware computing. This includes injecting molecules and drugs into organoids for testing. Whether aiding computing or organoid research, the potential achievements in this field are vast and exciting.

The findings of this study have been published in Frontiers in Artificial Intelligence and can be accessed here.