Swiss bioinformatics startup FinalSpark recently announced the launch of a cloud computing platform like no other, and for good reason: it is built around a bioprocessor composed of 16 synthetic human mini-brains who cannot think or feel anything, but still functional.
For decades, researchers have been trying to reproduce the functioning of the brain to exploit its formidable capabilities. Today, these efforts are mainly embodied by the artificial neural networks which underpin the systems of machine learning.
But bioinformaticians have not given up yet. Several laboratories continue to develop another approach based on organoids — small clusters of specialized living cells grown in a laboratory that mimic the functioning of an organ. Swiss startup Final Spark belongs to this category. His “computer” is built around several cerebral organoids, themselves composed of human neurons.
This is not the first time that such an entity has been created in the laboratory, far from it. In recent years, other teams have already created organoids which, once connected to another system, have proven capable of playing Pong, recognizing words or even solving mathematical equations.
Energy, a central issue in AI
FinalSpark, on the other hand, took the concept significantly further with a system composed of 16 distinct brain organoids. We can therefore consider it as a multi-core organic bioprocessor. The researchers believe that once mature, this approach could constitute an interesting alternative to systems of machine learning which are swarming at the moment.
Fundamentally, both technologies rely on the same mechanisms. They exploit a large number of subunits, virtual or biological neurons, which form networks whose architecture evolves when they are subjected to a signal. This is how humans or modern “artificial intelligences” learn to process data. But there are several major differences between the two approaches, including one that is absolutely crucial: the amount of energy they use.
Since they are unable to reproduce the nuances that make the brain’s architecture so efficient, computer scientists have only one approach available to emulate its functioning: rely on brute force. This is why they are developing very extensive artificial neural networks. For example, GPT-4 uses around 1670 billion parameters, compared to just 85 billion neurons in the human brain.
However, training a system of this type requires a significant amount of energy which increases exponentially with complexity. Training large language models (LLMs) like GPT, for example, is typically measured in tens of gigawatt hours; enough to power a few thousand homes for an entire year. A very important consideration, knowing that these systems are becoming more and more common. Last year, a study suggested that machine learning could consume nearly 4% of global electricity in 2030. So there is an urgent need to find ways to curb the all-consuming appetite of AI models.
Bioinformatics, a potential alternative?
To date, the industry is mainly looking for software solutions. Many researchers consider that it will be possible to enormously reduce energy bills by optimizing current algorithms, or by developing entirely new architectures. FinalSpark, for its part, seems convinced that biological media could be part of the solution. After all, if evolution has taken the trouble to optimize brain energy consumption for millions of years, why reinvent the wheel?
It must be recognized that the idea makes sense. The human brain, for example, is much more economical than an AI model. Even if it consumes around 20% of the body’s total energy, this only represents around a hundred kWh per year and per individual. Likewise, the startup claims that at equal power, an organic bioprocessor is a million times less energy-intensive than a traditional processor.
A “neuro-cloud” at the service of research
Obviously, it is not tomorrow that general public computers will be able to operate with this type of bioprocessor. But to be able to start exploiting them in real work, we must first bring the concept to maturity. And this is where FinalSpark’s work becomes particularly interesting.
In fact, the researchers connected their bioprocessor with 16 organoids to an online platform that allows researchers around the world to use it to conduct artificial intelligence experiments remotely. A sort of neuro-cloud computing in the service of research, in short. It will therefore be appropriate to follow the results of this approach which could open the way to some quite fascinating scientific publications over the coming months.
The text of the study is available here.
-
-
