What still has a hell of a head start on artificial intelligence? The human brain. And precisely, researchers plan to use human brain cells in biocomputers to design the AIs of tomorrow.
Since the advent of ChatGPT, we talk regarding it all the time. Artificial intelligence is now everywhere. She knows how to code, write essays, draw, make music, speeches and even write books. But AI has also shown its limits. So to go further and make sure we better understand what we have in our heads, scientists at Johns-Hopkins University in the United States have added a bit of real “brain juice”.
They have designed what they call an organoid intelligence, or IO. It may be scary, but there really are human brain cells in this IO. The organoids in question would be clusters of living tissue that come from stem cells. They would be nourished by microvasculature with fluids. Their electrical signals would then be picked up by micro-electrodes. Finally, the result would be analyzed by algorithms that rely on AIs as we know them. By comparing several of these pieces of “brain”, the researchers imagine that the power might ultimately be similar to that of the human brain. This biocomputer with 3D cellular structures is described in an article published in the journal Frontiers in Science. The lab has even developed rudimentary prototypes of IO. They are not even vascularized and remain very fragile.
A biocomputer to consume little with colossal power
To communicate directly with this organoid, the team developed a brain/computer interface inspired by EEG (electroencephalogram) helmets. Developed last year, this flexible shell that envelops the neuronal matrix uses an incredible density of tiny electrodes to pick up signals from the organoid and see what it is “thinking”. For the moment, given its small size, this pseudo-brain remains far from ChatGPT’s performance. It only has 50,000 cells, which is already quite a feat. But it would take more than 10 million, or 125,000 billion synapses for the IO to be able to pulverize the current AIs. Each of these synapses would be able to store up to 4.7 bits of information. Better than the most powerful supercomputer in the world with its 58 billion transistors that can store 4 bits each.
But, as the publication explains, this IO would however have a big advantage, that of low power consumption. It might approach that of a human brain which, with 1.4 kilos, consumes only 20 watts for a computing power of approximately 1 exaflops. On the machine side, this record was only reached by a supercomputer last summer. But it took a phenomenal amount of energy to achieve this, without even mentioning the colossal dimensions of the whole. So, by combining the best of these two worlds, the laboratory considers that learning the IO would be enormously easier compared to a conventional machine, with in addition this reduced energy consumption. But, for the moment, current AIs have plenty of time to carry out their revolution. That of these OIs should not begin for at least ten years according to the researchers.