Human brain organoid-based AI processors get closer to becoming reality
The company FinalSpark is now offering researchers access to their experimental human brain organoid processors for $500 per month
Last year I wrote about the emerging field of “Organoid Intelligence” — a field that aims to use living clusters of human brain cells (brain organoids) at the heart of advanced AI systems.1
This felt quite speculative at the time, but just over a year later the company FinalSpark is providing researchers 24/7 remote access to actual brain organoid-based processors.2
FinalSpark has been making waves for a few months now with a computational platform based on brain organoids. Back in May, researchers with the company published details of their remotely accessible platform for research in “wetware computing” in the journal Frontiers in Artificial Intelligence.
And just recently, they announced that researchers in universities can access the platform for just $500 per month.
To be clear, this is not the start of some massive integration of human brain tissue into the next generation of AI. Rather, it’s a chance to crowdsource research into how to take advantage of these highly novel biological processors.
According to FinalSpark, the big advantage of their organoid-based platform is massively lower energy overheads compared to AI systems based on digital computing.
For instance, it took around 1.3 GWh of energy for OpenAI to fine tune GPT3 according to the Stanford-based Human-Centered Artificial Intelligence initiative (HAI). In contrast the human brain typically uses a mere 0.3 KWh per day.
In other words, a person would need to live for nearly twelve thousand years for their brain to use as much energy as it took to fine tune GPT3!3
The potential energy savings are a big deal as AI systems become ever-more power hungry. But I suspect that the advantages of AI powered by brain organoids will far surpass the energy use advantages, if only because the underlying processes will be markedly different from those in a artificial neural network.
This is perhaps most obvious in the connections between neurons and how this impacts behavior and performance. In a digital neural net, the influence of these connections — the weights within the model in essence — are adjusted during training, and can be rewritten at will.
The same doesn’t hold for a collection of biological neurons. Here, the “weights” and the network of connections grow and change over time. And they can’t simply be tweaked and re-written.
The upshot of this is that new approaches to utilizing these “wetware” neural networks are needed. And this is exactly why FinalSpark are making their platform openly available to researchers.
For $500 per month users get to experiment with the how these brain organoid processors work, and how to develop and “code” them — all through an online Python-based interface.
As well as being a fantastic research opportunity, the cumulative knowledge will help better-understanding how to use these biological processors in ways that allow them to be scaled and used in novel ways.
The processors themselves that FinalSpark are making available consist of four organoids, each with eight electrodes attached to them. As the researcher, you write code that feeds electrical signals (stimulations) to the organoids and reads the responses — and then work on teasing out the relationships between the two.
And you do all of this via the Open Neuroplatform online interface.
By using the platform researchers are able to set up complex organoid array stimulations and read/analyze the resulting responses, much as neuroscientists do when stimulating or monitoring neurons in living brains.
To give you a taste of what the experimental setup looks like, FinalSpark has a webcam set up on their organoid arrays, and live feeds from the electrodes. It doesn’t mean much just looking at the system. but it does show you that real stuff is happening in real time!
(As well as seeing the array in real time by clicking the image above, you can also access an alternative feed here.)
Just being able to compute with inputs and outputs to an array of human brain organoids is impressive in itself. But these arrays have another trick up their sleeve which underlines what makes working with living brain cells so unique.
As well as an electrical read/write system, it’s also possible to program the release of molecules into the organoids that modulate their behavior — much as chemicals in our brains modulate neuron behavior.
The process is referred to as “molecular uncaging.” and as FinalSpark describe in their paper:
It is also possible to release molecules at specific timings using a process called uncaging. In this method, a specific wavelength of light is employed to break open a molecular “cage” that contains a neuroactive molecule, such as Glutamate, NMDA or Dopamine. A fiber optic of 1,500 μm core diameter and a numerical aperture of 0.5 is used to direct light in the medium within the MEA chamber. The current system, Prizmatix Silver-LED, operates at 365 nm with an optical power of 260 mW. The uncaging system is fully integrated into the Neuroplatform and can be programmatically controlled during experiment runs via our API.
This in effect gives researchers the ability to temporally change the behavior of these organic neural nets in ways that hve no direct equivalence in digital neural nets.
At this point it’s hard to predict where this technology will lead — and we’re unlikely to have a good sense of this until people start to experiment extensively with the systems FinalSpark is making available.
But there is clear promise here, as well as some potential pitfalls.
The promise — if organoid-based computing turns out to be viable and scalable — is AI systems that are exceptionally energy efficient, and use substrate architectures that open up possibilities beyond those available in non-biological systems. This includes the possibility of shifting AI technologies closer to more human-like capabilities.
And the downsides? These are likely to be complex and hard to identify in the near term. But any technology that depends on connecting living clusters of human brain cells to compute substrates with the aim of making them a pillar of advanced AI, raises questions that range from the ethics of using what are in effect “proto-brains” to the responsibility that comes with developing AI systems that are not only biological, but use human tissue as their foundation.
It’s also worth asking what the longer term consequences might be if we crack using human neurons at scale to power AI systems. If we can do this with brain organoids growing in petri dishes (or MEMS4 devices), why stop there? What might be possible if we tap into real brains inside the heads of real people?
This may sound like scare-mongering sci-fi, and to be sure there’s no talk at the moment of any research using anyone’s brain as an AI processor. But given the progress that companies like Neuralink and others are making toward brain computer interfaces that connect to tens of thousands of neuron clusters, I wonder how long it will be before someone considers monetizing these by suggesting users quite literally rent out their brain-time …
That’s a can of worms that we’re probably decades or more from opening. But if and when human brain organoid computing at scale becomes a reality, it’s a question that we’ll probably need to address.
In the meantime, it’s going to be interesting to see what researchers achieve with their $500 per month subscriptions to FinalSpark’s existing brain organoid processors — especially if it ushers in a new wave of organic AI.
I also wrote about a paper in December last year that showed the use of brain organoids in constructing simple neural nets that can be trained to carry out basic functions. More here.
Many thanks to my colleague Sean Dudley for alerting me to this!
These estimates differ from those in the FinalSpark paper, but the end result is similar — brain tissue is potentially way more efficient as an AI substrate than GPUs.
MEMS — Micro-Electro-Mechanical Systems.
You're earlier writing on this topic inspired my inclusion of organoid computing in my second Sci-Fi novel about AI and what it means to be human!
You can find it on Amazon: https://amzn.to/474x5Ss
Maynard writes, "It’s also worth asking what the longer term consequences might be if we crack using human neurons at scale to power AI systems."
These are the kind of questions that interest me, thanks.
Presumably we would move to using human neurons to power AI systems if that method delivered some advantage over using digital systems. We would use neuron powered systems if they were faster, cheaper, more powerful etc.
The question I always want to have us explore is....
How much farther can we safely travel in the direction of obtaining ever more powerful tools? How much power can human beings successfully manage?
The entire field of AI development (and other emerging tech) seems built upon an assumption that more knowledge and power is always better. The dangerous irony is that while those pushing such fields forward are obviously cutting edge engineers, they don't really think like engineers. That is, they seem unable or unwilling to take all factors involved in to account.
As example, while we might happily give our ten year old son a bicycle, we're probably not going to give him a racing motorcycle. Using only common sense, we take all factors in to account, view the situation holistically, and understand that one part of the equation, the son, is not suitable for a racing motorcycle partnership.
We take this kind of reasoning to be an obvious given in regards to children, but as soon as the children become adults we throw it all out the window and blindly assume that more power is always better.
Cutting edge engineers are not bad people. They're just bad engineers.
Example, if I was building a new space station for the government, and I didn't bother to look for the single points of failure, I would be a bad engineer, and I would be fired.
Simple. Obvious. Common sense. And beyond our reach.