What?! Is there a “strong AI” that can defeat GPT4?

The ChatGPT and GPT-4 large language models recently released by OpenAI have almost set off a powder keg in the national discussion on artificial intelligence. I believe that many people’s Moments have been flooded with messages containing “GPT”.

The friends around me are also divided into two camps. The radicals believe that the singularity towards strong artificial intelligence has arrived. After another increase in data volume and computing power in the future, it will be able to replace the vast majority of text workers, and then all non-creative jobs will face unemployment. The conservatives believe that it is just a top productivity tool, but it cannot master human innate abilities such as consciousness, emotion, and creativity, and it is difficult to become strong artificial intelligence.

I am a lazy person, so I chose to ask ChatGPT directly. Here is its answer:

Image source: Screenshot from ChatGPT

It is still unknown whether strong artificial intelligence can be achieved, but there is no doubt that the human brain is still the most intelligent and complex system on the entire planet.

Some people may ask: Why don’t we just use biological neurons as the basic unit of the network and replicate an intelligent system similar to the human brain from scratch? Would this be a shortcut to achieve strong artificial intelligence? Good question, welcome to the theme of this article - " Organoid Intelligence (OI)".

01

Carbon-based and silicon-based work side by side

I just want to ask you if you are afraid!

The term organoid intelligence is a new concept proposed by Thomas Hartung's team at Johns Hopkins University on February 28 this year.

Simply put, organoids are stem cells that are cultured in a three-dimensional environment in vitro into tissue analogs with certain structures and functions . Currently, a variety of organ tissues have been successfully constructed, including the small intestine, stomach, colon, bladder, liver, heart, pancreas, kidneys, and brain.

Perhaps in the near future, if there is a problem with a part of your body, you can use organoids to repair or replace it, such as getting a liver replacement (bushi) if your liver is damaged from working late at night.

Organoid intelligence is the use of brain organ tissues cultured in vitro as biological hardware, which is connected to external electronic devices to achieve biological computing . The brain we artificially construct is the CPU + GPU of the entire computer. Isn’t it sci-fi?

Copyright image, no permission to reprint

In fact, many studies have proven the feasibility of this path. Take the recent news, for example, Andrew Dou's team at the University of Illinois at Urbana-Champaign has cultivated more than 80,000 neurons obtained by reprogramming mouse stem cells, and placed them between optical fibers and electrode grids to receive 10 different types of electrical pulses and light signal stimulation. These components are placed in a constant temperature box to maintain the activity of the neurons.

After one hour of training, the researchers found that this group of neurons sent out the same signal every time they received the same stimulation pattern. The researchers also used the F1 score to quantify the efficiency of the neural network's pattern recognition. In simple terms, 0 is the worst and 1 is the best. In the end, the computer built by the living neural network had an F1 score of 0.98, which can be considered quite intelligent.

In addition to pattern classification tasks, the researchers also connected the "brain in a box" to a robot made of living muscle tissue, allowing the brain to sense changes in the surrounding environment through the muscles and process this information.

Combining living cells and reservoir computing technology can enable neurons and chips, or carbon-based and silicon-based intelligent units, to collaborate to complete signal recognition and processing tasks. This is the original form of the integration of carbon-based and silicon-based biological intelligence, although it currently looks rather ugly.

An organoid intelligent robot with biological neurons placed in the middle. Image source: Reference [3]

Going further back, there are also some mind-blowing studies.

For example, last December, Australian biotech startup Cortical Labs used a "Dish Brain" grown from human brain cells to learn how to play table tennis.

The research team integrated human neurons induced and differentiated from human stem cells with high-density multi-electrode arrays and computers, so that the electrical signals in the "ping pong" game were transmitted to the microelectrode array, which in turn told the neurons the location of the "ping pong ball". After the neurons reached a consensus through the exchange of electrical signals with each other, they controlled the movement of the "racket" and hit the "ping pong ball" back.

Miraculously, this group of "brains on a dish" learned the game in just 5 minutes, while an artificial neural network of similar size might take about 90 minutes to learn it.

Although the first author of the paper insists that the "brain in a dish" has already inserted itself into the racket in the game, whether organoid intelligence can "autonomously and consciously" absorb, distinguish and respond to external information like humans remains a question worth exploring.

A “brain on a dish” that can play table tennis. Image source: Reference [4]

02

Organoid intelligence technology discovered by accident

Brain organoid technology was not originally intended to achieve "organoid intelligence." Since the breakthrough in related research in 2019, most of the work has focused on studying brain development and disease, or repairing missing parts of the brain.

For example, a paper published in the main journal Nature last October showed that scientists transplanted human brain neurons into rat brains for the first time and formed connections, thereby controlling the rat's behavior.

Four months later, research published in the Cell journal showed that human brain-like organs not only achieved effective connections after being implanted into the rat brain, but also responded to visual stimuli , indicating that they had been integrated into the brain and were functioning. This result can be used as a therapeutic strategy to restore cortical function.

However, there are always some creative scientists who like to build some eye-catching technological networks, such as this organoid intelligent technology.

Of course, this technology is still a nascent baby. When it gradually matures and is compared with the current deep learning-based artificial intelligence technology, will it become a castle in the air that is always "to be expected in the future" in the field of scientific research, or will it become a trendsetter in the new wave of brain-like intelligence and head straight for strong artificial intelligence? Let us wait and see.

Copyright image, no permission to reprint

As a network composed of biological intelligence units, neurons, the advantages of organoid intelligence are concentrated in the following aspects:

1. Low energy consumption

The brain of a zebrafish larva in the water, which successfully captures prey and escapes from the eyes of its natural enemies, consumes only 1 microwatt of power. The brain of an adult human consumes only 20 watts of power . Currently, the power consumption of server clusters that rely on deep learning is usually around 1 million watts. The Frontier supercomputer with the most powerful computing power consumes nearly 21 megawatts, which does not sound very environmentally friendly.

2. Few-shot learning

Organisms are usually able to learn using fewer observations. Humans can complete a simple "object type is the same or different" task using about 10 training samples , and insects such as bees only need 100 training samples to learn, and this amount of training data often has poor classification results in deep neural networks.

The AlphaGo system received training data from 160,000 Go games. A player who trained for five hours a day would need to play Go for 175 years, rain or shine, to complete so many games. This shows that the brain can achieve high training efficiency in learning activities without too much data, and its storage capacity is amazing (about 2,500 megabytes).

3. Provide excellent brain-computer interface

The electrical signals output by the computer or electrode are transmitted to the organoid brain tissue, and the organoids implanted in the human brain can be fully integrated into the brain tissue to perform their functions. This interface has good physiological properties and can almost minimize the damage of the interface to the brain. At the same time, it can integrate machine intelligence and natural intelligence to achieve a new form of intelligence.

Of course, how to connect electrodes to small and three-dimensional organoids is also a problem that needs to be considered, which requires high-quality 3D electrodes as support. If humanoid robots are to be built in the future, it will be more convenient from a physiological point of view to use neurons connected to simulated muscle tissue.

Organoid intelligent architecture for biological computing. Image source: Reference [2]

03

Cruel or advanced, that is the question

The 21st century is not only the century of biotechnology, but also the century of information technology and the century of interdisciplinary science.

Currently, several research teams have used gene editing and optogenetics to build neural networks with specific functions, and have used nanotechnology and bioprinting to build more complex organoid frameworks. It is expected that brain tissues cultured in the future will have more sophisticated structures and more specific functions.

The original intention of "organoid intelligence" is to use the advantages of biological computing, which is faster, more efficient and low-energy, to build a living computer, so that it can achieve better performance than traditional silicon-based computers in more complex tasks. It can also send or receive instructions to computer chips through electrical pulse signals, realize coordinated computing of carbon-based and silicon-based neural networks, and integrate the relative advantages of the two to create a more intelligent computing system.

But the current challenges of organoid intelligence can be divided into two main areas: ethics and technology.

First is the ethical issue.

After 10 weeks of culture, the organoids will show characteristics of a 20-week fetus, such as myelination, and stimulation with information input will affect the development of the organoids, making them more complex in structure. Will they be conscious if they receive input, generate output, interact with the surrounding environment, and establish primitive memories? Will the electrical signal stimulation input from the outside world cause "pain" in these brains? These are issues that ethical organizations have been paying close attention to.

Finding the necessary and sufficient physiological conditions for consciousness is one of the most difficult problems in neuroscience . Work is currently underway to reveal the neural basis of consciousness, which will provide a good reference for the ethical regulations of organoids. Organoid intelligence itself is not intended to reconstruct human consciousness, but to provide a functional basis for biological learning, cognition and computing.

If the realization of strong artificial intelligence requires a large number of neurons as a basis, how is this intelligence different from our human intelligence? In an extreme case, we directly cut off the embryonic neural tube that has not yet formed autonomous consciousness (this part will form the human brain in the future), and then put it in an incubator to cultivate a network system with 86 billion nodes, and connect it to computer chips or servers to handle various tasks. Is this realization of strong artificial intelligence humane?

The author believes that even if this technology is successfully developed, it will be prohibited by laws and regulations due to quite complex ethical issues, just like the cloning technology and gene editing technology used in humans.

Strong artificial intelligence in the anime "Psycho-Pass" - Sibyl System. Image source: Anime "Psycho-Pass"

The second is technical issues.

The single brain organoid used by Thomas Hardon's team mentioned above contains about 50,000 neurons, while Andrew's team used 80,000. Although small, it has everything. The currently cultivated brain organoids can reproduce the organizational structure and function of the brain, with myelinated axons, spontaneous electrophysiological activities, complex oscillatory behaviors, high cell density and layered patterns, and even multiple cell types such as oligodendrocytes, microglia and astrocytes.

But the question is, how to make these organoids learn and use their computing power? And, what is the number of neurons required to show high intelligence? In the next stage, researchers will expand the culture scale based on the existing differentiation scheme and build a living computer with 10 million nerve cells. As for the level of intelligence shown, let us wait and see.

At the same time, how to adjust the connections between neurons so that they can better realize their functions also needs to be considered. Although brain organoids may realize the spatiotemporal characteristics of molecular characteristics, they cannot reflect the topological structure of human brain regions and the complexity and specificity of neuronal circuits, which may be the basis for realizing advanced brain functions. How to reasonably wire a large number of messy neuronal connections and use molecular signals to induce the generation of related functional circuits is also a problem that needs to be considered in the future production of specific function-oriented biological neural networks.

In addition, it is still impossible to efficiently and completely record the signal input and output of human brain organoids. Researchers are working on developing 3D brain-computer interfaces and corresponding probes specifically for brain organoids, such as 3D microelectrode arrays (MEAs), neuropixel silicon probes, etc., to achieve precise docking with external information ports. The advancement of these technologies can solve this problem to a certain extent.

In general, although current artificial intelligence is far less comprehensive and efficient than the human brain in learning ability, and can only perform well on tasks that have been pre-trained on a large scale, organoid intelligence is still difficult to reproduce the simplest brain. If we want to "beat" silicon-based biology led by the GPT series, I am afraid there is still a long way to go, especially the more biological neurons used, the greater the ethical challenges faced. At the end of the exploration of intelligence, it is very likely that carbon-based and silicon-based intelligence will come together.

Roadmap for achieving organoid intelligence. Image source: Reference [2]

Before realizing strong artificial intelligence, we might as well ask ourselves again, why do we want to realize strong artificial intelligence?

If all we need is an obedient and easy-to-use tool, then we only need to train an intelligent neural network that can perform well in specific tasks. Consciousness, emotion, and creativity are all unstable factors. They are not necessary for any low-tech and highly repetitive work, and will also lead to an increase in social management costs.

In this way, models like ChatGPT may be the most ideal intelligent system for humans. I hope that you who read this article can make good use of this tool and maximize your productivity in the new wave of artificial intelligence.

Humans and androids. Image source: game "Detroit: Become Human"

References:

[1] Smirnova L., Caffo BS, Gracias DH, et al. Organoid intelligence (OI): the new frontier in biocomputing and intelligence-in-a-dish. Front Sci 1:1017235. 2023. doi: 10.3389/fsci.2023.1017235

[2] Morales PIE, Smirnova L., Muotri AR, et al. First Organoid Intelligence (OI) workshop to form an OI community. Front. Artif. Intell. 6:1116870. 2023. doi: 10.3389/frai.2023.1116870

[3] Andrew D. 80,000 mouse brain cells used to build a living computer. NewScientist Physics. 2023. From: https://www.newscientist.com/article/2363095-80000-mouse-brain-cells-used-to-build-a-living-computer/

[4] Brett JK, Andy CK, Nhi TT et al. In vitro neurons learn and exhibit sentience when embodied in a simulated game-world. Neuron 110, 2022. doi: https://doi.org/10.1016/j.neuron.2022.09.001

[5] Paola Arlotta et al. Individual brain organoids reproducibly form cell diversity of the human cerebral cortex, Nature. 2019. doi: 10.1038/s41586-019-1289-x

[6] Pașca, SP, Arlotta, P., Bateup, HS et al. A nomenclature consensus for nervous system organoids and assembloids. Nature 609, 907–910, 2022. doi: https://doi.org/10.1038/s41586-022-05219-6

[7] Dennis Jgamadze rt al. Structural and functional integration of human forebrain organoids with the injured adult rat visual system. Cell Stem Cell. 2023. doi: 10.1016/j.stem.2023.01.004.

Produced by: Science Popularization China

Author: Qian Yu (Center for Excellence in Brain Science and Intelligence Technology, Chinese Academy of Sciences)

Producer: China Science Expo