Sony's AI Research Chief Wants to Build Robots That Could Win a Nobel Prize
In recent years, artificial intelligence and machine learning systems have proven to be a boon for scientific research across various academic fields. They have helped scientists identify genomic markers ready for advanced treatments, accelerate the discovery of effective new drugs and therapies, and publish their own research.
However, throughout this period, such systems have often been limited to processing large datasets and performing brute-force calculations rather than directing research. Still, Dr. Hiroaki Kitano, CEO of Sony AI, plans to create a “hybrid form of science that will take systems biology and other sciences to a new level,” by building artificial intelligence as capable as today's leading scientific minds.
To do this, Kitano aims to develop an AI smart enough to win a Nobel Prize by 2050.
“What sets this task apart is the inclusion of the system in an open domain for exploring important discoveries, rather than rediscovering what we already know or trying to imitate supposed human thought processes. The vision is to reformulate scientific discovery itself and create an alternative form of it. The value lies in developing machines that can make discoveries continuously and autonomously. An AI scientist will generate and test as many hypotheses as possible, expecting that some of them may lead to major discoveries on their own or become the basis for major discoveries. The ability to exhaustively generate hypotheses and test them efficiently is the core of the system”
Today's AI itself is the result of decades of scientific research and experimentation, dating back to 1950, when Alan Turing published his foundational work “Computing Machinery and Intelligence.” Over the years, these systems have evolved from laboratory curiosities into vital tools for data processing and analysis, but Kitano wants to take one more step forward by effectively creating “a constellation of software and hardware modules that dynamically interact to perform tasks.”
Hiroaki Kitano told Engadget: “Initially, this will be a set of useful tools that automate part of the research process, both experiments and data analysis. For example, automating the lab as a closed-loop system, rather than isolated automation, is one of the first steps. A great example of this is the robot scientist Adam-Eve, developed by Professor Ross King, which automatically generates hypotheses about the genetics of baker's yeast, plans experiments to confirm or refute them, and carries out the experiments. Gradually, the level of autonomy can increase to generate a wider range of hypotheses and tests. Nevertheless, it will remain a tool or assistant for human scientists, at least for the foreseeable future.”
According to him, if an AI scientist takes on the heavy intellectual work of generating hypotheses for research, his human colleagues will have more free time to focus on research strategies and decide which hypotheses to actually pursue.
As always, preventing the black box effect and implicit bias (both in software design and in the datasets it is trained on) will be of primary importance for establishing and maintaining trust in the system. Kitano stated: “For scientific discoveries to be accepted by the scientific community, they must be accompanied by convincing evidence and reasoning. AI scientists will have components that can explain the mechanisms behind their discoveries. AI scientists that lack such explanatory abilities will be less preferred.”
Some of the greatest scientific discoveries in history, from radiation and microwaves to Teflon and the pacemaker, came about through experimental errors. But when superintelligent AIs begin inventing their own incomprehensible language, researchers are “rushing to pull the plug.” What happens if and when an AI scientist makes a discovery or designs an experiment that humans cannot immediately understand, even with an explanation?
“When AI scientists become experienced enough to handle complex phenomena, there is a chance they will uncover things that human scientists do not immediately grasp. Theoretically, someone could direct highly autonomous AI scientists without restrictions and without caring whether their discovery is understandable. However, this may come at a high cost and must be justified. When it is recognized that such an AI scientist is already making important scientific discoveries, I am confident that a graphical interface will appear to ensure safety and prevent misuse”
The emergence of an AI scientist capable of working alongside human researchers may also raise some thorny questions about who should be credited with the discoveries made: the AI that formulated the hypothesis and conducted the experiment, the person who oversaw these efforts, or the academic institution or legal entity that owns the operation? Kitano points to a recent decision by an Australian court, which recognized DABUS (“Device for the Autonomous Bootstrapping of Unified Sentience”) as the inventor of patent applications, as one example.
Conversely, the scientist recalls the case of Satoshi Nakamoto and his invention of blockchain and bitcoin: “There was a case where a decisive contribution was simply published as a blog post and taken seriously, but no one ever met him, and his identity remains a complete mystery.”
“If an AI scientist's developer were determined to create a virtual scientist persona with an ORCID iD to showcase technological achievements, promote a product, or for some other motive, it would be nearly impossible to distinguish the AI from a human scientist”
But what if this problem is a truly revolutionary achievement in medicine, say a cancer cure or a nanorobot surgeon? Does it matter whether a human or a machine conducted the experiment?
