I can appreciate the beauty of a flower, and at the same time, I can see more about the flower. I can imagine the cells, the complicated actions in there, which are beautiful too. All kinds of interesting questions that science finds answers to add to the excitement, the mystery, and the awe of a flower. It only adds. It does not subtract.
—Richard Feynman—

In this InfoFire episode, I am in conversation with eScience pioneer Tony Hey on the topic of “AI for Science.” Our fireside chat commences with the profound words of Richard Feynman, who had a significant influence on Tony Hey, and centers around the two books— Feynman Lectures on Computation: Anniversary Edition and Artificial Intelligence for Science: A Deep Learning Revolution, edited/coedited by Tony Hey and published in 2023. Tony Hey intricately weaves a mosaic illustrating the evolution of the AI for Science topic, drawing upon short vignettes from his life and career, beginning with his post-doctoral days under Richard Feynman and many other prominent scientists at Caltech, editing his “Lectures on Computation” book, and more.

Tony Hey, a distinguished Fellow of the Royal Academy of Engineering, embarked on his academic journey with post-doctoral work at Caltech under the guidance of Nobel laureates Richard Feynman and Murray Gell-Mann following his Ph.D. under the guidance of PK Kabir at Oxford. He then went on to work with John Bell and Jacques Weyers at CERN in Geneva before taking up a faculty position at the University of Southampton.

With a theoretical physics background from Oxford, Dr. Hey seamlessly transitioned into computer science. He has made substantial research contributions, including numerous papers and nearly a dozen books. Beyond his academic endeavors, Tony Hey has played a pivotal role in shaping science policies and steering its trajectory. He notably served as the Director of the UK e-Science Initiative and continues his impactful work with various institutions, bridging the realms of academia, government, and corporations across the globe.

Igniting Passion for Science

Curiosity, the insatiable desire to unravel the mysteries of the universe, is the driving force behind scientific endeavors. Scientists are propelled by an innate sense of wonder and a profound intellectual thirst for understanding. Tony Hey shares his journey, recounting the individuals, events, and institutions that kindled his passion for science, offering glimpses into his multifaceted career across disciplines, continents, and sectors. This exploration into his professional trajectory also touches upon the motivating factors and challenges of propelling scientific research. He talks about the pushes and pushbacks in terms of their ethos in propelling scientific research, with some examples.

For young Tony, the journey began with British astronomer Sir Fred Hoyle’s radio talks and science fiction. As a teenager, he found excitement in astronomy and astrophysics, profoundly influenced by Hoyle, who formulated the theory of stellar nucleosynthesis. Hoyle’s radio talks on astronomy for the BBC in the 1950s, compiled in the book “The Nature of the Universe,” significantly impacted young Tony. Books like “Relativity for the Layman” by James Coleman and “The Strange Story of the Quantum” by Banesh Hoffmann sparked his interest in relativity and quantum mechanics—distinctive ways of perceiving the world that became Tony’s passion. Opting for Oxford over Cambridge for his undergraduate and postgraduate studies in Physics, he first encountered Feynman’s iconic work, “The Feynman Lectures on Physics,” during his undergraduate years. These lectures, influencing generations of physics teachers and graduate students, captured Tony’s interest and marked the beginning of his journey.

Shifting Gears: From Experimental Particle Physics to Theoretical Physics to Phenomenology

In contrast to experimental physics, which uses experimental tools to probe natural phenomena, theoretical physics employs mathematical models and abstractions of physical objects and systems to rationalize, explain, and predict natural phenomena. Phenomenology is the philosophical study of objectivity—and reality more generally— as subjectively lived and experienced. Applying theoretical physics to experimental data by making quantitative predictions based on known theories is phenomenology in physics. Phenomenology applied to particle physics forms a bridge between the mathematical models of theoretical physics (such as quantum field theories and theories of the structure of space-time) and the results of high-energy particle experiments.

Phenomenology in physics is thus a philosophical and methodological approach to understanding and describing physical phenomena. In the context of particle physics researchers study subatomic particles and the fundamental forces of nature.  Physicists develop theoretical models and descriptions of observed phenomena, make predictions, and test these models through experiments. This approach helps bridge the gap between theoretical physics and experimental data. Another colleague at Caltech, Geoffrey Fox, was also focused on particle physics phenomenology, especially emphasizing the modeling of strong interactions and their application to both “soft” and “hard” (high transverse momentum) processes. His overarching vision was to apply theoretical frameworks to empirical data.

Delving in more detail on his journey, Tony Hey acknowledges the great scientists who shaped his thinking and work, starting in experimental particle physics, transitioning to theoretical physics, and becoming a phenomenologist. At Oxford, he felt that some of the experimental scientists who lectured about particle physics were intimidated by the theoretical framework of quantum field theory. So, he switched to theoretical physics for his doctoral research with P.K. Kabir at Oxford and later wrote a best-selling graduate textbook on “Gauge Theories in Particle Physics” with Oxford theoretician Ian Aitchison as an introduction for experimental physicists. But moving to Caltech was more exciting as it changed how he looked at particle physics theory. Tony believes he got a whole new scientific research education at Caltech, which was very exciting. People at Caltech were only interested in solving the really significant problems of the day. He cites an interaction with George Zweig, who invented the quark model independently at the same time as Murray  Gell-Mann, as illustrative of this Caltech philosophy.

Tony Hey succinctly expresses his lifelong approach: “Phenomenology is what I have been doing all my life—understanding the theory but checking it out against experiment. I realized that I needed to do some detailed calculations to see how well – or otherwise – the theory fits the measurements from the experiments.”

Tony Hey posits the belief that both Richard Feynman and Hans Bethe were fundamentally phenomenologists in their scientific approach. The English philosopher Ben Trubody contends that Feynman, despite being frequently characterized as a physicist with a notable aversion to philosophy, should be acknowledged as an unacknowledged philosopher of science, and particularly a phenomenologist. Feynman was famously skeptical of the philosophy of science, asserting that the core purpose of scientific inquiry is to rediscover the truth directly from experience rather than relying on handed-down experiences from the past. Quoting Feynman from his book ‘The Pleasure of Finding Things Out’ (p. 185), Ben Trubody highlights its alignment with phenomenology’s foundational principle of commencing inquiries from personal experience.

…to find out ab initio, again from experience, what the situation is, rather than trusting the experience of the past in the form in which it was passed down.

Another great influence was Carver Mead, the Gordon Moore professor of computing at Caltech. Tony says he was the man who understood why Moore’s Law worked, and in a lecture in 1981, he said, “There were no engineering obstacles to things getting smaller, faster, and cheaper for the next 20 or 30 years.” This convinced Hey to believe that we need to go about doing things in different ways in computing.

In conclusion, Tony Hey says that Feynman, Gell Mann, George Zweig, Geoffrey Fox, and Carver Mead at Caltech all influenced his thinking and career trajectory. From his time at CERN, John Bell, who was first to devise a test that has ruled out Einstein’s preferred ‘local hidden variable theory’ of quantum mechanics, influenced Tony’s interest in the measurement problem of quantum mechanics and in quantum computing.

Computational Solutions in Quantum Chromodynamics: A Journey with Transputers

Tony Hey embarked on a transformative journey in quantum chromodynamics, where necessity begets invention and inspiration fuels accomplishment. Inspired by Carver Mead, he ventured into the realm of computational solutions for quantum chromodynamics by pioneering parallel processing with small computers. Confronted with limited access to supercomputers, Tony Hey found himself at the University of Southampton, a fortuitous move that led him to the world of transputers—remarkable chips with the potential for groundbreaking parallel computing applications.

The transputer, a brainchild of Iann Barron and manufactured by INMOS—a company funded by the UK Government—captured Tony Hey’s attention. These microprocessors, named from “transistor” and “computer,” served as both standalone computers and integral components of larger parallel systems when interconnected. Collaborating closely with INMOS, Tony and his colleagues at Southampton played a pivotal role in developing and harnessing the power of the transputer for scientific applications—an area that the computer science community had neglected.

In the early 1980s, facing the performance limit of conventional central processing units (CPUs), the path forward was clear: embrace parallelism. Transputers, equipped with a floating-point processor, a central processor, memory, and communication hardware, emerged as a pioneering chip, years ahead of its time. This technological leap allowed the construction of one of the first parallel-processing distributed memory computers, marking a monumental shift that eventually surpassed the dominance of expensive supercomputers.

Reflecting on this paradigm shift, Tony Hey humorously alluded to Seymour Cray’s famous analogy: “If you were plowing a field, which would you rather use? Two strong oxen or 1024 chickens?” The answer for Tony Hey and his team was clear—ultimately, the chickens (transputers) emerged victorious in transforming the landscape of parallel processing for scientific applications.

Feynman, Computation, and AI

The enormous contribution of Richard Feynman to modern physics is well known.  His book Feynman Lectures on Physics became a classic, along with the Feynman diagram approach to quantum field theory and his path integral formulation of quantum mechanics. However, his long-standing interest in the physics of computation is less well-known. Feynman lectured on computation at Caltech for most of the last decade of his life, first with John Hopfield and Carver Mead and then with Gerry Sussman.

Richard Feynman was always interested in computers and calculations. At Los Alamos, he led the computer team (which was not using real computers but IBM adding machines) to do the final calculations for the “Fat Man” plutonium atomic bomb. Feynman worked in the theoretical division headed by Hans Bethe at the Los Alamos Laboratory (Project Y of the Manhattan Project) and was involved in calculating figures such as the energy released from the bomb’s detonation. Along with Bethe, he produced a formula to determine the energy yield of a nuclear fission bomb known as the Bethe-Feynman Formula.

Feynman delivered his influential computation course at Caltech from 1984 to 1986. Although led by Feynman, the course also featured, as occasional guest speakers, some of the most brilliant men in science at that time. Eric Mjolsness, a graduate student, was one of Feynman’s teaching assistants for the computing course.  Eric worked on a neural network model for computing, storing, and retrieving data under his supervisor, John Hopfield of the Hopfield Network fame. So, there was certainly interest in neural networks at Caltech when Feynman put together material on the computing course. There were visiting scholars like Marvin Minsky from MIT, who explained his views on artificial intelligence, and Charles Bennett from IBM Research who gave lectures on reversible computing. Gerald Sussman, a professor at MIT with Minsky, a traditional, symbolic AI person, who had come to Caltech for a sabbatical, worked with Feynman on developing the lecture course.

Feynman asked Tony Hey to adapt his lecture notes into a book, which was eventually published in 1996. This timeless book, titled “Feynman Lectures on Computation,” provides a Feynmanesque exploration of both standard and some not-so-standard topics in computer science, including the controlled-NOT gate (CNOT) and quantum computers (Hey, 1999). The CNOT gate, often called the Feynman gate (in recognition of Feynman’s early notation for quantum gate diagrams in 1986), holds particular significance in quantum computing. These gates are a fundamental component for entangling and disentangling Bell states, akin to classical logic gates in conventional digital circuits. Tony Hey recounts how these lectures came to be written up as the Feynman Lectures on Computation.

Recently, in 2023, the 25^th anniversary edition of the classic was edited by Tony Hey and published by CRC Press. Tony Hey has updated the lectures with an invited chapter from Professor John Preskill on “Quantum Computing 40 Years Later”. Another update for this edition is a chapter by Eric Mjolsness (Feynman’s teaching assistant then) and now a professor in the Departments of Computer Science and Mathematics at the University of California, Irvine, on “Feynman on Artificial Intelligence and Machine Learning” which captures the early discussions with Feynman and looks toward future developments.

Though Feynman never lectured on AI and neural networks, he was interested in them. He was skeptical of the prevailing traditional symbolic artificial intelligence that was then the preferred approach of most of the computer science community. As Mjolsness states in this chapter, Feynman’s role in sparking the development of quantum computing and his affinity for calculating machines is well known. Less well known is that he was seriously interested in artificial intelligence and neural networks in their 1980s incarnation as an approach to artificial intelligence. He concludes that many of Richard Feynman’s ideas about the future of neural network-like machine learning systems have come to pass. As Tony says, whether you call it intuition or luck, Feynman’s preference for neural networks has been validated and such networks are revolutionizing AI.

Uncertainty Quantification and AI for Science

Tony Hey has coedited the book Artificial Intelligence for Science: A Deep Learning Revolution.  This book, published by World Scientific in April 2023 (Choudhary et al., 2023), is a unique collection that introduces the AI, Machine Learning (ML), and deep neural network technologies that are leading to scientific discoveries from the datasets generated both by supercomputer simulations and by modern experimental facilities. The book sets the scene with articles on Data-Driven Science and diverse application domains from astronomy, energy, health, and others. It also includes a section on The Ecosystem of AI for Science.

The power of AI rests with its ability to analyze and interpret humongous amounts of data—quickly and accurately. Responding to a question on the challenges of accuracy of data and data modeling, Tony Hey talks about the field of uncertainty quantification and refers to an article in the book “Uncertainty Quantification in AI for Science” by Tanmoy Bhattacharya, Cristina Garcia Cardona, and Jamaludin Mohd-Yusof.

Uncertainty quantification (UQ) is the scientific discipline dedicated to quantitatively characterizing and estimating uncertainties in both computational and real-world applications. Its primary objective is to ascertain the likelihood of specific outcomes when certain aspects of a system are not precisely known. By doing so, UQ aims to improve the reliability and robustness of predictions and decisions in the presence of inherent uncertainties. The applications of uncertainty quantification span various fields, including engineering, finance, environmental science, climate modeling, physics, and numerous others.

According to Tony Hey, uncertainty quantification (UQ) is necessary to address the accuracy problem for applications of Deep Learning neural networks. While acknowledging that the problem hasn’t been resolved to everyone’s satisfaction, he emphasizes the ongoing need for extensive research. Expanding on the topic of UQ, Tony introduces the critical distinction between correlation and causality within neural network models. He highlights that these models often reveal correlations rather than establishing causations. Medical AI therefore needs to develop a principled and formal uncertainty quantification (UQ) discipline, just as we have seen in fields such as nuclear stockpile stewardship and risk management (Begoli et al., 2019).

Scientific Machine Learning Benchmarks

Machine learning methods, particularly deep neural networks (DNNs), have been extensively employed in various scientific domains to analyze large experimental datasets. Their transformative impact is increasingly recognized across scientific communities (Hey et al., 2020). Nevertheless, selecting the most suitable machine learning algorithm for a specific scientific dataset poses a challenge, given the diverse range of machine learning frameworks, computer architectures, and models available. Addressing these challenges has historically involved benchmarking computer applications, algorithms, and architectures (Thiyagalingam et al., 2022).

Tony Hey, along with Thiyagalingam, Shankar and Fox, authored a paper titled “Benchmarking for AI for Science” in the book by Choudhary et al. (2023). Expanding on the insights from this paper, Tony Hey emphasizes the significance of curated, large-scale scientific datasets encompassing both experimental and simulated data. These datasets are crucial for developing meaningful ML benchmarks tailored for scientific applications. Tony Hey specifically highlights the SciMLBench approach, a project undertaken by Scientific Machine Learning Group members at the Rutherford Appleton Laboratory. Collaborating with researchers from Oak Ridge National Laboratory and the University of Virginia, the team has contributed to an international consortium called ML Commons, overseen by industry experts.

SciMLBench stands out as an open set of benchmarks, available for download and deployment by anyone. Noteworthy is the team’s commitment to incorporating significantly large datasets, exceeding a terabyte of data. These datasets are expansive and designed to be open, accompanied by metadata, and adhere to the FAIR principles (Findable, Accessible, Interoperable, Reusable). This commitment ensures that the scientific community has access to high-quality, standardized benchmarks that facilitate robust evaluation and comparison of ML algorithms across diverse scientific applications. It encourages people to explore which is the best algorithm to get the most science out of their data sets.

AI for Health

The book “Artificial Intelligence for Science: A Deep Learning Revolution” delves into various domain applications of AI. Our conversation focused on AI for Health, with particular attention to two papers: “AI and Pathology: Steering Treatment and Predicting Outcomes” by Rajarsi Gupta, Jakub Kaczmarzyk, Soma Kobayashi, Tahsin Kurc, and Joel Saltz, and “The Role of Artificial Intelligence in Epidemiological Modeling” by Aniruddha Adiga, Srinivasan Venkatramanan, Jiangzhuo Chen, Przemyslaw Porebski, Amanda Wilson, Henning Mortveit, Bryan Lewis, Justin Crow, Madhav V Marathe, and the NSSAC-BII team.

AI and Pathology

In the wake of AI’s recent success in computer vision applications, there is a growing expectation that AI will play a pivotal role in digital pathology. Computational pathology (CPATH), an evolving branch of pathology, involves leveraging computational analysis on patient specimens to study diseases. This includes extracting valuable information from digitized pathology images, often in tandem with associated metadata, utilizing AI methodologies such as deep learning. The advent of deep learning has notably facilitated image-based diagnosis, leading researchers to believe that AI, especially deep learning, holds the potential to enhance various repetitive tasks and significantly improve diagnostic capabilities (Kim et al., 2022)

While AI’s involvement in pathology is relatively recent, it is rapidly maturing as a collaborative effort among researchers, medical professionals, industry experts, regulatory bodies, and patient advocacy groups. This collective innovation aims to introduce and integrate new AI technologies into healthcare practices.

The paper titled “AI and Pathology: Steering Treatment and Predicting Outcomes” delves into histopathology, concentrating on the quantitative characterization of disease states, predicting patient outcomes, and guiding treatment strategies. The convergence of AI and high-end computing capabilities marks a paradigm shift where comprehensive morphological and molecular tissue analyses are achievable at cellular and subcellular resolutions. AI algorithms are instrumental in exploring and discovering novel diagnostic biomarkers grounded in spatial and molecular patterns that hold prognostic significance.

Tony Hey summarizes and encapsulates this transformative moment by referencing Geoffrey Hinton’s 2016 proclamation: “People should stop training radiologists now. It’s just completely obvious that within five years, deep learning is going to do better than radiologists.” As Hey says, this was a bit of a cavalier statement. Nonetheless, with image analysis, there is great progress in using DNNs to identify tumors better in some cases than trained ‘experts’. This statement underscores the potential of AI, particularly deep learning, to surpass traditional diagnostic approaches and revolutionize the field of pathology within a relatively short timeframe. The convergence of AI, computational pathology, and advanced computing is poised to usher in a new era of precision medicine and diagnostic accuracy.

Radiology is fundamentally rooted in image analysis and is particularly suited for neural networks. That’s how these deep neural networks were found, says Tony Hey, referencing Fei Fei Li, for her establishment of the ImageNet competition which was pivotal in driving significant advancements in computer vision and deep learning research during the 2010s.

Turning data to discovery: The SEER Program

In response to my request to elaborate on the latest trends in deploying AI for steering treatment and epidemiological modeling, Tony Hey cites the example of SEER repositories and how AI technologies are turning cancer data into discoveries. The Surveillance, Epidemiology, and End Results (SEER) Program provides information on cancer statistics to reduce the cancer burden among the U.S. population. SEER is supported by the Surveillance Research Program (SRP) in NCI’s Division of Cancer Control and Population Sciences (DCCPS). It is an authoritative source for cancer statistics in the US.

The example that Tony Hey cites, and with which he is associated, is the Department of Energy (DOE) and NCI Collaboration on three pilot projects that will impact future cancer research and guide advances in scientific computing. Tony Hey outlined how NLP, machine learning tools and High Performance Computing are being deployed in these joint projects to automate data capture, and overcome challenges in precision oncology at the molecular, patient, and population levels. These tools are also supporting better understanding of the care trajectory by enhancing the SEER data to better characterize each cancer patient. Natural Language Processing (NLP) pilot studies are evaluating the new NLP tools, focusing on specific data elements (e.g., biomarkers, recurrence) to extract information automatically from pathology reports and other clinical documents.

Acknowledging the challenges of AI in healthcare, Tony Hey remains positive about the promise and possibilities of AI in improving healthcare. This improvement is envisioned by analyzing health data and examining a population’s distribution, patterns, and determinants of health and disease conditions.

Data-Driven Revolution: AlphaFold and OpenFold

One of the highly discussed advancements in AI is DeepMind’s AlphaFold, which has achieved a monumental breakthrough in solving protein structures and is widely hailed as a game-changer. In late 2020, DeepMind, a London-based artificial intelligence (AI) company now under Google’s parent company, Alphabet Inc., announced the success of its AlphaFold 2 program. This program demonstrated a remarkable performance, surpassing other methods in the biennial Critical Assessment of Protein Structure Prediction (CASP) competition, marking a pivotal moment in the data-driven revolution in biology and medicine (Thornton et al., 2021)

In response to a question regarding the article “AlphaFold—The End of the Protein Folding Problem or the Start of Something Bigger?” by David T Jones and Janet M Thornton, featured in his edited book AI for Science, Tony expresses admiration for the remarkable achievements of AlphaFold. He highlights the commendable initiatives of keeping the code open and launching the AlphaFold Protein Structure Database in 2021, a collaborative effort between AlphaFold and EMBL-EBI. Tony credits his colleague at the University of Washington, David Baker, for insisting on the scientific openness of these initiatives. Baker firmly believed in transparency, advocating that research findings should be accompanied by disclosing the exact methodology. This principle greatly influenced the decision to keep the AlphaFold code open.

Quoting from a recent article in the Communications of the ACM titled “AlphaFold Spreads Through Protein Science,” Tony Hey observes that AlphaFold has gained widespread adoption. Researchers are actively working on refining and expanding AlphaFold’s capabilities by incorporating additional elements to enhance predictions for proteins that were previously challenging. The community is also embracing OpenFold, an open-source version of AlphaFold that provides biologists with valuable tools that they can adapt for their research.

OpenFold is being used to retrain AlphaFold2, providing fresh insights into its learning mechanisms and generalization capacity (Ahdritz et al., 2022). Of particular interest is integrating large language models into this process and developing protein language models (PLMs). Although individual PLMs may not match the accuracy of AlphaFold alone, their combination shows promise for the future. Computer scientists are collaborating with biologists to merge traditional AlphaFold approaches with deep learning NLP methods, incorporating Google’s BERT.

The dynamic landscape in protein science is marked by such ongoing developments, as highlighted in the article by David Jones and Janet Thornton

The authors, David Jones and Janet Thornton, have dedicated their careers to studying protein folding. David Jones played a crucial role in the original AlphaFold project, and Janet Thornton serves as the director of the European Bioinformatics Institute, a subdivision of EMBL located near Cambridge. Together, they provide an insightful perspective on the current state of protein folding research, identify areas that still require attention, and explore potential applications.

Their paper emphasizes the impact of open-sourced AlphaFold, highlighting how this has spurred continuous improvement and advancement in the field. The community’s collaborative efforts and contributions are pivotal in refining and extending AlphaFold’s capabilities. This openness and collaboration not only contribute to the ongoing evolution of AlphaFold itself but also hold promise for a multitude of applications in the broader scientific community.

Managing Big Data in Science: Semantic Web and Schema.org

In AI for Science, establishing a robust ecosystem and developing corresponding infrastructure are as pivotal as the advancements in artificial intelligence itself. Jim Gray, the proponent of the fourth paradigm of data-intensive scientific discovery, advocated for significant funding dedicated to data infrastructure, highlighting the crucial role of metadata and systems that facilitate data interoperability.

According to Tony Hey, Gray’s conceptualization of the fourth paradigm of science was driven by a need to support experimentalists grappling with overwhelming volumes of data, creating what is now known as scientific data infrastructure (called cyberinfrastructure in the US and eScience in the UK). While the origins of science trace back to systematic observations and experimentation, the subsequent paradigms introduced were the second paradigm of theory (led by figures like Maxwell, Newton, and Schrodinger) and a third paradigm of computation (coined by physics Nobel Prize Winner Ken Wilson), wherein computational simulations and modeling became integral to scientific inquiry. However, the third paradigm falls short in addressing the challenges of “data management,” leading to the emergence of the fourth paradigm, which centers around the sheer scale of data. Initially, eScience leveraged traditional machine learning techniques for data management, but contemporary approaches now employ deep learning methodologies.

The vision of a semantic web, first articulated by Tim Berners-Lee in 1999, envisioned a landscape where ontologies and linked data would impart semantic depth to the web. Schema.org, a mechanism designed to introduce a minimal amount of semantic information to websites, thereby fostering interoperability, has emerged as a significant player in this realizing this vision.

The AI for Science book incorporates a noteworthy paper titled “Schema.org for Scientific Data” by Alasdair Gray, Leyla Castro, Nick Juty, and Carole Goble.

In response to an inquiry about this article, Tony Hey underscored the success of schema.org, highlighting a rare consensus between major entities like Microsoft and Google on its adoption. The openness of its community model further contributes to its widespread success, he added. The paper describes the Bioschemas project which has extended the vocabulary supported by the original Schema.org standard to include terms relevant to biological data.  The Bioschemas project endeavors to enhance the discoverability of life sciences resources on the web, including datasets, software, and training materials. This objective is pursued by encouraging stakeholders in life sciences to incorporate the extended Schema.org markup for biology in their websites and databases, rendering them indexable and more findable by search engines and other relevant services.

AI and the Future

Tony Hey envisions a future where AI is a collaborative partner rather than a dominating force. He draws inspiration from JCR Licklider’s “Man-Computer Symbiosis” concept, emphasizing a harmonious relationship in which AI accelerates human capabilities and facilitates meaningful dialogue.

While acknowledging that this perspective might be overly optimistic, Tony contrasts it with the less likely scenario of fully independent intelligent systems akin to R2D2 in Star Wars. He anticipates the presence of simple robotic systems, particularly in care homes, but dismisses the notion of machines capable of generating independent thoughts and actions.

Tony is optimistic that, akin to how societies have addressed challenges like nuclear threats and designer babies, collaborative efforts will lead to solutions for managing the potential risks posed by AI. He notes that although issues like the atomic bomb and CRISPR technologies remain unsolved, the perceived severity of these threats is not as pressing at the moment.

Richard Feynman’s assertion that “The greatest value of science is the freedom it provides us to doubt” prompts a crucial question about whether AI possesses a similar capability for skepticism. In response, Tony Hey candidly admits that the current understanding is uncertain. The possibility that AI may lack this inherent ability is considered both a positive and potentially concerning aspect.

Hey acknowledges the complexity of injecting doubt into AI systems, a facet yet to be fully understood. The uncertainty surrounding AI’s capacity for doubt is deemed noteworthy. Hey, refrains from providing a definitive answer but promises to revisit the question in a year or so, recognizing the ongoing quest to unravel and enhance the nuanced cognitive aspects of artificial intelligence.