The new generation of large language models (LLMs) certainly do some nifty and useful things, but their applicability to health and life sciences still seems as immature, and yet as deceptively stunning, as ELIZA was 60 years ago. Back then, even the late, great Carl "extraordinary claims require extraordinary evidence" Sagan thought the ELIZA program meant the coming end of human psychotherapy as we knew it (Weizenbaum, 1976). We humans are easier to fool than we like to think - and maybe that is a good thing for the goals of creating a general machine intelligence.
Fast forward to today, and we are hearing much of the same hype that was heard back then. Yet a closer look shows us that these LLMs are really good at manipulating language that seems to catch us off guard and has us read deeper meaning into it, not unlike a psychic or mentalist. This is useful for a variety of applications across non-health industries and maybe even administrative tasks in health, but it doesn't cut it for evidence-based medicine when lives are on the line. Artificial intelligence (AI) has a pretty weak track record so far in health and life sciences, with less than 2% of clinical AI having been clinically validated (Fleuren et al, 2020) - a number which has likely dropped precipitously with the rapid influx of LLM approaches introduced in the past year with not enough time (and little inclination and funding) to do proper clinical validation (i.e. clinical trials). Add to this the difference between optimizing to a known end (AI applied in most industries) and optimizing from the unknown (biomedical science and its application in evidence-based medicine), and I'm not sure we'll be seeing reliable artificial general intelligence (AGI) for medicine or anything else any time soon.
Minding the gap
Over the past several years, I have delved deep into the history and philosophical background of artificial (née machine) intelligence along with cybernetics and other precursors. More than just an interest in history, I've sensed a fundamental gap between my current home in the tech world and my profession of origin as a neuroscientist. This sense of a gap has proved out as I've found in both theme and practice the original entwining of the exploration of natural and machine intelligence was somewhat abruptly bifurcated more than half a century ago.
Historically there was a richer interdisciplinary intersection of mathematics and engineering on one hand and neuro-, psycho-, cognitive sciences on the other (with a healthy dose of somewhat overlapping philosophy for each) in the late 1930s through early 1950s. Following the 1956 Dartmouth conference on AI, this interdisciplinarity began to bifurcate with the engineering lane accelerating narrowly with only a superficial grasp of human reasoning and the neuro-related camp spreading into a vast (and sometimes irreproducible) reductionism with limited time or attention to converting what was known into machine intelligence (a chasm that persists today).
As Patricia and Paul Churchland, the quintessential neurophilosopher couple, note in their joint foreword to John von Neumann's The Computer and the Brain (2nd edition, 2000),
"Curiously, however, these two kindred sciences - one focused on artificial cognitive processes and the other on natural cognitive processes - pursued their parallel concerns in substantial isolation from each other, from the 1950s right through to the present. The people who earned advanced degrees in computer science typically learned little or (more often) nothing about the biological brain...Equally, the people who earned advanced degrees in neuroscience typically learned little or nothing about computational theory, automata theory, formal logic..."
The whys of this split are complex - one crucial factor was Cold War funding in the US going largely to engineering exclusive AI (the defense and intelligence community literally got confused about the usage and meaning of the term "intelligence" at one point in the 60s) while the Soviets and others in their influence leaned into the more holistic cybernetics - including its somewhat bloody pinnacle in Chile in the early 1970s. Once humans are hardened into our political-cultural camps, our halo biases often hamper our ability to consider tools and ideas cross-culturally once they have been associated with one or another team.
But going back further it was the critical early loss of several key minds that sought to bridge these disciplines (and one intellectual divorce) that really impacted their co-evolution.
Four funerals and a divorce
Kenneth Craik (1914-1945) - His book The Nature of Explanation (1943) was a foundational work for cognitive science, and his posthumously published Theory of Human Operators in Control Systems (1947-48) represented the type of thinking that inspired the British cyberneticist Ratio Club (1949-58) - where his named was invoked. He was struck by a car while on his bicycle in the celebratory mayhem of Britain as WWII came to a close, passing away the next day on VE day.
Alan Turing (1912-1954) - Another Ratio Club member, the group in part may have inspired his conversion from his stance in the 1930s and early 40s that machine intelligence was not achievable to his later views that it was (articulated most clearly in his Computing Machinery and Intelligence published in 1950 in the journal Mind). His wide-ranging contributions to computing, and the British War effort, along with his tragic death and shameful treatment leading up to it are now well known, though his shift to focus on mathematical biology in the early 50s is sometimes overlooked.
John Von Neumann (1903-1957) - Like Turing, well known for various contributions to computing and WWII efforts, he also spent a considerable amount of effort on trying to align our knowledge of computers and brains. He was a regular at the Macy Conferences on Cybernetics and even co-hosted the initial similarly themed gathering (along with Norbert Weiner) at Princeton in January of 1945, before the name cybernetics had even been introduced for the topic. He is recorded as having cautioned the group about the oversimplification of the brain activity they were modeling as "neural networks". His last work, as he lay dying of bone cancer at Walter Reed, was an unfinished lecture on this topic for the Yale Silliman lecture series that was published posthumously as, Computer and the Brain. Like Craik's absence was noted at the first Ratio Club, von Neumann's absence at the initial AI meeting at Dartmouth in 1956 due to illness was similarly noted by the organizers.
Walter Pitts (1923-1969*) - From the early work on the 1943 paper A Logical Calculus of Ideas Immanent in Nervous Activity, he and Warren McCulloch, along with Norbert Wiener and the Research Laboratory of Electronics team at MIT, had set out to, "build a computational brain from the neuron up." This homeless runaway who as a child wrote to Bertrand Russell to critique his Principia Mathematica (and got graduate fellowship offer from him in return) was referred to as the "genius" of the Macy's cybernetics group and in 1954 was named one of Fortune's top 20 scientists under 40 along with Claude Shannon and James Watson. In 1959*, after thinking the labs' experimental findings negated his earlier work, he literally set fire to his graduate thesis and notes (no back-ups in those days; MIT even offered his lab mates a bounty if they could reconstruct any of it) and became a recluse, drinking himself to death over the next decade.
* Pitts' mental withdrawal from the field began with the Wiener/McCulloch split in 1951 and was accelerated in 1959 when his own work identified the analog/digital hybrid nature of the brain and he burned his doctoral thesis thinking it was now irrelevant.
And the split...
Norbert Wiener & Warren McCulloch (1951*) - Following their meeting in 1942 and respective foundational papers in 1943 (McCulloch's Logical Calculus... with Pitts mentioned above and Wiener's, Behavior, Purpose, and Teleology with Arturo Rosenbleuth and Julian Bigelow), these two spent most of the next decade aligning the neurophysiological and computational underpinnings of cybernetics and what Turing would coin in 1950 as "machine intelligence", the manifestations of which were seen in the not only the Ratio Club and Macy Conferences, but also the First Draft of a Report on the EDVAC and von Neumann's framing of the computer in neurological terms. The relationship and the associated team were aligning at MIT in the newly minted Research Laboratory of Electronics (informally known as the "Experimental Epistemology" unit) in late 1951. Suddenly, tragically and almost soap operatically (Conway & Siegelman, 2005 describe in detail the factors that may have influenced this) Wiener sent an angry telegram to his colleagues at MIT from Mexico that simply stated,
“Please inform [Pitts and Lettvin] that all connection between me and your projects is permanently abolished. They are your problem. Wiener.”
He never spoke to them again. And never explained why.
In the wake of these deaths and the academic divorce, the field of cybernetics and the intersecting new field of machine intelligence rapidly fractured. At the same time John McCarthy led a few of his colleagues in the newly named field of artificial intelligence with a kick-off conference in 1956 at Dartmouth. McCarthy's focus was significantly narrower than any of its precursors - aiming to avoid intersections with the neurological and even purely mathematical to focus on engineering approaches to software solutions for problem solving - artificial intelligence. This becomes evident from his proposal and invitations for the event, and especially so when you follow his thought processes relating to the event that he gave in subsequent interviews to Pamela McCorduck for her book, Machines Who Think. (McCorduck passed away in 2021 and gifted her interviews and related materials for the book to the CMU library archives, where I have had the pleasure of reviewing both the digitized and hard copy files of the not yet completely catalogued collection).
Moving forward
Whatever the reasons for the split and current chasm between the pursuit of creating artificial cognitive processes and understanding natural cognitive processes (as the Drs. Churchland framed it), it does seem like there is value and opportunity in bridging this gap. What is known about the brain and its processes, much less the intelligence it demonstrates, is quite limited. On the flip side, the rapid advance of engineering, on what is basically the same limited functional understanding of a "neural network" that McCulloch and Pitts gave us in 1943, has been phenomenal - and yet, with respect to general intelligence, it is still seemingly the equivalent of a child pedaling a tricycle faster and faster and thinking they can reach the moon.
By more fully understanding the neurophysiology of intelligence and the brain in general and applying it to an aligned effort towards machine intelligence, we may get to a far more advanced place with the application of computers to aid human problem solving (as Allen Newell and Herb Simon framed it). I'm not claiming that none of this is happening, the neuromorphic computing community is doing amazing work, but there is far less alignment than one would expect given the human and monetary resources at play. I think there are lessons to be learned in what came before that can apply to the more effective interdisciplinary alignment that is needed to reach the AGI, or the general machine intelligence, orbit.
Sean,
Great writing.
Thanks for the virtual, posthumously introduction to Pamela McCorduck. I'm looking forward to your treatment of Charles Babbage And Ada Lovelace, which I haven't gotten to ... yet.
Thrilled to see you brining in Herb Simon (as did McCorduck as well as the Nobel Prize committee). The Herbert Simon concept of Satisficing dramatically improved psychology. Although he passed away in 2016, I'm told that most classical economists would still like to kill him, since he took simplistic mathematics away from them. (Since Amos passed away, everyone is more kind toward Danny for rocking their worlds with science.)
I loved your opening bit on Eliza, the psychotherapist. Interested readers should know that most versions of the text editor Emacs still have the a reasonable version of Eliza in them, available to run and get free psychotherapy for yourself !
I recently verified that for GNU Emacs; Eliza is called 'doctor' and can be accessed with "M-x doctor", that's escape X followed by doctor which produces this:
I am the psychotherapist. Please, describe your problems. Each time
you are finished talking, type RET twice.
Thanks again Sean!