fort, if not general agreement, with the remarks I made. I did get a few “I beg to differ” responses though, all centering on recent advancements relating to optimizing functions, which are key to the successful training of neural networks (such as results on stochastic gradient descent, dropouts, and new activation functions). The objections stemmed from not having named them as breakthroughs (in AI). My answer: They all fall under the enabler I outlined earlier: “increasingly sophisticated statistical and optimization techniques for fitting functions.” Follow up question: Does it matter that they are statistical and optimization techniques, as opposed to classical AI techniques? Answer: It does not matter as far as acknowledging and appreciating scientific inquiry and progress, but it does matter as far as explaining what just happened and, more important, forecasting what may happen next.

Consider an educated individual sitting next to you, the AI researcher, on a plane; I get that a lot. They figure out you do AI research and ask: What are the developments that enabled the current progress in AI? You recount the function-based story and lay out the three enablers. They will likely be impressed and also intellectually satisfied. However, if the answer is, “We just discovered a new theory of the mind,” you will likely not be surprised if they also end up worrying about a Skynet coming soon to mess up our lives. Public perceptions about AI progress and its future are very important. The current misperceptions and associated fears are being nurtured by the absence of scientific, precise, and bold perspectives on what just happened, leaving much to the imagination.

This is not to suggest that only a
new theory of the mind or an advance
of such scale would justify some of the
legitimate concerns surrounding AI. In
fact, even limited AI technologies can
lead to autonomous systems that may
pose all kinds of risks. However, these
concerns are not new to our industrial-
ized society; recall safety concerns when
the autopilot was introduced into the
aerospace industry and job-loss con-
cerns when ATMs were introduced into
the banking industry. The headline here
should therefore be “automation” more
than “AI,” as the latter is just a tech-
nology that happened to improve and
speed up automation.h To address these
concerns, the focus should be shifted
toward policy and regulatory consider-
ations for dealing with the new level of
automation our society is embarking
on, instead of fearing AI.

On Objectives and Success

Let me now address the third reason for
the current turn of events, which relates
to the change in objectives and how we
measure success as a broad AI com-
munity. This reason is quite substantial
yet goes largely unnoticed, especially by
younger researchers. I am referring here
to the gradual but sustained shift over AI
history from trying to develop technolo-
gies that were meant to be intelligent and
part of integrated AI systems to develop-
ing technologies that perform well and
are integrated with consumer products;
this distinction can be likened to what
has been called “Strong AI” vs. “Weak AI.”
This shift was paralleled by a sharp-
ening of performance metrics and by
progress against these metrics, partic-
ularly by deep learning, leading to an
increased deployment of AI systems.
However, these metrics and corre-
sponding progress did not necessarily
align with improving intelligence, or
furthering our understanding of intelli-
gence as sought by early AI researchers.i
One must thus be careful not to draw
certain conclusions based on current
progress, which would be justified only
if one were to make progress against
earlier objectives. This caution particu-
larly refers to current perceptions that
we may have made considerable prog-
ress toward achieving “full AI.”
Consider machine translation, which
received significant attention in the early
days of AI. The represent-and-reason ap-
proach aimed to comprehend text before
translating it and is considered to have
failed on this task, with function-based ap-
proaches being the state of the art today.
In the early days of AI, success was mea-
sured by how far a system’s accuracy was
h See also the first report of the One Hundred
Year Study on Artificial Intelligence (AI100) for
a complementary perspective; https://ai100.
stanford.edu/
AI there are metrics for evaluating task per-
formance but not for evaluating the fit among
an agent, its goals, and its environment. Such
global metrics may be needed to assess and
improve the intelligence of AI systems.

putational speed, and estimation techniques. This includes recognizing and localizing objects in some classes of images and certain tasks that pertain to natural language and speech. The third development, which goes largely unnoticed, is that we gradually changed our objectives and measures for success in ways that reduced the technical challenges considerably, at least as entertained by early AI researchers, while maintaining our ability to capitalize on the obtained results commercially, a point I discuss further later in the section on objectives and success.

Interestingly, none of these developments amounts to a major technical breakthrough in AI per se (such as the establishment of probability as a foundation of commonsense reasoning in the late 1980s and the introduction of neural networks more than 50 years ago).f Yet the combination of these factors created a milestone in AI history, as it had a profound impact on real-world applications and the successful deployment of various AI techniques that have been in the works for a very long time, particularly neural networks.g

‘I Beg to Differ’

I shared these remarks in various contexts during the course of preparing this article. The audiences ranged from AI and computer science to law and public-policy researchers with an interest in AI. What I found striking is the great interest in this discussion and the com- f Research on neural networks has gone through many turns since their early traces in the 1940s. Nils Nilsson of Stanford University told me he does not think the pessimistic predictions of the 1969 book Perceptrons: An Introduction to Computational Geometry by Marvin Minsky and Seymour Papert was the real reason for the decline in neural network research back then, as is widely believed. Instead, it was the inability to train multiple layers of weights that Nilsson also wrestled with at SRI during that time “but couldn’t get anywhere,” as he explained to me.

g A perspective relayed to me by an anonymous reviewer is that science advances because instruments improve and that recent developments in neural networks could be viewed as improvements to our machine learning instruments. The analogy given here was to genomics and the development of high-throughput sequencing, which was not the result of a scientific breakthrough but rather of intense engineering efforts, yet such efforts have indeed revealed a vast amount about the human genome.