"Weak AI" is Likely to Never Become "Strong AI", So What is its Greatest Value for us?

Abstract

AI has surpassed humans across a variety of tasks such as image classification, playing games (e.g., go, "Starcraft" and poker), and protein structure prediction. However, at the same time, AI is also bearing serious controversies. Many researchers argue that little substantial progress has been made for AI in recent decades. In this paper, the author (1) explains why controversies about AI exist; (2) discriminates two paradigms of AI research, termed "weak AI" and "strong AI" (a.k.a. artificial general intelligence); (3) clarifies how to judge which paradigm a research work should be classified into; (4) discusses what is the greatest value of "weak AI" if it has no chance to develop into "strong AI".

Full text

arXiv:2103.15294v1 [cs.AI] 29 Mar 2021
1
“Weak AI” is Likely to Never Become “Strong AI”,
So What is its Greatest Value for us?
⋆Bin Liu
First posted March 30th, 2021
Abstract
AI has surpassed humans across a variety of tasks such as image classification, playing games (e.g., go,
“Starcraft” and poker), and protein structure prediction. However, at the same time, AI is also bearing serious
controversies. Many researchers argue that little substantial progress has been made for AI in recent decades. In
this paper, the author (1) explains why controversies about AI exist; (2) discriminates two paradigms of AI research,
termed “weak AI” and “strong AI” (a.k.a. artificial general intelligence); (3) clarifies how to judge which paradigm
a research work should be classified into; (4) discusses what is the greatest value of “weak AI” if it has no chance
to develop into “strong AI”.
Index Terms
Artificial intelligence, artificial general intelligence, deep learning, weak AI, strong AI
I. INT RO DUC TION
The last decade has seen impressive applications of AI represented mostly by deep neural networks,
i.e., deep learning [1]. The striking point lies in that the computing agent has reached and even surpassed
humans in many tasks, e.g., image classification [2], speech recognition [3, 4], games [5–7], protein
structure prediction [8]. Even ten years ago, it was hard to imagine that AI would achieve so many
amazing breakthroughs.
On the other side, AI is also bearing serious controversies during the same period. Among the critics,
Judea Pearl, a pioneer for probabilistic reasoning in AI and a winner of the Turing award, argues that
“... all the impressive achievements of deep learning amount to just curve fitting,” and a necessary ability
to be supplemented for AI is causal reasoning [9, 10]. Gary Marcus, a professor of cognitive science,
⋆B. Liu is with Zhejiang Lab, Hangzhou, China. e-mail: bins@ieee.org or liubin@zhejianglab.com.

2
summarizes ten limitations of deep learning [11], namely, “... it is data-hungry, ... it has limited capacity
for transfer, ... it has no natural way to deal with hierarchical structure, ... it struggles with open-ended
inference, ... it is not sufficiently transparent, ... it has not been well integrated with prior knowledge,
... it cannot inherently distinguish causation from correlation, ... it presumes a largely stable world, in
ways that may be problematic, ... it works well as an approximation, but its answers often cannot be fully
trusted, ... it is difficult to engineer with”. In a recent issue of the journal Frontiers in Psychology, another
cognitive scientist J. Mark Bishop argues that AI “is stupid and causal reasoning will not fix it” [12].
In this paper, I attempt to concisely respond to current controversies about AI. Specifically, I emphasize
discrimination between two paradigms of AI research, namely “weak AI” and “strong AI” (Section II);
provide a conceptual guide to judge which paradigm a research work should be classified into (Section
II-A), explain why controversies about AI last (Section III), present major views on whether “weak AI”
will grow into “strong AI” (Section IV) and discuss what is the greatest value of “weak AI” if it has no
chance to become “strong AI” (Section V).
II. WH AT DO “WE AK AI” AND “ST RONG AI” ME AN?
“Weak AI” and “Strong AI” are two terms coined by John Searle in the “Chinese room argument”
(CRA) [13]. CRA is a thought experiment as follows: “Searle imagines himself alone in a room following
a computer program for responding to Chinese characters slipped under the door. Searle understands
nothing of Chinese, and yet, by following the program for manipulating symbols and numerals just as
a computer does, he sends appropriate strings of Chinese characters back out under the door, and this
leads those outside to mistakenly suppose there is a Chinese speaker in the room” [14]. The term “strong
AI” entails that, “... the computer is not merely a tool in the study of the mind; rather, the appropriately
programmed computer really is a mind, in the sense that computers given the right programs can be
literally said to understand and have other cognitive states.” In contrast, the term “weak AI” implies that
“... the principal value of the computer in the study of the mind is that it gives us a very powerful tool.” J.
Mark Bishop summarizes that ‘‘weak AI focuses on epistemic issues relating to engineering a simulation
of human intelligent behavior, whereas strong AI, in seeking to engineer a computational system with all
the causal power of a mind, focuses on the ontological ” [12].
I borrow the terms “weak AI” and “strong AI” here without an intent to discuss CRA. See related
discussions in e.g., [15–18].

3
Simply put, “weak AI” represents computational systems that exhibit as if they own human intelligence,
but they do not. In contrast, “strong AI” represents computational systems that have human intelligence.
Correspondingly, all AI research can be categorized into two paradigms: one is targeted for realizing
“strong AI”; and the other produces advanced “weak AI” systems to meet a variety of practical needs.
A. How to Judge a Research Work Belongs to Which Paradigm?
The biggest motivation for realizing “strong AI” is to answer the question: what are the generation
mechanisms of humans’ intelligence and how to implement these mechanisms with a machine. Therefore,
given a research work, it is easy to judge whether it belongs to the “strong AI” paradigm. If this work
provides any new and useful clue for us to answer the above question, it falls within the “strong AI”
paradigm; otherwise, it belongs to the “weak AI” paradigm.
Based on the above method, part of the (especially early) works on neural networks that deepen our
understanding of the working mechanism of biological neural systems, surely belongs to the “strong AI”
paradigm. On the other hand, most research works that involve artificial neural networks and deep learning,
even if they are proposed under the inspiration of research on neuroscience, cognitive science, behavior
psychology, they belong to the “weak AI” paradigm as long as they do not give us any new insight on
the generation mechanisms of humans’ intelligence or on how to better implement mechanisms that have
already been found.
III. WHY CO NTROVE RSI E S ABOUT AI LAST?
In controversies about AI, party A believes that AI has made substantial progress in the past decade;
party B doubts or even negates the development of AI.
I argue that controversies arise mainly because these two parties mix two different concepts, “weak
AI” and “strong AI”, together, when they talk about AI. The fact is that “weak AI” has made substantial
progress in the past decade, while “strong AI” has not. Party A thinks that “weak AI” is an important
member of the AI family; progress gained from “weak AI” also belongs to this AI family. In contrast,
in the mind of Party B, there always exists one ideal form of AI, namely a realized “strong AI”, and
the “distance” between current AI and this ideal AI is treated as a criterion for evaluating current AI.
Compared with decades ago, current AI still lacks basic human-level abilities such as causal reasoning
[9], robust decision making [19], commonsense utilization [20], and knowledge transfer, which implies

4
that the “distance” between the realized AI and the ideal “strong AI” has not been remarkably shortened.
Therefore, it is reasonable for party B to doubt or even negate the development of AI.
A natural question arises: how breakthroughs of “weak AI” have come out in the past decade? Judea
Pearl argues that “... all the impressive achievements of deep learning amount to just curve fitting”.
However, the point is that, different from previous fitting methods, deep learning permits to do an
extraordinary fitting - fitting multi-modal big data in an end-to-end way. This deep learning type of
fitting requires a big consumption of both computing and storage resources but avoids labor-intensive
feature engineering. Big data, big computing, and big storage are three requisites that make deep learning
surpass humans in playing Go, image classification, speech recognition, and so on. The luck for deep
learning is that the past decade happens to witness great improvements in sensing technologies, wireless
mobile phones, cloud computing, computing devices, computer storage, and databases, which give birth
to big data, big computing, and big storage required by deep learning.
IV. WILL “WEAK AI” GROW INT O “STRONG AI”?
A metaphor is often used to reply to this question: the relationship between “weak AI” and “strong AI”
is like that between flying machines and birds. Flying machines are not developed by accurately mimicking
birds’ flying. Birds perform much better in maneuvering than the most advanced flying machine today.
Birds can flexibly re-purpose their behaviors while flying machines cannot. But the appearance of flying
machines has met demands of speedy transportation and others. People may think that, since it is unlikely
and not necessary for flying machines to develop into birds, then similarly, “weak AI” is unlikely and not
necessary to grow into “strong AI”.
To formally consider whether “weak AI” will grow into “strong AI”, let recall the Turing test [21] and the
CRA (mentioned in Section II). An example statement of the Turing test is as follows [22]: “Originally
known as the Imitation Game, the test evaluates if a machine’s behavior can be distinguished from a
human. In this test, there is a person known as the “interrogator” who seeks to identify a difference between
computer-generated output and human-generated ones through a series of questions. If the interrogator
cannot reliably discern the machines from human subjects, the machine passes the test. However, if
the evaluator can identify the human responses correctly, then this eliminates the machine from being
categorized as intelligent.” Through the lens of CRA, Searle argues that the Turing test has serious flaws,
as passing the test does not indicate that the machine has consciousness or understanding. The absence
of an effective evaluation method hampers the development of “strong AI”.

5
Besides, philosophers and cognitive scientists often use G¨
odel’s first incompleteness theorem [23] to
argue that a machine cannot generate humans’ consciousness or understanding. See related discussions in
e.g., [12].
V. WH AT IS T HE GREATE ST VALU E OF “WEAK AI” FO R US?
In his most recent paper, Geoffrey Hinton states that “The difference between science and philosophy is
that experiments can show that extremely plausible ideas are just wrong and extremely implausible ones,
like learning an entire complicated system by end-to-end gradient decent, are just right” [24]. In [25], Judea
Pearl argues that “Modern connectionism has in fact been viewed as a Triumph of Radical Empiricism
over its rationalistic rivals. Indeed, the ability to emulate knowledge acquisition processes on digital
machines offer enormously flexible testing grounds in which philosophical theories about the balance
between empiricism and innateness can be submitted to experimental evaluation on digital machines.”
Combining their arguments, one can see that they both attribute recent deep learning’s success as a success
of empiricism which is data-driven, other than driven by philosophical theory or intuition.
A very important lesson that can be learned from the fast-pacing development and applications of AI
in the past decade is that deep learning running on big enough data can produce unexpected shortcuts
to solve extremely difficult problems. For example, by combining deep learning, reinforcement learning
[26], and Monte Carlo tree search [27], a computer program AlphaGO [28] can win the human champion
without having to understand any of the Go-playing strategies that have been accumulated by humans
for more than four thousand years. The Generative Pre-trained Transformer 3 (GPT-3) [29] can generate
human-like texts through deep learning without having to understand any syntax or semantics underlying
the texts. It is shown that the greatest value of “weak AI” represented by deep learning lies in that it
provides scalable, less-labor-involved, accurate, and generalizable tools for distilling, representing and
then exploiting patterns hidden from big data. Although such “weak AI” has no real intelligence, to a
large extent it meets urgent needs for scalable, efficient, and accurate processing of big data.
In a foreseeable future, “weak AI” is likely to become more robustly (with e.g., portfolio [19] or
dynamic portfolio methods [30–35]), while a big challenge is how to model “unknown unknowns”; it will
perform more automatically through e.g., auto machine learning [36], but it can not become completely
automatic provided that “strong AI” is realized [37]; it may perform as if it owns abilities of cognition
and understanding, but it does not.

6
VI. CO NCL USI O NS
AI has made great progress in the past decade. It has influenced almost all facets of human society
by providing more efficient algorithmic solutions to representation, management, analysis of multi-modal
big data. Controversies about AI last mainly because “weak AI” becomes so strong while “strong AI” is
almost as weak as it was decades ago. Almost all breakthroughs of AI that have attracted the public’s
attention in the past decade are within the “weak AI” paradigm. “Weak AI” is developing much faster
than expected. Even ten years ago, one could not imagine that a computer program would beat the human
champion soon in playing Go. In contrast, the “fruits” people have got from the “strong AI” paradigm are
not so striking as from “weak AI”. I suggest, when talking about AI in the future, one should better make
a statement in advance whether this talk is about “weak AI” or “strong AI”. In this way, more focused
and constructive discussions can be expected.
In a foreseeable future, “weak AI” cannot develop into “strong AI” (see why in Section IV), but it
provides a channel to synthesize advances obtained from related disciplines such as cloud computing,
computer storage, high-speed wireless mobile communications. Through this synthesis of technologies,
more advanced algorithmic tools will be developed in the “weak AI” paradigm, then “weak AI” will
continue to influence human society more profoundly, through big data. The man-computer symbiosis
world that Licklider predicted more than sixty years ago [38] is becoming a reality.
REFER ENC E S
[1] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, 2015.
[2] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” Advances in neural
information processing systems, vol. 25, pp. 1097–1105, 2012.
[3] W. Xiong, L. Wu, F. Alleva, J. Droppo, X. Huang, and A. Stolcke, “The microsoft 2017 conversational speech recognition system,” in
IEEE int’l conf. on acoustics, speech and signal processing (ICASSP). IEEE, 2018, pp. 5934–5938.
[4] G. Saon, G. Kurata, T. Sercu et al., “English conversational telephone speech recognition by humans and machines,” arXiv preprint
arXiv:1703.02136, 2017.
[5] D. Silver, T. Hubert, J. Schrittwieser et al., “A general reinforcement learning algorithm that masters chess, shogi, and go through
self-play,” Science, vol. 362, no. 6419, pp. 1140–1144, 2018.
[6] N. Brown and T. Sandholm, “Superhuman ai for heads-up no-limit poker: Libratus beats top professionals,” Science, vol. 359, no.
6374, pp. 418–424, 2018.
[7] O. Vinyals, I. Babuschkin, W. M. Czarnecki et al., “Grandmaster level in starcraft ii using multi-agent reinforcement learning,” Nature,
vol. 575, no. 7782, pp. 350–354, 2019.
[8] A. W. Senior, R. Evans, J. Jumper et al., “Improved protein structure prediction using potentials from deep learning,” Nature, vol. 577,
no. 7792, pp. 706–710, 2020.

7
[9] J. Pearl, “The limitations of opaque learning machines,” Possible minds: twenty-five ways of looking at AI, pp. 13–19, 2019.
[10] J. Pearl and D. Mackenzie, “Ai can’t reason why,” Wall Street Journal, 2018.
[11] G. Marcus, “Deep learning: A critical appraisal,” arXiv preprint arXiv:1801.00631, 2018.
[12] J. M. Bishop, “Artificial intelligence is stupid and causal reasoning will not fix it,” Frontiers in Psychology, vol. 11, pp. 1–18, 2021.
[13] S. John, “Minds, brains, and programs,” Behavioral and Brain Sciences, vol. 3, no. 3, pp. 417–457, 1980.
[14] D. Cole, “The chinese room argument,” https://plato.stanford.edu/entries/chinese-room/.
[15] G. Rey, “What’s really going on in Searle’s “Chinese room”,” Philosophical Studies, vol. 50, no. 2, pp. 169–85, 1986.
[16] M. J. Shaffer, “A logical hole in the chinese room,” Minds and Machines, vol. 19, no. 2, pp. 229–235, 2009.
[17] A. Sloman and M. Croucher, “How to turn an information processor into an understander,” Behavioral and Brain Sciences, vol. 3,
no. 3, pp. 447–448, 1980.
[18] M. A. Boden, Computer models of mind: Computational approaches in theoretical psychology. Cambridge University Press, 1988.
[19] T. G. Dietterich, “Steps toward robust artificial intelligence,” AI Magazine, vol. 38, no. 3, pp. 3–24, 2017.
[20] G. Marcus, “The next decade in AI: four steps towards robust artificial intelligence,” arXiv preprint arXiv:2002.06177, 2020.
[21] A. M. Turing, “Computing machinery and intelligence,” in Parsing the turing test. Springer, 2009, pp. 23–65.
[22] IBM Cloud Education, “Strong AI,” https://www.ibm.com/cloud/learn/strong-ai.
[23] P. Raatikainen, “G¨
odel’s incompleteness theorems,” https://plato.stanford.edu/entries/goedel-incompleteness/.
[24] G. Hinton, “How to represent part-whole hierarchies in a neural network,” arXiv preprint arXiv:2102.12627, 2021.
[25] J. Pearl, “Radical empiricism and machine learning research,” Causal Analysis in Theory and Practice (Blog), vol. 26, 2020.
[26] R. S. Sutton and A. G. Barto, Reinforcement learning: An introduction. MIT press, 2018.
[27] S. Gelly and D. Silver, “Monte-carlo tree search and rapid action value estimation in computer go,” Artificial Intelligence, vol. 175,
no. 11, pp. 1856–1875, 2011.
[28] D. Silver, A. Huang, C. J. Maddison et al., “Mastering the game of go with deep neural networks and tree search,” Nature, vol. 529,
no. 7587, pp. 484–489, 2016.
[29] T. B. Brown, B. Mann, N. Ryder et al., “Language models are few-shot learners,” arXiv preprint arXiv:2005.14165, 2020.
[30] B. Liu, Y. Qi, and K. Chen, “Sequential online prediction in the presence of outliers and change points: an instant temporal structure
learning approach,” Neurocomputing, vol. 413, pp. 240–258, 2020.
[31] Y. Qi, B. Liu, Y. Wang, and G. Pan, “Dynamic ensemble modeling approach to nonstationary neural decoding in brain-computer
interfaces,” Advances in neural information processing systems, pp. 6087–6096, 2019.
[32] B. Liu, “Robust particle filter by dynamic averaging of multiple noise models,” in IEEE Int’l Conf. on Acoustics, Speech and Signal
Processing (ICASSP). IEEE, 2017, pp. 4034–4038.
[33] Y. Dai and B. Liu, “Robust video object tracking via bayesian model averaging-based feature fusion,” Optical Engineering, vol. 55,
no. 8, pp. 1–11, 2016.
[34] B. Liu, “Data-driven model set design for model averaged particle filter,” in IEEE Int’l Conf. on Acoustics, Speech and Signal Processing
(ICASSP). IEEE, 2020, pp. 5835–5839.
[35] ——, “Instantaneous frequency tracking under model uncertainty via dynamic model averaging and particle filtering,” IEEE Trans. on
Wireless Communications, vol. 10, no. 6, pp. 1810–1819, 2011.
[36] F. Hutter, L. Kotthoff, and J. Vanschoren, Automated machine learning: methods, systems, challenges. Springer Nature, 2019.
[37] B. Liu, “A very brief and critical discussion on automl,” arXiv preprint arXiv:1811.03822, 2018.
[38] J. Licklider, “Man-computer symbiosis,” IRE Transactions on human factors in electronics, no. 1, pp. 4–11, 1960.