Preamble

This article was originally written for the June 2024 issue of the Bulletin of
European Association for Theoretical Computer Science, in the framework of the
“Logic in Computer Science” column administered by Yuri Gurevich. In the fol-
lowing pages, the article is reproduced as is.
arXiv:2409.14465v1 [cs.AI] 22 Sep 2024
Foreword by the columnist
The ongoing AI revolution raises many foundational problems. For quite a while, I
felt that the issue needs to be addressed in this column. Not being an AI expert, I
was looking for volunteers. This didn’t work, and so one day I took a deep breath
and started to write an article myself. Andreas Blass, my long-time collaborator,
was reluctant to join me, but eventually he agreed.
A hundred years ago, logic was almost synonymous with foundational studies.
I tried to rekindle that tradition in [5]. The goal of the following dialog is to
provoke young logicians with a taste for foundations to notice the foundational
problems raised by the ongoing AI revolution.

A whimsical picture of a neural net with a devil’s head by the OpenAI tool Dall-E

2
 On logic and generative AI
Yuri Gurevich and Andreas Blass
University of Michigan in Ann Arbor, MI, USA

I think the most beautiful thing about

deep learning is that it actually works.
—Ilya Sutskever [13, 29:46]

§1 Thinking fast and slow1

Q: I just learned that Daniel Kahneman, Nobel laureate in economics and the
author of “Thinking, fast and slow” [7], passed away on March 27, 2024. I heard
a lot about this book but have never read it. What did he mean by thinking fast
and thinking slow?

A: Daniel Kahneman and Amos Tversky discovered that human thinking is

driven by two distinct systems, System 1 and System 2.
System 1 supports fast thinking. It “operates automatically and quickly, with
little or no effort and no sense of voluntary control . . . The capabilities of
System 1 include innate skills that we share with other animals” [7, pp. 41-43].
System 1 is good at detecting patterns and reading situations on the fly. It
allows us to make snap judgments and decisions without deliberation, but it is
prone to biases and errors.
System 2 supports slow thinking which is deliberate, analytical, and conscious.
“System 2 allocates attention to the effortful mental activities that demand it,
including complex computations. The operations of System 2 are often associated
with the subjective experience of agency, choice, and concentration” [7, p 41].
System 2 can override the impulses and biases generated by System 1.

Q: I probably rely on System 1 more than I should.

A: System 2 is much slower than System 1 and requires more effort. “While
walking comfortably with a friend, ask him to compute 23 × 78 in his head, and
to do so immediately. He will almost certainly stop in his tracks ” [7, p 74]. It is
1
The freewheeling conversations [1, 4, 10] are broken into “chapters” for the reader’s conve-
nience. Following this example, our freewheeling dialog was broken into parts at the eleventh
hour.
Below Q is Quisani, a former student of the first author, and A is the authors.

3
not surprising that often we tend to be lazy and rely on System 1 more than we
should.

Q: Have either of you met Kahneman?

A: Andreas: No, I haven’t.

Yuri: Neither have I, even though, in the 1970s, when I was teaching at Ben
Gurion University, I had a good chance to see Kahneman and Tversky at the
Hebrew University of Jerusalem where they had their seminar in psychology. I
was at the Hebrew University on Wednesdays. My Jerusalem colleagues spoke
enthusiastically about the psychology seminar, and I intended to drop in some
day, but never did. Those Wednesdays were too busy for me: model theory
seminar, set theory seminar, and work with the renowned logician Saharon
Shelah.

§2 Is generative AI intelligent?

Q: Generative AI is getting a lot of press recently. Is it intelligent?

A: The AI revolution is impressive. Generative AI is used in many domains. In

particular, it is used to improve existing logic tools. Generative AI is changing
our lives. But is it already intelligent? The issue is controversial and super
interesting; it richly deserves a separate conversation.

Q: May we address it anyway, at least briefly?

A: Sure. So the thesis under discussion is that the current generative AI is

intelligent. Let’s first hear skeptics.
Edward Gibson is a psycholinguistics professor at MIT and the head of the MIT
Language Lab. In his recent interview with Lex Fridman, Gibson argues that the
current LLMs are all about the form, not meaning.

“I would argue they’re doing the form. They’re doing the form, they’re
doing it really, really well. And are they doing the meaning? No, probably
not. There’s lots of these examples from various groups showing that they
can be tricked in all kinds of ways. They really don’t understand the
meaning of what’s going on. And so there’s a lot of examples . . . which
show they don’t really understand what’s going on” [4, 01:34:33].

Q: Presumably, GPT-4 knows all grammatical rules of English in all existing

grammar books.

4
A: The rules of natural languages may have statistical character and may not be
written anywhere.

Q: Give me an example.

A: A typical American pronounces the one vowel in “Blass” as in “glass” and

stresses the first syllable in “Gurevich.” In fact, “Blass” is a German name, and
German doesn’t have the English “glass” vowel. Similarly, “Gurevich” is a
Russian name, and Russians stress the second syllable.

Q: This is about pronunciation. Give me a text-based example.

A: One has a cold, the flu, or influenza.

Q: Wow, I am glad that I don’t have to teach grammar.

A: Also, in America, one is “in the hospital” while, in England, one is “in
hospital.” In America, the government “proposes” a law, while, in England, the
government “propose” a law.

Q: You mentioned “skeptics.” Whom else did you have in mind?

A: Yann LeCun, along with Yoshua Bengio and Geoffrey Hinton, received the
2018 Turing Award for their work on deep learning. They are sometimes referred
to as the “Godfathers of Deep Learning.” LeCun says that LLMs, large language
models like GPT-4, aren’t truly intelligent yet (and cannot take us to
superhuman intelligence) because they essentially lack important
capabilities [10, 00:02:47]:

1. understanding the physical world,

2. persistent memory,

3. reasoning, and

4. planning.

“That is not to say,” adds LeCun, “that autoregressive LLMs are not useful,
they’re certainly useful; or that they’re not interesting; or that we can’t build a
whole ecosystem of applications around them, of course we can. But as a path
towards human-level intelligence, they’re missing essential components.”
At this point, we heard two skeptics. What do you think?

5
Q: Gibson’s argument did not convince me too much. I am thinking about clever
extraterrestrials analysing the wet chemistry of our nervous system and
wondering do humans really understand the meaning of what’s going on? There
are lots of ways humans can be tricked into irrational behaviour.
LeCun doesn’t look to me like a true skeptic. He points out various ways to
improve the current generative AI.
Also, there is a matter of definitions, in particular the definition of intelligence.
Maybe we should be speaking about degrees of intelligence.

A: OK, let’s turn our attention to the defence of the thesis. In a recent talk,
Geoffrey Hinton made a strong case for the intelligence of the current large
language models, LLMs.

“They [LLMs] turn words into features, they make these features interact,
and from those feature interactions they predict the features of the next
word. And what I want to claim is that these millions of features and
billions of interactions between features that they learn, are understanding
...
This is the best model we have of how we understand. So it’s not like
there’s this weird way of understanding that these AI systems are doing
and then this [is] how the brain does it. The best that we have, of how the
brain does it, is by assigning features to words and having features [and]
interactions” [6, 0:14–15].

Q: But these allegedly intelligent LLMs hallucinate.

A: So do we. “Anybody who’s studied memory, going back to Bartlett in the

1930s, knows that people are actually just like these large language models. They
just invent stuff and for us, there’s no hard line between a true memory and a
false memory” [6, 0:16].

Q: Interesting. I intend to listen to that talk of Hinton.

A: Another champion of the thesis is Ilya Sutskever, Chief Scientist at Open

AI [13, 14].
A brief discussion cannot do justice to an issue as complicated as the thesis.
Let’s move on.

Q: Wait, there is a closely related issue: What about the future? Many AI
experts say that generative AI is getting more intelligent than humans and may
become a danger to us.

6
A: The chances are that its intelligence will be incomparable to ours. Think of
airplanes versus birds. Airplanes fly faster, but birds can land and take off
almost anywhere. On the other hand, AI develops fast and we don’t know the
directions. “It’s tough to make predictions, especially about the future,” said the
baseball philosopher Yogi Berra.
In any case, you are right. Numerous AI experts worry that generative AI may
become a danger to humans. They point out that it may be used by bad actors
for manipulating electorates and waging wars.

“There are always misguided or ill-intentioned people,” says Bengio, “so it

seems highly probable that at least one organisation or person would —
intentionally or not — misuse a powerful tool once it became widely
available” [2].

In the talk [6], Hinton describes various scenarios he is worried about.

“But the threat I’m really worried about,” says Hinton, “is the long term
existential threat. That is the threat that these things could wipe out
humanity . . .
[W]hat happens if superintelligences compete with each other? . . . The one
that can grab the most resources will become the smartest. As soon as
they get any sense of self-preservation, then you’ll get evolution occurring.
The ones with more sense of self-preservation will win and the more
aggressive ones will win” [6, 0:21–24].

Q: This is frightening indeed.

§3 Reasoning and logic in AI

Q: I am fascinated by LeCun’s analysis above. Let’s look at it more carefully.

Persistent memory sounds technical, not as profound as the three other
capabilities.

A: Persistent memory would help with long-term coherence. If you heavily

interact with an LLM for a long time, it would be highly desirable that the LLM
maintains coherence and consistency. This is a challenge for the current systems.

Q: This is a challenge for some humans as well ,

A: Also, in humans, persistent memory is vital for incubation of ideas,

cross-pollination of ideas from different domains, and reasoning by analogy.

7
Q: Concerning LeCun’s item 3, what kind of reasoning is AI incapable of?

A: System 2 reasoning. An LLM is basically just two files. One is a huge data
file that reflects the information the model was trained on. The other is an
algorithm, typically succinct, which usually embodies the model architecture as
well as a probabilistic inference mechanism. That mechanism works with a
sequence of words. It starts with a given query, and runs in rounds. During one
round, it infers another word and appends it to the current sequence. (More
exactly, it works with subword tokens which are not necessarily full words.)
Sometimes the next word is obvious, but sometimes it is very hard to figure out
an appropriate next word. But the LLM cannot stop and think. In the AI
parlance, “it is allocating approximately the same amount of compute for each
token it generates” [1, 00:59:51]. In that sense it uses only fast thinking
(System 1). As far as we know, it is an open problem how to improve the process
by incorporating slow, deliberate thinking (System 2).

Q: Understanding of the real world seems to involve reasoning as well, both

System 1 reasoning and System 2 reasoning, from evading predators (fast
thinking) to quantum physics (slow thinking). Thus reasoning is ubiquitous in
AI, and AI needs many different kinds of reasoning. It would be natural if the
role of logic in AI would grow and grow. And yet, Yuri wrote that “the golden
age of logic in artificial intelligence is behind us” [5, §4]. How come?

A: At the early stage of AI, the logic approach was dominating. A lot of good
work was done; see the article “Logic-based artificial intelligence” in the Stanford
Encyclopedia of Philosophy [15]. But then an important competitor arose. Let
us quote LeCun on the issue2 .

“In the 1950s, while the heralds of classical AI, based on logic and
tree-based exploration, were pushing back its limits, the pioneers of
learning started to make their voices heard. They defended the idea that,
if we want to make computer systems capable, like animals and humans, of
complex tasks, logic is not enough. We must get closer to the functioning
2
The French original: “Dans les années 1950, tandis que les héraults de l’intelligence artifi-
cielle classique, basée sur la logique et l’exploration arborescente, en repoussent les limites, les
pionniers de l’apprentissage commencent à donner de la voix. Ils défendent l’idée que, si l’on
veut rendre les systèmes informatiques capables de tâches complexes, à la manière de l’animal et
de l’homme, la logique ne suffit pas. Il faut se rapprocher du fonctionnement du cerveau, et donc
rendre les systèmes capables de se programmer eux-mêmes, en s’inspirant de ses mécanismes
d’apprentissage. Cette partie de la recherche fondée sur l’apprentissage profond (deep learning)
et les réseaux de neurones (artificiels) est celle à laquelle je me consacre. Elle est à l’œuvre dans
toutes les applications spectaculaires actuelles, à commencer par la voiture autonome” [9, p. 23].

8
 of the brain, and therefore make the systems capable of self-programming,
drawing inspiration from their learning mechanisms. This part of research
based on deep learning and (artificial) neural networks is the one to which
I am devoted here. It is at work in all of today’s spectacular applications,
starting with the autonomous car” [9, p. 23].

§4 Real-world understanding

Q: Coming back to real-world understanding, robots need to navigate around

obstacles and handle objects without causing damage to those objects, to
themselves, and of course to people. This requires understanding of the physical
world. But, unlike robots, LLMs don’t walk around and they have read much of
what there is to read. Shouldn’t all that knowledge give them a near perfect
understanding of the physical world?

A: Not really.

Q: Can you illustrate that?

A: Sure. Henri Lebesgue had an idea that we can build upon: there are cases
where arithmetic does not apply. “Suppose I put two animals in a cage and then
two more. How many animals will the cage then contain? [8, p. 14]” Let’s present
a version of this scenario to Google’s Gemini.

— Suppose that I put 2 animals of one kind and 3 animals of another kind into a
closed space. How many animals will be there altogether?

— Adding the number of animals of each kind together, you will have 2 + 3 = 5
animals altogether.

Q: This is ridiculous. Gemini doesn’t realize that the result depends on the
kinds of animals. For example, the two animals of the first kind may consume
some or all of the three animals of the other kind. Maybe, Gemini takes into
account only the moment that the you close the door of the closed space.

A: OK, let’s check your hypothesis.

— What if I check the total number of animals not immediately but in one hour, one
day, or one month?

9
— The total number of animals (5) wouldn’t change regardless of whether you check
on them in one hour, one day, or even one month. As long as no animals enter or
leave the closed space, the initial sum of 2 and 3 will remain the same.

Q: One month is not a trivial period. Even if none of our original animals is a
predator, some of the animals may die or have young.

A: Gemini doesn’t have your real-world experience3 .

Q: Have you seen imperfections with other LLMs?

A: In §3, we mentioned an article on logic-based AI in the Stanford Encyclopedia

of Philosophy (SEP). Before turning to SEP, we asked GPT-4 to suggest a good
survey article on logic-based AI. GPT-4 hallucinated a combination of an existing
title and well-known authors who could write an article with this title but didn’t.

Q: This harks back to the issue that an LLM can’t stop and think.

A: In this particular case, the necessary thinking is algorithmic: just check the
known sources. So this hallucination seems to be a bug.

§5 Moravec’s paradox

Q: It is ironic that Gemini can do impressive intellectual work but misses

obvious things about the physical world.

A: You rediscovered Moravec’s paradox from the 1980s. Here’s an instructive

quote from Hans Moravec’s 1988 book [11, pp. 15–16]:

“It seemed to me that, in the early 1970s, some of the creators of

successful reasoning programs suspected that the poor performance in the
robotics work somehow reflected the intellectual abilities of those attracted
to that side of the research. Such intellectual snobbery is not unheard of,
for instance between theorists and experimentalists in physics. But as the
number of demonstrations has mounted, it has become clear that it is
comparatively easy to make computers exhibit adult-level performance in
3
The chat with Gemini occurred on April 29th. Long before that, we presented to Gemini
another scenario (a coin tossed down from a tall building) that also exposed Gemini’s lack of
real-world knowledge. In early May, we again presented both scenarios to Gemini and discovered
that Gemini improved in the meantime. Despite the improvements, Gemini still showed, albeit
in a less dramatic way, a lack of real-world knowledge.

10
 solving problems on intelligence tests or playing checkers, and difficult or
impossible to give them the skills of a one-year-old when it comes to
perception and mobility.
In hindsight, this dichotomy is not surprising. Since the first multicelled
animals appeared about a billion years ago, survival in the fierce
competition over such limited resources as space, food, or mates has often
been awarded to the animal that could most quickly produce a correct
action from inconclusive perceptions. Encoded in the large, highly evolved
sensory and motor portions of the human brain is a billion years of
experience about the nature of the world and how to survive in it. The
deliberate process we call reasoning is, I believe, the thinnest veneer of
human thought, effective only because it is supported by this much older
and much more powerful, though usually unconscious, sensorimotor
knowledge. We are all prodigious olympians in perceptual and motor
areas, so good that we make the difficult look easy. Abstract thought,
though, is a new trick, perhaps less than 100 thousand years old. We have
not yet mastered it. It is not all that intrinsically difficult; it just seems so
when we do it.”

Q: I don’t buy that slow thinking is not all that intrinsically difficult. He also
seems to suggest that — given a chance and time — evolution will make slow
thinking more efficient. Taking into account how wayward evolution is, the time
in question may be humongous. But I digress.

A: In the Lex Fridman podcast, LeCun further develops Moravec’s argument.

“Those LLMs are trained on enormous amounts of texts, basically, the

entirety of all publicly available texts on the internet, right? That’s
typically on the order of 1013 tokens. Each token is typically two bytes, so
that’s 2 · 1013 bytes as training data. It would take you or me 170,000
years to just read through this at eight hours a day. So it seems like an
enormous amount of knowledge that those systems can accumulate, but
then you realize it’s really not that much data. If you talk to
developmental psychologists, they tell you [that] a four-year-old has been
awake for 16,000 hours in his or her life, and the amount of information
that has reached the visual cortex of that child in four years is about 1015
bytes [10, 00:04:08] . . .
What that tells you is that through sensory input we see a lot more
information than we do through language and that, despite our intuition,
most of what we learn and most of our knowledge is through our

11
 observation and interaction with the real world, not through
language” [10, 00:05:12].

§6 Thinking fast

Q: All this makes the fast-thinking grasp of the real world even more interesting
and relevant. Clearly, LLMs would need such a grasp.

A: They need it already. Have you recently spoken to an LLM-powered chatbot?

Q: I try to avoid them and talk to humans.

A: This will be harder and harder to do.

Q: It is already plenty hard and sometimes virtually impossible; try to talk to

OpenAI support.
But, OK, let’s suppose that I am chatting with a bot representative of some
company.

A: Imagine that they sent you something, say a laptop, and it arrived damaged,
that you waited for that laptop longer than promised, and that you need one
right now. A human would quickly realize that you are upset and would try to
be extra empathetic. So should the bot.

Q: Yes, and this requires reading the situation on the fly, fast thinking.
Now that I have some idea about fast thinking, I understand Yuri’s question
what the logic of fast thinking is in the February 2021 column. But is there is a
logic of fast thinking?

A: Much depends on what we mean by logic. George Boole discovered an

algebra of universal “laws of thought” [3], that is, universal laws of slow thinking.
Is there a useful algebra of universal laws of fast thinking? We doubt it. Fast
thinking is rather far from exact reasoning.

Q: Are there any meaningful rules of fast thinking?

A: Oh, yes. For example, we intuitively judge the frequency of an event by how
easily examples come to mind, which can be influenced by recent exposure and
emotions [7, p. 244]. Such rules of thumb help us, given inconclusive perceptions,
to produce quickly an action that is often beneficial but may be detrimental.
Arguably logic should study the laws of thinking, including fast thinking,
whatever they are. The intention, in the question about the logic of fast

12
thinking, was to provoke young logicians with a taste for foundations to notice
the nascent study of fast thinking.

Q: Studying fast thinking may expand logic tremendously. Do you know

historical examples where logic expanded so greatly?

§7 The dawn and heyday of logic

A: One example is related to the emergence of logic as a separate discipline in

the classical Greek period.

Q: Was there any logic before that?

A: Various forms of structured argumentation were used before the classical

Greek period, certainly in China and India. But it was in Greece that logic
became a separate discipline.

Q: Does this have something to do with Athenian democracy?

A: The practice of Greek democracy wasn’t limited to Athens, though Athens is

the best-documented case. The Athenian democracy was characterized by the
participation of its (free male) citizens in the Assembly and courts. There were
many occasions where one had to convince numerous citizens rather than talking
to gods or to a few aristocrats. The necessity of convincing argumentation drove
the development of logic as well as of rhetoric and demagoguery. The
contributions of Aristotle and his school were particularly significant in the birth
of logic as a discipline. Aristotle’s system of logic was virtually unchallenged
until the 19th century.

Q: Mathematics greatly advanced in the period from Aristotle to the 19th

century. Mathematical arguments went beyond Aristotelian syllogisms. They
must have used much more sophisticated logic, and logic seems to precede
mathematics, logically speaking ,

A: Yes, already in the time of Aristotle mathematics involved logic beyond

Aristotelian syllogisms. But for a long time mathematicians used deduction
principles implicitly. After all, those principles are essentially trivialities. One
learned them in the course of learning mathematics.

Q: Still, there were many developments in mathematics, between Euclid’s time

and 19th century.

13
A: The most dramatic development was the invention of calculus. Its use of
infinitesimals and related notions, like infinite sums, certainly raised new logical
issues. Reasonable looking computations could contradict each other.

Q: Give me an example.

A: Consider the infinite series 1 − 1 + 1 − 1 + 1 − 1 + . . . One way to sum it is

(1 − 1) + (1 − 1) + (1 − 1) + . . . = 0. Another way is
1 + (−1 + 1) + (−1 + 1) + . . . = 1. Which way, if any, is the correct one? The
foundations of mathematics became problematic.

Q: Today we know that the series does not converge.

A: It took a while for mathematicians to work out notions like convergence. The
problem was actual infinity. Mathematicians used to work with finite objects and
potential infinity, never actual infinity. E.g., for Euclid, the only lines were finite
line segments.

Q: In my experience, people mention infinitesimals but don’t use them.

A: Eventually, the language of infinitesimals was replaced by the precise

language of epsilon and delta, and thus the foundations of analysis got rid of the
fuzziness induced by infinitesimals.
The introduction of set theory by Georg Cantor improved the situation further.
Cantor bravely dealt with actual infinity and made great progress.
Mathematicians quickly grew comfortable with set theory. In particular, it was
observed that ordered pairs can be coded as sets, which allowed modern
set-theoretic definitions of notions like function and relation purely in terms of
set membership.
But at the same time paradoxes of set theory were discovered, and a real crisis
arose in the foundations of mathematics. It wasn’t clear anymore which
convincing reasoning is valid and which is not.

Q: And, I guess, all this provoked the birth of mathematical logic.

A: Even before the discovery of paradoxes, important foundational work was

done, notably by George Boole, Giuseppe Peano, and Gottlob Frege. But the
foundational crisis put mathematical logic in the center of attention. The crisis
was resolved, for all practical purposes, by the development of axiomatic set
theory.

Q: Did this make logic popular?

14
A: It surely did. Andrei Kolmogorov and John von Neumann started as
logicians. Alan Turing wrote his PhD thesis in logic [16]. The universal Turing
machine suggested to von Neumann the idea of the universal electronic computer
which made logic even more popular.
In 1956, when the famous Dartmouth workshop [17] started AI as a field, logic
was at the zenith of its popularity. This is a partial explanation for the
dominance of the logic approach at the early stage of AI. By the way, the term
“artificial intelligence” was coined by John McCarthy, Marvin Minsky, Nathaniel
Rochester and Claude E. Shannon in their proposal for the Dartmouth
conference.

§8 Logic in the age of AI

Q: In both cases, the birth of logic and the resolution of the foundational crisis
in mathematics, there was a need, and logic rose to the occasion. Demand
preceded supply, so to speak. Do you think that, this time around, generative AI
provides sufficient demand?

A: It does. The demand is ubiquitous.

Q: Give me an example.

A: A quick glance at a person gives you a surprising amount of information. We

may learn the gender, estimate the age, etc. Such rapid assessment was
instrumental to our survival as a species.

Q: It seems that, as in object-oriented programming, a person P is viewed as an

object with various attributes.

A: But the set of attributes may depend on the assessor; for example, a runner
may see whether P is a runner. Also, the set of values that an attribute takes
may be dynamic; for example, a child may learn only whether P is a child or an
adult — in contrast to an adult’s more precise estimate of P ’s age.

Q: Surely, psychologists and neuroscientists are already studying rapid

assessment. What can logic add?

A: Rapid assessment is a case of fast thinking. In [5], Yuri asked what the logic
of fast thinking is. If and when laws of fast thinking are understood, at least to
some extent, they may help us to analyze rapid assessment.

Q: Do you think that logic will rise to meet the AI challenge?

15
A: This is an interesting question. To us, logic is the study of reasoning, any
kind of reasoning [5]. There is no doubt that the kinds of reasoning which are
relevant to AI will be studied.

Q: But will those studies be called logic? Maybe, AI logic?

A: At the moment this does not look probable. It is not predetermined what is
or is not logic. Conflicting social processes are in play. Much depends on how —
and whether — the logic community will address the challenge.
The story of infinitesimal calculus may be instructive. See a sketch of that story
in [5, §2]. After the invention of infinitesimal calculus, a question arose how to
reason with infinitesimals? For a couple of centuries, mathematicians struggled
with the problem and eventually largely solved it using the epsilon-delta
approach that essentially eliminated infinitesimals. In a sense, the mathematical
heroes of that story worked as logicians but this wasn’t recognized. This
particular story has a happy, albeit somewhat bitter, end of sorts as far as
logicians are concerned. Abraham Robinson came up with a proper logic of
infinitesimals, which he called nonstandard analysis and which allows you to
handle infinitesimals consistently [12]. But, in the meantime, mathematicians got
accustomed to the epsilon-delta approach and weren’t too much interested in
nonstandard analysis which is arguably more elegant and efficient.

Q: You wish of course that logic will be more successful in AI.

A: True.

Acknowledgments
Many thanks to Naomi Gurevich, Ben Kuipers, Vladimir Lifschitz, and Moshe
Vardi for useful comments.

References
[1] Sam Altman, “OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power,
and AGI,”
Lex Fridman Podcast #419, 18 March 2024, transcript,
https://lexfridman.com/sam-altman-2

[2] Yoshua Bengio, “One of the ‘godfathers of AI’ airs his concerns,” The
Economist July 21st 2023

16
 [3] George Boole, “An investigation of the laws of thought,” Walton & Maberly
1854, https://gutenberg.org/ebooks/15114

[4] Edward Gibson “Human language, psycholinguistics, syntax, grammar, &

LLMs,” Lex Fridman Podcast #426, 17 April 2024, transcript,
https://lexfridman.com/edward-gibson-transcript
(At the time of writing, the transcript mistakenly assigned Gibson’s words to
Fridman and vice versa.)

[5] Yuri Gurevich, “Logical foundations: Personal perspective,” Bulletin of

EATCS February 2021, also arXiv:2103.03930 and Logic Journal of the
IGPL 33:6 Dec 2023 1192–1202

[6] Geoffrey Hinton, “Will digital intelligence replace biological intelligence,”

Romanes Lecture, Oxford, UK, 19 February 2024
https://www.youtube.com/watch?v=N1TEjTeQeg0&list=PPSV

[7] Daniel Kahneman, “Thinking fast and slow,” Farrar, Straus and Giroux 2011

[8] Henri Lebesgue, “Measure and the integral” (a translation of “La mesure
des grandeurs” and “Sur le développement de la théorie de l’intégrale”),
Holden-Day 1966

[9] Yann LeCun, “Quand la machine apprend,” avec la collaboration de

Caroline Brizard, Odile Jacob 2019. Russian translation: Ян Лекун, "Как
учится машина," ПРО Москва 2021. There is no English translation yet, it
seems.

[10] Yann LeCun, “Meta AI, Open Source, Limits of LLMs, AGI & the Future of
AI,”
Lex Fridman Podcast #416, 7 March 2024, transcript,
https://lexfridman.com/yann-lecun-3-transcript

[11] Hans Moravec, “Mind children,” Harvard University Press 1988

[12] Abraham Robinson, “Non-standard analysis,” North-Holland 1966

[13] Ilya Sutskever, “Deep learning,” Lex Fridman Podcast #94, 8 May 2020,
https://www.youtube.com/watch?v=13CZPWmke6A (At the time of writing,
only an auto-generated transcript is available.)

[14] Ilya Sutskever, “The Exciting, Perilous Journey Toward AGI,” TED Podcast
20 November 2023,
https://www.youtube.com/watch?v=SEkGLj0bwAU

17
[15] Richmond Thomason, “Logic-based artificial intelligence,” Stanford
Encyclopedia of Philosophy (Spring 2024 Edition), E.N. Zalta & U.
Nodelman (eds.),
https://plato.stanford.edu/archives/spr2024/entries/logic-ai/

[16] Alan Turing, “Systems of logic based on ordinals,”

PhD Thesis, Princeton University 1938,
Proceedings of the London Mathematcal Society, Series 2 45, 161–228, 1939

[17] Wikipedia contributors, “Dartmouth workshop,”

https://en.wikipedia.org/wiki/Dartmouth_workshop.
The version, last edited on 12 April 2024, was retrieved May 10 2024.

18