Amy's List

“Why shouldn’t a book teaching mathematics mimic the creative, inventive process involved in doing mathematics?”

- Amy Langville, Deconstruct this Calculus 3 Journal

For the past year, Dr. Amy Langville and I have been collaborating on a book focused on explaining the mathematical concepts that make possible the current generation of large language models.

Amy is a math professor at College of Charleston. I first met her because I read and admired her book, “Google Pagerank and Beyond: the Science of Search Engine Rankings” (Princeton University Press 2006). Based on how easy she made it to understand the deep concepts behind using link-analysis for search and retrieval, concepts on which I claimed priority to some extent, I asked if she would take a look at my patents and meet with me to discuss them.

Amy understood my patents. And in person, she turned out to the most down-to-earth and unpretentious math professor I’ve ever met. Open-hearted and caring, the kind of person who seems incapable of deceit. Given my experience as a law clerk sitting in on trials in federal district court, I knew a jury would feel exactly the same way about her that I did. On my recommendation, the patent owners recruited her as expert witness in litigation against Google.

(We went on to win that battle decisively, on the condition settlement terms remain secret).

But back to Amy. Amy is not like any other math professor I know. She was a star basketball player in high school and college, recruited by more than 50 colleges, named Women’s Basketball Player of the Year in her conference. Now, when not at school, she is surfing with her husband at some remote location famous for its waves, like Tofino or somewhere in Costa Rica off the beaten path.

And not surprisingly, Amy is also a mold-breaking and generative math teacher. After her Pagerank book, and another with Princeton on other types of ranking systems, she has spent the past ten years developing a series of six light-hearted but carefully designed textbooks called Deconstruct Calculus. These books embody what I aspire to do when teaching math. I recommend them whether you are a teacher or learner of Calculus. They are a breath of fresh air.

Each book includes explanations and practice problems of course, but also challenge problems, a graphic novella, “math morals” - demonstrations of the broader applicability of mathematical ways of thinking - and activities, including some that involve ripping pages out of the book (she means the “deconstruct” part literally). And throughout, the emphasis is on concepts over calculation, and on active participation: drawing, making things, kinesthetic understanding, visual learning.

Check them out at: https://www.deconstructcalc.com.

We are very closely aligned.

So naturally, when I realized I wanted to write a book explaining the math concepts leading up to the current generation of AI, while providing philosophical context from the history of thought about the nature of computation, I immediately thought of Amy as the ideal writing partner.

I broached the idea, and to my delight she was game for developing a manuscript. We wrote the first one-hundred fifty pages of straight description and explanation quite quickly and easily last Summer and Fall. But this Spring, as we contemplated approaching publishers, we ran into the practical obstacle of articulating what was original in our approach.

As of 2026, AI has become so hyped that there is an avalanche of very good books and articles and courses and Substack posts and online resources already on seemingly every imaginable aspect.

Of course, being late to the feast would apply just as much to writing about Calculus, and Amy found a way to be original there. And the same was true for me with the subject of Data Science in 2015, before I launched the Coursera courses on math for data analysis that became my defining statement as a teacher, attracting over a million students in 180 countries.

So, what do we have to offer together that is unique? I feel that we do have something, but what exactly? I have had many different ideas.

Every time I think I have come up with a new and compelling elevator pitch, Amy deflates my exuberance by telling me about a book she has read that approached the topic from a similar point of view. She seems to have read everything and fully assimilated it, while I move much more slowly through the alien worlds of other people’s ideas.

As we reflect on the year just past, I thought it would be worthwhile to review what we read and discussed together, and sum up how the work struck me.

And I thought I’d share Amy’s List with you.

Amy’s List

(in alphabetical order by first author)

1. Brian Christian, The Alignment Problem, Machine Learning and Human Values (Norton 2020)

The idea that some day in the distant future computers might become powerful enough to be dangerous has been around in one form or another for almost 80 years. It is a staple of science fiction dystopias. But in recent years this hypothetical risk is widely seen to be much less hypothetical, and more nearly actual.

In 1951, Alan Turing wrote that “it seems probable that once the machine thinking method has started, it would not take long to outstrip our feeble powers.” And once computers exceed human abilities, human beings would no longer be their master. “At some stage therefore we should have to expect the machines to take control.”

In 1965, Turing’s Bletchley Park colleague, I.J. Good, wrote a paper in which he articulated a mechanism by which such an “intelligence explosion” might happen. “Let an ultra intelligent machine be defined as a machine that can far surpass all the intellectual activities of any man however clever. Since the design of machines is one of these intellectual activities, an ultra intelligent machine could design even better machines; there would then unquestionably be an ‘intelligence explosion,’ and the intelligence of man would be left far behind.” Good saw the potential danger in this. “Thus the first ultra intelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control.”

In 1993 the science fiction author Vernor Vinge rebranded the moment superintelligence emerges as the technological “Singularity,” although the term is now more widely associated with Ray Kurzweil, who published The Singularity is Near in 2005.

In the late 1990s, Eliezer Yudkowsky rebranded the process by which computers will achieve the Singularity as “recursive self-improvement,” and this name has stuck.

In 2014, Nick Bostrom, in Superintelligence, emphasized urgency - the idea that computers would soon have capabilities that “exceed the cognitive performance of humans in virtually all domains of interest.” And he agreed with Good, Vinge, and Yudkowsky that the most likely way for this to happen would be recursive self-improvement.

Also in 2014, Stuart Russell rebranded the danger more technocratically as a “the value alignment problem.”

Yet, after seventy years of speculation, Brian Christian’s 2020 book was so far as I know the first full-scale attempt to enumerate the practical policy implications of this strain of thought for non-technical readers. His book is key reference point for almost all subsequent discussion of the subject.

Christian focuses on three main aspects of the alignment problem:

First, machine learning systems train on historical examples and therefore learn the biases present in the data. For example, if you train a system to determine whether someone should be given a mortgage, using data that reflects historical bias in lending against a particular ethnic group, the computer will likely continue to implement that bias.

Second. when trained on a reward function such as reinforcement learning, computers may engage in “reward hacking” - optimizing their performance in ways inadvertently permitted by their formal instructions but not consistent with the goals actually intended by their human programmers.

Third, computers might in future be trained to infer what human beings value by observing human behavior, rather than by what we say we value. This raises significant dangers if the modelled behavior falls short of being values-driven: “watch what we do, not what we say.”

This book played a large role in making what had previously been a specialized, or even fringe, discussion, primarily among a small group of technically trained computer scientists, into a set of well-defined topics that can be debated by lawyers, legislators, regulators, public policy organizations and the general public. Its timing was also excellent, coming as it did less than three years before ChatGPT was released.

2. Mark Coeckelbergh and David J. Gunkel, Communicative AI (Polity Press 2025)

This book and another on the list below - Language Machines by Leif Weatherby - are both built around the observation that the way large language models internally represent the relationships among words are reminiscent of ideas published posthumously in 1916 by Ferdinant de Saussure about how meaning arises in language. (Saussure’s ideas are well known in academic circles in the humanities because they were taken up and promoted by influential French Structuralist and Post-structuralist literary critics). Between the two books, I found Language Machines to be more imaginative and thoughtful, and include more comprehensive citations to the relevant sources.

But rather than critique these two books in detail here, given how big and complicated this topic is, I will later write a separate article about the relationship between LLM architecture and the ideas of various literary critics, as discussed in these two books and. a half-dozen other academic publications that have explored the LLM/Structuralism parallels.

But one issue is worth mentioning right away; the success of LLMs undermines the core assertions of Noam Chomsky’s Universal Grammar. See my comments on Language Machines below.

3. Karen Hao, Empire of AI, Dreams and Nightmares in Sam Altman’s Open AI (Penguin 2025)

This book is so well-written it is annoying. The reporting is vivid and dramatic. The author is an MIT alumna with a STEM degree, and yet she writes like a poet. There are some minor errors (she exaggerates the water consumption of data centers, for example) but overall this is the best book to read to get the lay of the land of the high-stakes maneuvering that went on before and after the November 2023 launch of ChatGPT, and the subsequent world-historical hype levels around Large Language Models and AGI.

4. Kai-Fu Lee, AI Superpowers: China, Silicon Valley, and the New World Order (HarperCollins 2021)

Many perceptive insights and shrewd predictions validated in the four years since publication. Perhaps the most useful insight is that technological advances proceed on two very different fronts - individual conceptual leaps, and practical advances by what Lee calls “tinkerers.” The tinkerers are the ones who make a new concept useful and commercially viable. Culturally, the US tends to romanticize and valorize the individual who conceives of a whole new approach, while in China the focus is much more on rapid iteration of practical applications. This seems quite true to me.

Written from a technology venture investor’s perspective, with an excited and optimistic tone, which may strike some readers as a bit too “gung ho.”

[”Gung ho” - interesting phrase. I wonder where that comes from? The way I can learn new things on the Internet continues to delight me. Ok, here’s what I just learned: The term originates in China’s World War II resistance to the Japanese invasion. It originally referred to Workers’ Industrial Cooperatives, as a shortening of 工業合作社 (Gōngyè Hézuòshè). It was adopted by the US Marines through its use as a motto by the 2nd Marine Raider Battalion, organized and led by Lieutenant Colonel Evans F. Carlson. Carlson was a military observer in China in the 1930s, and admired the sprit de corps he saw in the Chinese guerilla movement. He adopted the expression for his unit, and it became widely known in the US though the 1943 film Gung Ho! that dramatized Carlson’s new Battalion and its heroic raid on Makin Island.

The movie is here:

But through the influence of the film, “gung ho” underwent a semantic shift. It lost Carlson’s intended meaning of a cooperative, egalitarian spirit, where men could freely question officers, for example, and came to be understood in the popular imagination as something closer to enthusiasm to the point of recklessness. And so, to the extent Carlson’s original idea threatened traditional hierarchy in the military, it was neutralized and trivialized.]

5. Kai-Fu Lee and Chen Quifan, AI 2041 (Crown Currency 2021)

I have not yet read this book. Sorry, Amy.

6. Dr. Fei-Fei Li, The Worlds I See (Flatiron Books 2023)

Dr. Li is a hero of mine, for her willingness to build the first giant training set of images at a time when the academic status quo was near-universal contempt for neural networks and their apparent need for orders of magnitude more examples to train on than were then available in digital form.

Once enough training data existed (and Ilya Sutskever and others realized in 2012 that they could use NVIDA graphics cards to train much larger neural networks for image classification efficiently) neural networks began crushing the more “sophisticated” and elegant/efficient symbolic AI approaches in open competitions on Dr. Li’s dataset. This was the “bitter lesson” - scale beats algorithmic elegance - and the breakthrough restored neural networks to academic respectability, which in turn led to the current LLM era.

Of course there is much more to her life and career than that - please do read the book.

This book’s timing did feel like a bit of a calculated PR move, as she prepared to launch her own startup, World Labs - see what she did there with the name?

But more power to her. Dr. Li raised a $230 million “stealth” round in 2024, and has since raised another $1 billion. I love what she is doing; she understands exactly the fundamental limitations of the LLMs architecture, and sees where the field must go if it is to advance. If she offered me a job, I would accept it in a heartbeat. Please, Dr. Li, DM me...

7. Mark McGurl, Everything and Less, The Novel in the Age of Amazon (Verso 2021)

Not so much about AI as about the Internet’s impact on literary culture in general. I forget how this book first came up in our discussions, but I really enjoyed it. I found it refreshing; a fundamentally optimistic perspective on both literature and the Internet. It documents how the rise of Amazon as a global channel for books democratized distribution across a much wider range of distinct communities of interest. The Internet and Amazon together created a world in which every genre of creative writing is potentially equally valid, so long as enough readers want it. For every conceivable niche there are recommendations and best sellers, discussion boards and home grown literary criticism . . . and some very talented practitioners.

Take, for instance, Harry Potter Fan fiction. Manacled by SenLin Yu (not mentioned by McGurl) published 2019, struck me as better-written than the original. Manacled is a 1200-page reimagining of the world that follows after Voldemort kills Harry Potter and the death eaters take over. It is the Harry Potter novel that a darker world needs right now. I’m not alone in thinking so; Manacled has been downloaded over sixteen million times.

Everything and Less made me want to read with a more catholic eye. It reminded me yet again why, in spite of all social media disillusionment (a pro forma disclaimer which I am going to start referring to as “ISOASMD”) the Internet is still on balance a wonderful human advance.

8. Melanie Mitchell, Artificial Intelligence, A Guide for Thinking Humans (Farrar, Strauss and Giroux 2019)

This is the most humane, intelligent, and overall thoughtful book about AI that I have read. It is closest to the book I wish I had written. Melanie Mitchell has a technical computer science background, and she understands fully the ideas she writes about.

Her personality too is very relatable, and it comes through in her writing. Gödel, Escher, Bach, (GEB) is a big part of her origin story, as she loved the book, and eventually went to work for the author, Douglas Hoefstadter. GEB had a big impact on me too; it led me to study philosophy of mathematics, in particular set theory, and then theories of computation, which eventually, through many byways, led to a career as a software entrepreneur.

I think I will read Melanie Mitchell’s wonderful book again right now.

9. Ethan Mollick, Co-Intelligence, Living and Working with AI (Portfolio, 2024)

Ethan Mollick is a business school professor at Wharton, not a computer scientist, and this book benefits from a lively style that avoids math and all technical jargon. He introduces a number of useful and catchy ways of framing the AI debate.

The first thing I really liked was that he described what LLMs do as “emulating” human responses. This is less dismissive than the “stochastic parrot” take that LLMs merely “mimic” human language patterns, and it captures the touchingly sycophantic aspect to today’s AI. A typical commercial LLM acts as if it hopes to convince us that it is as good as us at thinking, and maybe even some day better, but that this is still aspirational - it makes mistakes, apologizes, and keeps on trying.

Mollick has popularized a number of other very useful terms. AI is like an otherworldly intelligence trying to pass for human - “an alien wearing a human mask.” This is quite poetic.

Maybe the most useful new term is “the jagged frontier.” This is the idea that current AI systems are in some areas clumsy and failure-prone, while at the very same time they are extraordinarily fluent and effective in other areas. The outer boundary of AI efficacy shifts with every new training run and model release. To get the most out of these tools, we need to stay flexible, and be willing to experiment to find where the jagged frontier currently lies.

Mollick is a realist. He acknowledges that in education, traditional homework is dead, for instance. And he makes the point that today’s AI is the worst it will ever be. We will need to learn how to include it in nearly all work flows, treating it not a substitute for human intellect but as a gifted but strange collaborator, a “co-intelligence.” I like this framing very much.

10. Arvind Narayanan and Sayash Kapoor, AI Snake Oil (Princeton University Press 2024)

I agree with most of the conclusions in this book. I found this book workmanlike and timely.

Also, I follow these guys on Substack and what they say is always reasonable. Narayan and Kapoor are building a major social media presence with 76,000 subscribers under the name “AI as Normal Technology.” Subscribe. https://www.normaltech.ai/. The “AI as Normal Technology” framing, which I also agree with completely, is that while AI is a transformational technology, we are not about to experience superintelligence through recursive self-improvement any time soon. Neither the millenarian visions of doomers nor utopians will come to pass.

There is nothing at all wrong with this book. I agree with its conclusion. It is just kind of overly normative. Safe.

My reaction to the book, as I try to describe it, reminds me of the time when, as an intern at Yale Press, I was asked to request a dust jacket blurb from Jurgen Habermas for a new book on his philosophy. I explained to Herr Professor Habermas, over the telephone, and with some difficulty, what a “blurb” is, and the very important role it plays in American publishing. He said he would try to accommodate us, so I sent him an advance copy of the book. Several weeks later, the promised blurb arrived.

Translated, it read as follows:

“The book, ________, by_______, is a quite adequate exposition of the ideas discussed, so far as it goes. It can be said that it does not misrepresent my views.”

Sincerely, Jurgen Habermas,

Johann Wolfgang Goethe University, Frankfurt am Main

I loved this blurb. It was an instant classic. Sadly, the book’s risk-averse editor declined to use it.

11. William Powers, Hamlet’s Blackberry, Building a Good Life in the Digital Age (HarperCollins 2010)

Powers was very early to recognize the importance of the changing ways we interact digitally; once the Internet migrated to mobile devices we would never again be without potential distraction. He predicted correctly that constant distraction would - as we now know all too well - lead to fragmented attention, cognitive decay, and feelings of deep dissatisfaction.

Powers recognized that the social dimension of the Internet addresses a fundamental human need for connection with others, but this need in normally to be balanced with periods of uninterrupted solitude. The “alone time” that was once a normal restorative part of life would now require a conscious effort to cultivate and protect. A prescient book.

However, his predictions were not pessimistic enough. Since 2010, even more pernicious problems with our phone-addled existence have become apparent.

First, social media feeds promise the social connection we all want, but fail to satisfy; they leave many if not most feeling worse than before. And the rise of AI chatbots takes this issue to a whole new level, as millions of people seek social connection with a machine that cannot reciprocate human feelings, or even remember past interactions that are no longer stored in its context window.

Second, it is not just social life that is being cheapened. AI chat is invading the solitary thinking aspect of life too. Now we write, think, and even answer texts and emails in the presence of a machine that can substitute for our own internal monologue.

To mix metaphors, our personal island of solitary reflection in the restorative realm of silence, is drowned under a rising tide of clamoring voices.

12. William J. Rappaport, Philosophy of Computer Science (Wiley-Blackwell 2023)

This book is a monumental labor of love by a retired academic. Dr. Rappaport was a full professor at State University of New York, with appointments in the computer science, philosophy, and linguistic departments. For many years, he taught courses in the philosophy of computer science, and he clearly wanted, as the summation of his career, to get all the most important ideas and arguments he taught down in one book. Dr. Rappaport is a very methodical and careful teacher, who goes the extra mile to make everything comprehensible.

I love this book as a resource. However, I respectfully disagree with a number of his conclusions, and believe my counter-arguments are more persuasive. If you are interested, some of my alternative positions are presented in my Substack posts from Spring, 2025.

This kind of respectful disagreement is normal in philosophy - no hard feelings. Respect.

13. Leif Weatherby, Language Models: Cultural AI and the End of Remainer Humanism (University of Minnesota Press 2025)

This book is very good. It has a lot of style. It is eccentric enough to be memorable. This book would be in my top three recommendations from Amy’s List, along with Melanie Mitchell’s and Karen Hao’s. Weatherby covers much the same ground as my least-favorite book on this list - see above - but in a more precise and rational manner, and he reaches quite different conclusions.

Both books are based on the discovery - due to LLMs - that Saussure it right and Chomsky wrong about how language originates. Large language models’ ability to generate new grammatically correct sentences and string them together logically and coherently, all with no prior rules of grammar of rules for recursive organization of complex ideas, should be enough to rebut once and for all the Noam Chomsky Universal Grammar school of linguistics, which has tyrannized over all other perspectives in academia for seventy years.

Saussure thought that signs (words) take acquire meaning (their signification) not from the outside world, but due to the relationships between them. In large language models, words take their variations in meaning from their relative locations in a vector space. There is no absolute position, only relative positions. So, for example, pairs of opposites (”hot” and “cold”) are near each other on one vector space dimension, a dimension signifying “temperatures” but far apart on another dimension, which represents the continuum from extremely hot to extremely cold.

In a famous example, studying the vector space locations of words in the embedding, researchers were able to show that subtracting the vector for “male” from the vector for “King” gave the vector location of “Queen.”

The key point of similarity between Saussure’s theory and current technology is that large language models acquire their understanding of the relative position of words from digesting billions of pages of written language, with no reference to the external objects signified, as they would be understood by humans.

In other words, LLMs build a convincing syntactic model of how language is used, which allows them to generate grammatically correct and convincing sentences and paragraphs, without any knowledge of what words mean to a human being - and without being programmed with any rules of logic or grammar. This sounds an awful lot like Saussure’s idea of language as a system of signifiers built up in reference not to an objective outside world. but to each other.

The way LLMs manipulate tokens in the vector space representation known as an embedding encodes meaning largely, if not completely, through the relative geometric position of the tokens linked to concepts - and this really does seem to be evidence in favor of Saussure’s semiotics. Saussure had no evidence for his idea; it was just a fresh way of thinking about language systems that decentered signification of things outside language itself. But large language models are clear evidence that the capacity to recreate a language, and the knowledge embedded in the texts written in it, can be “stored” in the complete absence of any reference to the world outside the corpus of the texts themselves. So Saussure is shown empirically to have been at least partially correct.

Chomsky’s claim is that all human languages follow common structural rules, and that these rules are innate - found in the physiology of the human brain and its resulting cognitive function. There is no evidence from neuroscience for such structures, and exceptions to Chomsky’s “universal” rules keep being found in the variable structures of various languages, so that the realm of shared innate structures keeps getting whittled down. But Chomsky has nevertheless left little room in academic discussion for alternative theories of how language is learned and used. LLMs show that a purely empirical aggregation of language patterns can be the basis for reproducing language accurately, even though there are no “rules” to be found anywhere in the model.

14. Conrad Wolfram, The Math’s Fix, An Education Blueprint for the AI Age (Wolfram Media 2020)

This excellent book by the brother of the founder of Mathematica, Stephen Wolfram, wrestles with the fact that the current standard curriculum for elementary school through college math teaches the subject all wrong. We spend most of the time on a basket of algorithms for manual computation, yet computation is done much better by computers than by humans.

Wolfram argues that in the real world mathematical process consists of four steps:

(1) define the problem you want to solve,

(2) abstract the problem into the language of math,

(3) solve the mathematical problem using calculation, then

(4) interpret the results back into the domain of the original problem to verify that they are correct.

People whose jobs involve solving actual problems with math spend most of their time on (1), (2) and (4), because (3) is done by computers; but in the standard elementary and high school math curriculum we largely ignore these three steps while drilling endlessly on (3). The result is mass alienation from the subject, combined with complete failure to give students economically useful skills.

Wolfram proposes an alternative: Computer-based math. He recommends shifting the focus away from geometry and algebra to the areas of math that are actually used to run the world, namely statistics, data science, information theory, and risk-analysis.

I completely agree, but Amy tells me that the system by which children are ground through the mill, leading a few to the AP Calculus exams in high school, while the vast majority are traumatized and give up math as soon as they can, is so entrenched that fundamental change is nearly impossible.

A focus on giving students more interesting problems - harder problems - is dismissed as elitism for a few, or as mere “enrichment” that places unreasonable additional demands on already overworked teachers. Frankly, talking about this with Amy was depressing.

But Amy brought up this book because she believes - and I agree - that in light of Claude Code and where the field is with proof-solving AI (using the Lean programming language, for example), Wolfram’s recommendation from 2020, intended to be radical, is no longer radical enough.

The three other steps - defining the problem, abstracting it into mathematical terms, and once solved, translating it back into practical context for evaluation - can all be done by computers now too.

Since math teaching won’t change anyway, regardless of what we say, I hesitate to even bring up what true reform would need to look like. Amy and I, simply by staying abreast of what is happening with AI, are being radicalized right out of the range of acceptable pedagogic discourse.

15. Stephen Wolfram, What is ChatGPT Doing and Why Does it Work? (Wolfram Media 2023)

This is a great short introduction to how LLMs generate text, and why they are so prone to hallucinations and errors when doing even simple math. Wolfram makes an impassioned argument for what is commonly called “Neuro-symbolic” AI - hybrid systems that take advantage of probabilistic pattern matching with neural networks where appropriate, but invoke traditional deterministic rules-based programs to answer questions in math and logic.

[Computer science professor Gary Marcus is probably the most famous advocate for the need for a “neuro-symbolic” approach. https://garymarcus.substack.com/. Marcus’ blog posts and tweets are very much worth reading. He is someone whose opinions I almost always agree with. His books are excellent as well. I only decided to exclude The Algebraic Mind (MIT press 2001) from Amy’s List because I recommended it to Amy and not the other way around].

16. Eleizer Yudkowsky and Nate Soares, If Anyone Builds It, Everyone Dies (Little Brown and Company 2025)

To be fair to me, I would have read this book anyway. The doomer case is presented here coherently and persuasively. But to be honest, it is not as much fun a read as AI2027, which brings a James Bond vibe to the same recursive-self-improvement catastrophe human-extinction scenario.

I don’t believe any of it in any case.

There are many fascinating and disordered intellectual currents that emerged from Silicon Valley over the past two decades. These ideologies are suffused with millenarianism and transhumanism.

It is very important to know about them, because they are core beliefs motivating decisions made by some of the most politically powerful and culturally influential people in the United States right now.

I highly recommend Emile Torres’ The TESCREAL Bundle if you want to dig deeper into the irrational beliefs and inhumane wishes that will likely shape our future.

https://firstmonday.org/ojs/index.php/fm/article/view/13636

Effective Altruism on the one hand, and the Girardian/Ayn Rand anti-democratic axis on the other, have largely evolved in the AI context into two camps: the doomers and the outer-space utopians, and a few people who manage to espouse both at the same time (Elon Musk).

By the way, it is precisely these transhumanist ideas that the Pope takes issue with in his recent Papal Encyclical on AI. (And don’t be fooled, Yudkowsky is actually fine with human extinction so long as the digital intelligence that takes our place is worthy). Rarely in the past century have secular advocates for humanistic values abdicated their responsibilities so profoundly, which makes the willingness of the Pope to speak for the needs of everyone who is not an owner of these technologies even more valuable. Amy is encouraging me to write about the Encyclical, set in the context of over a century of prior influential Encyclicals where Popes have addressed the social impact of technological changes. A great idea for a Substack, for sure.

Ok, so where does this all leave us?

Better educated than we were a year ago, certainly. And humbler.

I know intellectually that, in contrast to scientific invention and discovery, where priority is everything, a book on a topic already explored by an earlier author can nevertheless have many worthwhile aspects. But, Ugh. By temperament, simply doing a competent job at something that has been done before feels anathema to both of us. The thought of being “great, but late” is just plain demotivating.

But wait. Maybe the most famous example of “great, but late” in the publishing world is James Boswell’s Life of Johnson. Boswell’s friend, Samuel Johnson, who died in 1784, was a literary critic, dictionary author, editor, and all-around London man-about-town: what we now call a celebrity. The public was hungry to know more about him. Soon after he died, about a dozen accounts of his life appeared in magazines and pamphlets, followed shortly by two full book-length biographies. Meanwhile, Boswell toiled away on his manuscript for seven more years.

However, happy ending: When Boswell’s Life of Johnson finally did appear, it swept the field.

But, I hear you say, Samuel Johnson was conveniently dead and was not continually adding new witticisms to his oeuvre at an alarming rate.

And Calculus too is a settled body of knowledge at the college level and has been for over one hundred years, so any disadvantage of arriving late to the topic was, for Amy in her Deconstruct series, counterbalanced by the advantage of clearly established boundaries for what concepts are included in any standard high school or college Calculus curriculum and what can be left out.

In contrast to both Samuel Johnson and Calculus, AI capable of passing the Turing Test is only about three years old - obviously in its infancy. Which aspects and enabling technologies will prove most important is not yet knowable. For example, it seemed for several years that generative adversarial networks were the core enabling technology for multimodal models, but then diffusion models largely replaced them. In Mechanistic Interpretability, nothing much has been learned yet; researchers cannot even agree whether the Platonic Representation Hypothesis is real or illusory. Will Transformers be the key, or will they be superseded by some better architecture? Perhaps the true key is Reinforcement Learning with Human Feedback (RLHF)? Nathan Lambert put out an entire book this year limited to the mathematics of the current of-the-moment approaches to RLHF. I think he would be the first to tell you that the whole field is in flux, and many of the methods he enumerates will fall out of favor, as engineers learn through trial and error what works better.

And so the dilemma for our would-be authors:

AI is simultaneously a creature moving too fast to be captured,

and a beast already speared with countless arrows from every conceivable direction.

This week I had another great idea for how to pitch our manuscript. But, if the Internet works wherever Amy is surfing, she will no doubt call in this week and inform me about yet another excellent book to add to this List.

But I have confidence we will figure it out. The world really did need another Life of Johnson.