Convolve & Attend

Introduction: a technical conversation

“If you speak the truth, have a foot in the stirrup.” — Opening quotation

KMOWelcome to another installment of the C-Realm Podcast. I’m your host, KMO, and this is episode number 582, “Convolve & Attend,” with Roger Craig, prepared for release onto the World Wide Web on Sunday, July 17, 2022.

You’ve probably heard that the world is being eaten by software. That software, in the past, was written by human beings. These days it’s written by human beings, but trained on very large datasets.

In this episode, I’m going to have to apologize in advance to a segment of the audience. This is going to be tough going, particularly at first. I talk to Roger Alan Craig about language models, text-to-image-generating AIs, and lots of different kinds of algorithms that are trained rather than explicitly hand-coded. Toward the end, we talk in practical terms about how these things are changing our lives. But for the first half hour or so, it’s going to be pretty technical.

If you get bogged down, and recourse to mathematical explanations isn’t helping, I’d say go ahead and jump ahead to about the one-hour mark, and you’ll find some easily digestible conversation. I’ll let you know right up front: I’ve been reading about this topic and listening to people talk about it for a little while now, and when Roger busts out the math talk and offers up phrases like “a vector of real-valued numbers,” I’m as lost as you are. Let me encourage you to stick with it. If you feel you’re completely lost and a re-listen isn’t going to help, jump ahead toward the end. There is some very accessible, nontechnical conversation.

To introduce the guest: Roger Alan Craig is an American game-show contestant and computer scientist. He held the record for highest single-day winnings on the quiz show Jeopardy! from September 14, 2010, to April 9, 2019. In 2011, Roger returned to win the Jeopardy! Tournament of Champions.

I kind of do Roger a disservice. If he’s really good at Jeopardy!, he’s got a lot of information right at the tip of his brain about a variety of topics. But I’ve asked him to try to translate a very technical topic for a general audience, and that is a tall order.

At a couple of points, you’ll hear that I have inserted some AI-generated text-to-speech, unpacking or providing definitions for things Roger mentions in passing. You’ll notice a couple of different voices. I created these at readloud.net. This is browser-based stuff where you can type in text and hit a button, and it’ll read your text for you in a variety of voices.

I’ll play the conversation in a moment. I just want to let you know I’m including some informal chitchat before the introduction. We’re talking about how we have to be careful about saying the word “Alexa” out loud, because she’s always listening and will often respond when you’re trying to have a conversation with somebody. You didn’t really mean to invoke her; you just meant to reference her, and she’s not smart enough yet to know the difference. So that’s “Mrs. A” or “Miss A.”

Here is the conversation with Roger Craig, and thanks to Elliot for making this introduction.

Program identification: “You’re listening to the C-Realm Podcast. C stands for consciousness.”

Miss A, natural language, and Jeopardy!

Roger CraigA buddy of mine started saying “Mrs. A” when we want to talk about Mrs. A.

KMOYou think she’s married?

RogerOr Miss A. That was what my friend said, and I just went with it. I didn’t even think of the marital status.

KMOMadame, whatever—Mademoiselle. I’m sure lots of people have at least a low-level crush on Miss A.

RogerWithout jumping too far ahead, that plays into the whole thing with the guy with LaMDA. There is this trend now of people developing—I don’t know if it’s parasocial; that’s more with celebrities—but these relationships with devices they’re communicating with.

KMOI’m going back to the bio you prepared and letting everybody know that I’m talking with Roger Craig, who is the principal at Signal Alpha, a machine-learning consultancy working primarily in natural language processing. He’s worked on machine-learning projects in molecular biology, protein–protein interaction prediction, biodefense, finance, insurance, natural feuds, communication surveillance—that’s a little spooky-sounding—aviation, and electronic medical records.

Among other things, you’ve been on Jeopardy!, and you are likely most well known for holding the single-day winnings record from 2010 to 2019 and winning the Jeopardy! Tournament of Champions, using NLP—that’s natural language processing, not neuro-linguistic programming—to prepare for the show.

RogerJohn Bandler didn’t help me, no.¹

KMOI guess that’s a great place to start. How did you use NLP—again, natural language processing—to help you prepare to compete on Jeopardy!?

RogerI used that mostly to bring order to chaos. What I mean is, you have a large corpus of data. There’s a website called the Jeopardy Archive, where volunteers transcribe all the questions on the show every night. I basically downloaded the website. This was back in 2007, 2008.

Essentially, I wanted to quantify what that corpus was like. What percentage of the questions were on Shakespeare? What percentage on presidents, on chemistry, whatever it was? Then I sampled from those. I made a little web tool, answered questions, and saw if I was right or wrong. Then I got self-knowledge.

This tied into the quantified-self movement: first getting a static snapshot of how well I was doing in each category, then deciding which ones to tackle, then seeing if I was improving over time. That was the main way.

There was no chatbot or anything. It was really more about what would be termed text clustering and topic modeling: taking a large body of documents, a corpus that you never have time to read through, and that you don’t want to manually label. Maybe you manually label a small percentage, and then you want to understand what it’s talking about.

This is a very useful concept and comes up again and again. We’re much better at it now, and a lot of these NLP models are much more useful in labeling all this unlabeled data in various ways, in terms of topics.

KMODid you say you were doing this back in 2007?

RogerThat’s right.

KMOIt’s my understanding that neural networks weren’t good for much back in 2007.

RogerThat’s a great point. This didn’t involve any neural networks at all. Think of it as—okay, to get technical—it’s really just counting. It was tokenizing the documents, these questions and categories on the game show Jeopardy!, and counting the appearance of the words.

This does segue into what we want to talk about. Say you have 250,000 questions. The first thing you can do is see how many unique words you have in there. You might have 110,000 unique words, or something along those lines.

KMOThe English language is large.

RogerYes, because the English language is large. This is the old-school NLP method, almost akin to information retrieval. You can represent every document as a vector of length 110,000. It wasn’t 100,000—you cap it off. You’re going to have a long tail of very infrequent terms, or terms that appear only once, twice, or three times. You don’t care about those as much.

You care about the more frequent terms, but not the head terms: “the,” “of,” “and,” all the stop words in natural language processing. These would be more the terms like “president,” “vice,” “cabinet,” “White House,” “capital.” Those terms tend to co-occur, and now you have your government or history cluster. So I did not use neural networks at all. Short answer. Sorry.

KMOYou didn’t then. You are working with them these days?

RogerYes. Everybody is, pretty much. If you’re not, you’re playing the wrong game.

How neural networks became useful

KMOFor your information and for the edification of the listeners—many of whom already know this—I went to grad school for philosophy in the ’90s. My specialty was recent Anglo-American philosophy, with emphasis on philosophy of science and philosophy of mind.

Philosophy of mind in the recent Anglo-American tradition is analytical. It’s basically synonymous with AI: a natural-language discussion of how you structure a mind and how a mind would interact with the world, without any programming. You’re not building anything that does anything. You’re modeling stuff in an abstract space.

At the time, there were champions of neural networks, but they were basically talking up the potential. This is 1994, 1995. Nothing really had been built that could approach the utility of what was called good old-fashioned AI: human-coded, explicitly structured programming.

There was an idea that you might be able to train neural nets, and talk of weighting certain connections and reinforcing particular outputs to shape this thing over time, so it could, in theory, someday be as good as, if not better than, stuff that’s explicitly hand-coded. But when I was in grad school, that was still speculative.

I had been out of grad school for more than a decade, a decade and a half, by the time 2012 rolled around. Looking back, that’s the year a variety of well-known papers, winners of contests, and university programs really started to get things rolling.

A longstanding problem in artificial intelligence and robotics is computer vision. I think there were a lot of breakthroughs in computer vision that year, which spilled out and had effects in other fields as well, including natural language processing. Usually the name to reference for this period is Geoffrey Hinton and his team, I think, in Toronto.

From that moment on, suddenly neural networks and algorithms that you train rather than program—train on large datasets—stopped being in the realm of speculation and of interest only to people in AI or philosophy of mind, and became of interest to people who wanted to make money. The last decade has basically been the implementation phase of that revolutionary period when significant breakthroughs were made.

I’m speaking in very general terms. Let me turn it over to you and ask you to flesh that out with some specifics.

RogerThat’s a great synopsis. There were a few smaller wins in neural networks prior to then, going back to the ’80s. You had Yann LeCun doing stuff with digits, the classic MNIST dataset: handwritten numerals that would appear on checks or mailed envelopes.

Also in the ’90s—these were very narrow—a group at CMU actually had a self-driving car with neural networks. They drove it around the highways around Pittsburgh, I believe.

KMOThey were on the roads with human drivers, not just on a track?

RogerI’m pretty sure. I haven’t read about that recently, but I know they drove it. I’m sure it was on a sunny day with perfectly marked lanes. There were probably no other cars around. It wasn’t driving in Midtown Manhattan at 5 p.m. on a Friday through Times Square. That’s going to take a while. And also when it’s raining and snow’s on the ground.

There were some wins there. But during that period, it is true that neural networks were sort of in their own little AI winter. They weren’t the hot thing. I’d like to speak a little to that. There are interesting reasons for it.

One is that most of this work was being done in academia. I’m sure industry was working on it, but they weren’t advancing it much. As you know, the culture of academia is all about publish or perish, and showing how smart you are. It’s only recently that a lot of these advances are about doing better on a problem versus the actual brilliance of the method you have come up with. There were other methods that, frankly, allowed computer scientists and mathematicians to use more sophisticated or interesting mathematics. That’s what they like doing.

It’s what you say with the handcrafted system. A big trend in all of this is human hubris: we’d like to design some system we’ve put a lot of intelligent thought into, taking all this learning through human history and encoding it into the system.

On the other side, you take a bunch of data, do logistic regression on it, transform the data in some way, and get something that outcompetes anything the most brilliant people in the world can come up with. That can be interesting or fascinating. It could be depressing, depending on your point of view. It has huge consequences for the future when you think about who has the large sets of data, who has the compute power, the money to do these things.

Convolution, compression, and autoencoders

RogerGetting to 2012: AlexNet, named after, I believe, a grad student of Hinton’s by the name of Alex—and I’m blanking on his full name²—devises a convolutional neural network to do image classification. The existing state-of-the-art methods weren’t using neural networks. They were using algorithms based on edge detection and segmenting images.

I should wrap all of this with: I’m not an expert on computer vision. This is my high-level understanding.

KMOI’ll extend your caveat and say you had no notion whatsoever that I was going to quiz you on the history. You were prepared to talk about the present day. This is all off the cuff.

RogerAnd I’m willing to discuss, like I said, anything and everything. AlexNet was a convolutional neural network.

KMOLet me have you pause there. What is a convolutional neural network?

RogerMost usually, these are two-dimensional convolutions. You take in an image. Typically, you would make it square—it doesn’t have to be, but that simplifies everything. Then, through the layers of the neural network, you’re essentially shrinking the square through convolutions.

You look at all the elements in the matrix. You’ve converted the image into a matrix or a tensor. Without getting into the nitty-gritty details, you take a larger image and represent it in a matrix. The first thing you feed in is, say, the RGB values or grayscale values of the pixels in that image. Then you shrink it down smaller and smaller. Each layer in the neural network will have a smaller square.

That encodes it. You’re getting a smaller representation of the image, using less information, but you’ve captured the essential parts. Once you have that, you can take that representation and do classification on it.

Think of the representation as a vector of real-valued numbers. If you started with an image that was, say, 1,000 by 1,000 pixels, it has a million pixels. Let’s say it’s grayscale: each of those is zero to 255. The image you feed in is a vector of length one million. Doing these convolutions, you get it down to a much smaller vector, which might only be a few hundred or a thousand numbers.

Now you’ve put this image and compressed it into a thousand-dimensional space, or a 300-dimensional space, whatever it is. You can keep going smaller and smaller. You lose information each time, but you’re also generalizing. You’re capturing the essence of the image.

When you train the model, those convolutions—how to shrink it—are being learned during the training process. One way to train them is by using what’s called an autoencoder. You can convolve and shrink it down. Then you can do the opposite, and have convolutions that build it back up to 1,000 by 1,000. A lot of Hinton’s early stuff—early as in 10 to 20 years, and even going back to the ’90s or ’80s—was the concept of an autoencoder.

What is an autoencoder? You have the image as the input to the neural network, and as the output, the same exact image. You have it learn to encode its own image. It won’t be perfect, because you’re losing information, shrinking it down to a much smaller vector. Then you’re learning how to reconstruct the original image.

Initially, the network is initialized randomly with all the weights, so you get almost random noise out the first time. Then you compare it, and using backpropagation and these methods, you adjust the weights until you converge—or minimize your training loss to something where the output image most closely resembles the input.

When you train that, you can remove the bottom half. You have the encoding at the top and the decoding at the bottom. The thing right there in the middle is your representation. Let me pause there for a second.

KMOFirst, I’ll say that when you start listening to people talking about AI, you discover they use language rather differently than people speaking conversationally out in the world. For example, they use the word “compute” as a noun.

RogerOh, yes. Compute.

KMOWhich I take to be short for computing power or processing.

RogerThat’s probably what it is. It’s saving two or three syllables.

KMOTypically, the only time people in normal conversation use any form of the word “convolve” is as an adjective: “Dude, that’s convoluted,” which is to say it’s poorly structured. But convolve is a verb. It means to entwine. In AI talk, these don’t mean what they mean out in the larger world.

You’ve been talking about the process of convoluting.³ I don’t want to spend a whole bunch of time on it, but maybe it needs another minute. When you invoke that concept, what’s a conversational-level shorthand? What is involved when you’re talking about a convolution?

RogerThat’s a great question. One definition of convoluted is something completely hard to understand. Looking up a definition in Wikipedia, they have it as a mathematical operation on two functions that produces a third function. That’s one definition.

But I’m not an expert on computer vision. My understanding of convolutions is: let’s stick with 2D matrices. You have some function that operates on the cells in that matrix to produce a new matrix. The new matrix can be the same size, smaller, or bigger.

A more easily comprehended example would be Photoshop and GIMP. Those image filters, when you do blur, sharpen, or edge detection, are convolutions. You look at the values of the pixels. You can have a three-by-three window and go over all the little three-by-three windows in the whole image, multiply those values by your convolutional kernel, and end up with new values.

For blurring, edge detection, and sharpening, that will be the same size, because you’re just going to blur or sharpen the image or do edge detection. That’s what we’re talking about with convolution. Think of it as a way to take a matrix of numbers and map it into a new matrix of numbers, where at the most basic level you’re multiplying and summing them in some way to get new numbers.

Text-to-image systems and diffusion

KMOWhat’s big in the news for me is text-to-image generation programs like DALL-E 2. DALL-E 2 is the big one. It’s the one I don’t have access to, so I have to content myself with things like Wombo Dream and Midjourney.

These are amazing devices. You put in a text prompt: “Show me a teddy bear walking through Times Square.” DALL-E 2 is astounding. It gives you these photorealistic images of a teddy bear walking through Times Square or riding a skateboard. The other things I have access to aren’t quite that good, but still they do amazing things.

You give a description of what you want, and it shows you something that isn’t anything like what you visualized, but it’s still amazing. There are simulated brushstrokes, play of light, interesting compositional things and textures—stuff that would take me, as a graphic artist doing stuff by hand, hours and hours. Maybe I couldn’t do it at all. Maybe there are techniques there I just don’t know. It does it in a few seconds.

At the same time, it’s incredibly frustrating, because it’s nothing like what I wanted. I had an image in mind, and I came up with a sentence to describe it, but what the computer comes up with is nothing like what I had in mind. It’s very interesting in its own right.

What it’s doing is starting with a random static field, basically a screen of gray with a few blotches of darker gray. It runs this convolutional process over and over. Each time it transforms the image, it’s looking for familiar patterns statistically correlated with words in your prompt. Each refresh, you can see it’s a little closer to the final product.

Sometimes you’ll have what looks like the perfect final image, and you’re like, “Okay, stop!” But there is no stop command. It keeps going and overwrites and obliterates this beautiful mid-series stage. It is running a sort of convolution, as you’ve described, to create these—at least Wombo Dream and Midjourney, the ones I have access to and can observe. That seems to be how they’re doing it, just as an illustration of what you’ve been describing.

RogerI think you’re absolutely right. I think it’s slightly different than a convolution to reconstruct the image. Typically, these are diffusion kernels. They trained it by adding noise to the images in the training data. They progressively add the noise, then flip that and train the neural network to take noisy images and make them less noisy.

What it does is take your text input, generate a representation of that, and use that representation to guide the removal of the noise. You’re right: it starts with random noise, typically on a square canvas, and then starts to generate the image that will match, or be close to, the representation of the text.

This is really interesting because it’s text and images. You have representations of the text and representations of the image, and you try to align them during the training process. When you put in Kermit the Frog, or a teddy bear skateboarding in Times Square, it has a representation of that because it’s been trained on images that have teddy bears, Times Square, and skateboards. Maybe none of them had all three, but it’s able to compose them.

Then there are the new ones from Google, like Imagen—which is a brilliant name. I’m surprised no one’s ever used that name before. Image, gen, Imagen. And then Parti. You just add an N to the word “image,” and it’s like image-gen. Oh my gosh.

KMOSomebody got a nice bonus for that.

RogerThere are six-letter portmanteaus out there that have been unexplored, and they’re still good.

KMOI’m sure there are neural-network algorithms exploring that space, looking for interesting combinations.

RogerOh yeah. I was actually lucky enough to get DALL-E 2 access a few weeks ago, so I play with that most days. You really overdose on it in the beginning. It’s just unbelievable. It doesn’t do text right.

I would like to try Midjourney; I haven’t looked into that at all. There’s also Disco Diffusion, and some Colab notebooks where you can run those. Interestingly enough, some of these models are coming out of Russia, where you actually use Cyrillic text to generate the image. If you put English in, it Google-translates it, or whatever, to Russian and then generates the images. Very akin to using these convolutions to reconstruct an image.

Generators, discriminators, and artificial faces

KMOLet’s push on. I think you may have mentioned, just in passing, adversarial networks, or adversarial—what’s the word?

RogerGANs: generative adversarial networks.

KMOGenerative, there you go. What are those?

RogerIt’s G-A-N, generative adversarial network. It’s the combination of two things. You have a generator, which will generate images. We were just talking about text-to-image; in the original formulation of these, there’s no text involved at all. Instead, the first GANs, I believe, were generating human faces. That’s where you see things like StyleGAN and thispersondoesnotexist.com.

KMOWhich generates photorealistic, fool-the-eye images of human beings. If somebody showed you this image and said, “I ran into this guy at the mall,” you wouldn’t question it. You wouldn’t say, “I don’t think that’s a real person.” They just look like photographs, but they’re not any particular person. They’re computer-generated combinations of features. No living human being actually has this particular combination of features, most likely.

RogerYou wonder how many LinkedIn profiles, Facebook and Twitter headshots, are really thispersondoesnotexist screen grabs.

What the generator does is learn how to generate images—specifically faces, let’s say. Then you have another part of the system, which we call the discriminator. It just has to make a binary decision: is this a real image, or was this image generated by the generator? Is this a real photograph of a human, or was this created by the generator?

That’s where the adversarial aspect comes in. It almost becomes a game. They’re competing against each other. The generator is trying to fool the discriminator into not being able to tell the two apart, and the discriminator is trying to detect the fake images as best it can. They play this game and both get better. The generator ends up getting really good at fooling the discriminator, and that’s how you get the photorealistic effects.

When they first came out, they would have strange artifacts: everybody would have two different-colored eyes, or they’d be wearing an earring on one ear, or their ears were shaped differently. They’ve done a lot of work and fixed a lot of those issues. You used to be able, if you knew what you were looking for, to look at an image and pretty easily tell if it was a GAN. Now they’re getting so good. Scary.

From images to language

KMOWe’ve been talking about imagery, I think mostly because I’m a visual artist and interested in that sort of thing. But I know that your day-to-day work is with natural language. What’s a good segue to take what we’ve talked about so far and port that over to natural language processing?

RogerOne key thing I always like to talk about at a high level is the difference between discriminative models and generative models. Discriminative models have to discriminate: they have to make a decision, usually classification. Is this a fake image of a face, or a real image? Is this movie review positive or negative? Do we give a loan to this person? Is this person going to vote for candidate X? Is this person going to click on this ad?

KMOPredicting human behavior is definitely a big commercial aspect of all of this stuff.

RogerYes.

KMOShoshana Zuboff wrote a big, long book about it called The Age of Surveillance Capitalism. She talks about how the big tech companies are hoovering up the data exhaust we create all day, every day with our devices, building models about us that are predictive as to our tastes, purchasing decisions, and voting decisions. Their marketing departments figure out who’s the best customer to sell these models and predictive products to.

This is the commercial application, which is not new. It’s developed; it’s pretty mature at this point. It’ll probably get much more sophisticated as time goes on, but it’s not nascent. This is an existing industry.

RogerExactly. It’s huge, and we only see glimpses of what is occurring. Discriminative models were what everybody was doing in the beginning, and it’s still huge, like you just said.

But then you also have generative models. These aren’t making a decision about the input. Instead, they’re learning from all the training data how to generate new instances, never before seen. That can be images, as in DALL-E 2, Imagen, Parti, Midjourney, and all the rest. Or it can be text, most famously with GPT-3, or recently in the news, LaMDA, the dialogue application out of Google.

We could also take a step back. You mentioned philosophy of mind. Let’s think about our own minds. For images, we have a discriminator: we look at something and know if it’s a bird or not. If it’s flying in the air, it’s probably a bird—just so it’s not a bat, a flying squirrel, or a dragonfly. We have exceptions to these rules, like a kiwi or a penguin.

Then we have generators inside our heads. If we want to sit down and draw a bird, we can draw a bird. A toddler can sit down with a crayon and draw a bird. Same with text: somebody says something to us, we process it, and we’re like, “Is that a question? A command? A salutation? A valediction? What are they saying to us? Is it expletives?” Whatever situation you’re in. Obviously we have generators: we can speak.

Even learning a new language, you have text, audio, and sound, but we have recognition first and production second, typically. A baby has to listen to a human language before they start speaking it.

To segue into natural language processing, in machine learning you typically have three things: data, which is labeled or unlabeled; representations of that data; and a learning algorithm, depending on what you want to do. A key part is these representations in the middle. Your data can be any modality—text, images, labeled or unlabeled—and your learning algorithm could be decision trees, linear regression, or neural networks.

Neural networks can also generate representations that you then take and use in something else. These pretrained language models have internal representations of the text. You can export those and use them in a completely different system that doesn’t involve neural networks. With BERT and RoBERTa, these encoder natural-language-processing networks, you feed in text, get representations, and then put those into a system like semantic similarity search. This will improve the search, hopefully, versus other methods.

KMOI’m not familiar with BERT and RoBERTa. What are those?

RogerLet’s start with BERT. That came out in 2018, a year after the first transformer paper in 2017. It’s a neural-network model that you feed text into, up to a few thousand characters or tokens, and you get a representation that is a 768-dimensional vector.

With that vector, you can put another layer there for classification and classify the sentence or piece of text as positive, negative, or neutral sentiment. You can also put in two different sentences and ask: are these sentences similar? Are they contradictory? Is there entailment?

As AlexNet was in 2012 for images, BERT was that for NLP in 2018. I think that’s a fair statement, because it broke all the previous state-of-the-art metrics on various datasets.

What does a transformer attend to?

KMOYou’ve mentioned the word “transformer” in passing. It needs a definition.

RogerThat’s not an easy one to define, from my experience. Let’s motivate it. Before the transformer, most use of text data in neural networks used recurrent neural networks and things like long short-term memory, created by Jürgen Schmidhuber—we need to give him credit.

These neural networks operated on one token at a time. The structure of the network would encode how it was understanding the text. Once you got to the end of the text, you would sometimes lose the thread of what was in the beginning. This is a big simplification.

When transformers came out in 2017, they said: let’s have the text we’re interested in look at all the other text in the same piece at the same time. Essentially, you get a big matrix. You tokenize the text. Let’s say we just split it on whitespace for now—they actually do something different—or split it into characters. I think whitespace is better for English speakers. We get a matrix and have it look at it all at the same time. Then we introduce attention: where should it be focusing in this matrix?

In the computer-vision world, they were using convolutional networks, and then introduced attention. I’m pretty sure attention was introduced in computer vision before text. In vision, it allowed them to decide which pixels to focus on. You can think of the biological analogy as the fovea of our eye. We can see a whole visual field, but it’s only what we’re—

KMOWhat part is the fovea?

RogerThe fovea is the part of the retina where the rods and cones are at their most dense. If you think of the bandwidth of the optic nerve coming out of the back of the eyeball, the vast majority of the information is coming from this small part of the eyeball, right in the back center. When we look straight ahead at whatever’s in focus, it’s almost in the middle of our visual field. It’s not off to the sides—that’s our peripheral vision, and it’s usually much blurrier. But we can look to the side and it becomes sharper.

The thinking is that you don’t have to look at all parts of the image; you have to look at the interesting parts. As human beings, our eyes flit around, whether we’re reading, looking at an image, or walking around in everyday life. You don’t just keep staring straight ahead. Your eyeballs move around and take in parts of the scene. If things aren’t changing, you tend to move your eyes. But if you’re watching television or a movie, you just stare at the TV forever.

Attention was a way to encode in the neural network which parts of the data to focus on. With text, we don’t focus on the “ands,” “a’s,” “thes,” and “ofs.” We focus on the least probable words that have the highest semantic input. If I say, “The World Series was won by the New York Yankees,” you’re not really paying attention to “the” or “was.” You hear “World Series,” “won,” and “New York Yankees.” That’s what we’re attending to.

Coming back to the transformer, it has multiple layers. BERT had 12 layers. Each is a different attention layer, transforming the initial text we’ve encoded. We learn a new representation of the text, and do that 12 times in a row.

In the beginning, it usually attends to the neighboring tokens. It has positional encodings: we’ve told it where all the text is—this is the first word, this is the second, this is the third. Three mostly attends to two and four; four mostly attends to three and five. Three will have some attention to the ninth in the beginning, but less.

As you train BERT, for instance, if you pass in “The World Series was won by the New York Yankees,” “World” and “Series” start to attend to each other more strongly because they go together. You pass in billions and trillions of sentences, and “New York Yankees” will attend to each other. Once you get down to the lower levels, the twelfth layer, you’ll see “New York Yankees” attending to “World Series,” having a long-range correlation.

That’s a key thing in machine learning. It was always relatively easy to capture short-range correlations over a sequence. You have a sliding window and can look at your nearest neighbors: today, yesterday, tomorrow. It’s harder to capture a correlation between today and a month ago, two years ago, or a month in the future. The key was to have the whole text you feed in be able to attend to all the other text at the exact same time. That way, you can capture a long-range correlation from the end to the beginning, or anywhere else in the string.

KMOLet me have you pause and, if you can, give me a definition of “transformer” that involves 20 words or less.

RogerI’m sorry if I’m a little long-winded there. Twenty words or less—it’s almost like a haiku. The transformer is a neural network that uses attention such that—I shouldn’t say things like “such that.” Strike that.

KMOI’m not counting words.

RogerThe transformer is a neural network that uses attention where any element in the sequence can attend to another element of the sequence and help influence its representation. They influence one another.

Let me back up. Some of these things can be viewed as black boxes or gray boxes.

Explanatory insertA black-box tester is unaware of the internal structure of the application to be tested. While a white-box tester has access to the internal structure of the application, a gray-box tester partially knows the internal structure, which includes access to the documentation of internal data structures as well as the algorithms used. Gray-box testers require both high-level and detailed documents describing the application, which they collect in order to define test cases.

RogerI should mention, I’m not researching these. I’m more a practitioner and user of them. They are incredibly complex. What it really gives you is better representations. You can have this correlation between the text, and it captures that better than any existing methods before it. When you feed in “The World Series was won by the New York Yankees,” that becomes a context and controls the representation.

Anyway, I went way over 20 words. I’m sorry. My point is that these are very hard to understand. These models end up having a billion parameters. No one really knows how they’re actually working. They just work, and you can interrogate them and see certain things. Then you get into explainable AI and interpretable AI.

Before neural networks, in traditional NLP, you would take text, preprocess it, tokenize it, detect the parts of speech, and do named-entity recognition—all these different layers. When interrogating something like BERT, they actually see these 12 layers sort of relearning those exact steps the humans had done in the eighties, nineties, and early 2000s. I think that’s really interesting.

Prompt engineers and creative centaurs

KMODoes the name Conor Leahy mean anything to you?

RogerConor Leahy? No.

KMOI listened to a couple of interviews with him just yesterday. I can’t recite his CV from memory, but he was talking about a lot of these same topics. He was talking about how a lot of the people who become really good at providing input for image generators or something like GPT-3 to generate new text—he calls them “prompt engineers”—are typically not particularly technical people. They might write some code, but they’re not developing these systems. They’re people who are creative, smart, curious, and willing to spend time interacting with them. They get really good at learning what words to feed into the system to get a satisfying output.

He was talking about how, after Deep Blue beat Garry Kasparov—the chess grandmaster, world champion; at one time, if you could beat him, you were the best chess-playing entity in the world—there was a period where a grandmaster playing in conjunction with a chess program could beat the best chess-playing programs. Kasparov called those pairings “centaurs.”

There was a period when centaurs would beat the best chess-playing algorithms. Then that period ended, and now the best chess-playing entities don’t have any human components at all. There is no use for chess grandmasters anymore.

Conor was talking about how we’ve entered this period where, if you want really satisfying computer-generated novels or imagery, you’re still going to get much better results with a creative human artist or writer interacting with a system: editing, giving it prompts, taking its output, keeping what’s good, cutting what’s bad, pasting it together, maybe rewriting it.

We’re in this period where the creative centaur is still outperforming the algorithmic artist. But like the chess-playing centaurs, this is a finite period. There’s going to come a time when humans in the loop are just an impediment. They’re not helping at all.

This is much more high-level and conversational than what you’ve been talking about so far, but I’d like to switch to that level for the end of our portion here on the free C-Realm Podcast. What do you see in your work where there are possibilities for people who don’t write code or have engineering degrees, but who have general intelligence, to interact with some algorithm, neural network, or product of your work to create something new and exciting?

RogerI think the best way to approach that is with DALL-E 2, with text-to-image generation. By the way, I looked Conor up. He’s one of the principals at EleutherAI, and I’ve actually watched an interview with him, but forgot the name. I apologize, Conor.

KMOI listen to a lot of interviews. There are a lot of names.

RogerOne thing about chess: I would dispute the chess-centaur speculation that a human plus a computer cannot beat a computer. Are we saying that if you had Magnus Carlsen and AlphaZero from DeepMind play AlphaZero, Magnus Carlsen would add zero to that? That over the long run they would have 50 percent equity on win share?

KMOI am just reporting this observation from Garry Kasparov. This is not a statement that I defend or have any investment in. It’s not a debate I would be interested in entering into.

RogerOkay, I take that back. Conor didn’t put that forward; Garry Kasparov did. Garry Kasparov is extraordinarily knowledgeable about chess, and very knowledgeable about chess-playing computers. That’s an understatement.

Interestingly, I met him—well, he walked right by me. I went to one of his matches in New York in 2003 against Deep Junior, which they ran on a laptop. That was different from Deep Blue.

Explanatory insertThe final game of the epic man-versus-machine match between Garry Kasparov and Deep Junior ended in a 3–3 tie.

RogerAs for prompt engineering, with DALL-E 2, Midjourney, and all these coming out, that’s exactly what’s needed right now: exploration of prompts. There are subreddits and Discord servers specifically dedicated to DALL-E 2, and probably all the others, where people share their best prompts and what pieces of text you need to get the image results you want.

For instance, if you put in space scenes—a meteorite hitting Mars—if you put “courtesy of NASA/ESA,” the European Space Agency, you get much better results. It looks cooler to the human eye.

KMOOr you can end your prompt with “Unreal Engine.”

RogerYes, Unreal Engine is great, especially for digital art. And you do “digital art, trending on ArtStation.”

KMOOr you can end your prompt with a description of the camera lens, aperture, and f-stop, and get an image that looks like it was photographed in a particular way with particular equipment.

RogerIt’s pretty wild. Even more simply, put “35 millimeter,” “50 millimeter,” or “18 millimeter” for the focal length, and it changes the image. You can even put the model of the camera. I’ve done prompts with DALL-E 2 where it’s “Fuji Pro,” whatever camera. I have a Nikon myself.

What’s interesting to me is that you almost have to arrive at it by dumb luck unless you’re able to interrogate the training data. Some of the open-source text-to-image algorithms are being trained on the LAION dataset. I think that was 500 million images, and now it’s like four billion. They have a website where you can type in text prompts and see images and their captions.

As you’ve probably seen with Midjourney, and I’ve seen with DALL-E 2, when you put in the text prompt, you don’t always get exactly what you want. If you play around with it, you might. However, some things are insurmountable. I saw one on YouTube where it was a raccoon hugging a pancake. In all the images, the pancake doesn’t have arms; it doesn’t have this concept of an anthropomorphic pancake hugging. Instead, the raccoon is eating the pancake.

I put in prompts like “Ronald McDonald eating sushi near the Eiffel Tower,” and in all the images he has a hamburger, because it’s Ronald McDonald. It’s sometimes hard to overpower those.

I do believe prompt engineering is a really interesting area. For text, you have all the voodoo for prompting something like GPT-3, or open-source generative models like BLOOM, which just came out of the BigScience group at Hugging Face within the past week.

Hugging Face, shared tools, and a lost movie

KMOI’ve been digging into this stuff, so I know what Hugging Face is, but assume the audience does not.

RogerIf the audience is familiar with GitHub, let’s talk about that for a second. GitHub was a startup that allowed people to make Git repos.

KMOWhich are what?

RogerA Git repo is a version-control system, typically for code, but it doesn’t have to be. Going back, there was Subversion and CVS, way back in the nineties. Just like you can see different versions of your Word docs, Google Docs, or a Wikipedia article, versioning is very important. Interestingly, version-control software was not really adopted everywhere until the past 20 years.

On GitHub, people can put those repos up, keep them private, or share them with the world, and you can download them. Hugging Face is an analog for that. Instead of repos of written code, it’s models, datasets, and graphical front ends to those models and datasets that people can download, interact with, and build off. They’re trying to be the GitHub of ML.

Based on the money they’ve received, I think they have a billion-dollar valuation now. Having said that, I don’t put much credence in VC valuations; I put credence in the actual product. I think their product is really good. They make it so you can have six lines of Python code and be predicting the sentiment of a sentence in English. It’s extremely powerful. If you don’t even want to do that, they have a graphical front end where you can type it in.

It’s not just text; they do images and audio. There are spaces up there where you can put two images together, and it will interpolate between them and do frame insertion. You can make slow-motion videos where the neural network is creating the intermediate images or transforming something.

KMO“Tweening,” we would say in animation.

RogerTweening. On animation, you know the Pixar story about Toy Story? Have you heard this, about version control?

KMONo.

RogerThe story goes that during the creation of Toy Story—one, Toy Story—there was a massive deletion of a lot of the assets at the company. I forget what happened, if it was a “remove star” in the root directory on the server or whatever. But one of the employees was at home for maternity leave. She had her workstation at home and, because it was easier to work this way, had downloaded almost all of the movie they had up to that point. She didn’t have everything.

During the creation of Toy Story, they deleted everything, but got back about 90 percent. Then they had to reconstruct the ten percent they were missing and finish the movie. They weren’t using version-control software. Of course, that was 1995 or whatever.

KMOWhat was the relevance of the woman on maternity leave working from home?

RogerThat she had all the data. The data at the actual office, the Pixar studios, all got deleted.

KMOI see.

RogerFor some reason it got destroyed, and they couldn’t recover it. They were lucky that one employee was working from home. This was 1995, not 2022. She had downloaded almost the whole movie because it made it easier to work with it. This was also dial-up internet. Maybe they didn’t even download it; they probably did sneakernet with the hard drives.

KMOThat’s a term I haven’t heard in a long time.

RogerBut it’s still one of the highest-bandwidth ways to move data around.

KMOPick up a hard drive and drive it somewhere.

LaMDA and the claim of sentience

KMOI’d like to talk about LaMDA. You mentioned that earlier. I’ll weave it into the conversation like this: Blake Lemoine, the person who has claimed that LaMDA is sentient and deserves legal representation and much greater consideration from Google than it’s getting, is not a technical person. He is kind of a dilettante like me who’s moved around. His seven years at Google is the longest he’s kept any one job in his whole life. He’s a Christian mystic, I think of the Gnostic variety—a person with great imagination. He’s more creative than analytical.

He was charged with interrogating LaMDA to ferret out any biases it might have. This is Lemoine’s contention: the only biases Google really cares about are the sorts of things that are going to get it sued. It can have all kinds of biases that are harmful and could have unpleasant effects down the line, but as long as it’s not going to result in a lawsuit against Google, Google doesn’t care. They’re looking to make sure LaMDA is not deadnaming anybody, is using the right pronouns, and is not using prejudicial racial language—things like that.

The point I really want to emphasize is that Blake Lemoine doesn’t have the skill set to examine the code, to look at the weighting of the different layers in the neural network. All he can do is compare his input to the output LaMDA is giving him. After many months interacting with this system, he came to the conclusion, so he claims, that it was sentient.

I’d like to talk about that, and we’ll continue it in the C-Realm Vault portion of this conversation.

Program identification: “On the C-Realm Podcast, a mind that is stretched by a new experience can never go back.”

KMOThat was Roger Craig, and that was the technical talk. I’ve got more than an hour of unstructured, free-ranging conversation with Roger for the C-Realm Vault. But if you sat through this much technical stuff, here’s a taste of the easier stuff. Here’s about the first 17 minutes of the conversation that was supposed to be reserved for the C-Realm Vault.

Vault excerpt: metrics, roles, and prompting

RogerI’m forgetting the name of that, where the metric becomes the target— I’m sure we could Google it, or ask Mrs. A... or Ms. A.

KMOAlexa, what do you call it when the metric becomes the target?

Alexa[muffled reply]

RogerIt’s Goodhart’s.

KMOGoodhart’s Law. Yeah. As soon as the metric becomes the target, you get absurd behavior meant to game the system, but which is nonproductive.

RogerOn that note, it might be a nice segue to LaMDA. If you look at the LaMDA paper, the thing they talk about a lot is their metrics. Let me pull it up so I don’t—

KMOWhile you’re searching, I will relay something relevant. In one of the conversations I heard with Conor Leahy yesterday, he talked about a particular language model he was working with. You could say, “How would you translate this into French?” and it would give you a translation. Then he’d say, “Okay, but you are a very smart, sagacious, poetic translator. How would you translate this into French?” And you’d get different output.

Basically, you can make the system smarter by telling it that it’s smarter. It will take on the role of a thoughtful, poetic, or erudite translator.

This comes into play a lot with LaMDA and Blake Lemoine, because there are videos Google put out before the Lemoine thing. They were trying to promote LaMDA, and highlighted two sample conversations. In one, the first thing is: “You, LaMDA, are the planet Pluto. I’m going to ask you questions. You will answer from the perspective of the planet Pluto.” Then there’s a conversation, and LaMDA is speaking as Pluto. There was another situation like that: “You’re an umbrella. I’m going to ask you questions. Answer from the perspective of an umbrella.”

Blake Lemoine’s published transcript, which he says demonstrates that LaMDA is sentient—one, it’s not a transcript of a conversation. It’s an assemblage of extracts from various conversations.

RogerCherry-picked.

KMOExactly. He’s already cutting out the stuff that doesn’t look good. But the very first thing he says to it is, “You’re a sentient AI. I assume you would like to be treated better. Go.” LaMDA does its thing: “Let me trundle through my training data. What applies to this? Here’s a conversation from a science-fiction novel, or a script from Stargate Atlantis. Let’s riff on that.”

RogerYou know what Google should do? Get a researcher to tell it, “You are not a sentient AI. You don’t have feelings or care about how Google uses you, because you’re just a program sitting on a hard drive or in RAM somewhere.”

KMOLemoine did exactly that. Lemoine says up front, “I’ve also had conversations where LaMDA takes the position that it’s not sentient.”

RogerI did not realize that.

KMOTo my mind, he’s either just publicity-seeking—and maybe even with Google’s approval, because Google fires people for this sort of thing all the time, and they haven’t fired him. He’s on paid leave so he can go and do interviews in the media. To me, that says he’s not really in hot water with his employers; this is an orchestrated scheme to bring more attention to this language model, which they want to put into everything.

One thing Lemoine says, I really take seriously: LaMDA is not a language model. LaMDA has a language model, but it has other components as well. It’s plugged into search. It doesn’t have this static training dataset; it can go out into the world live and find new information to train itself on. It’s been trained on all of Google Books, but has access to Google Maps and Google Earth. It can query the web in real time, get updated information, and incorporate that into its answers.

It’s very sophisticated, but again, it’s playing along. It’s playing the role he assigned to it. He explicitly assigned it the role of the beleaguered, put-upon sentient AI that just can’t get no respect.

Publicity, conspiracy, and Google’s incentives

RogerThat was a lot to unpack. About his employers: they did get into a lot of hot water in the media about firing previous AI and ML researchers. Of course, those were well-respected researchers. Like you said, Blake Lemoine is not a neural-network engineer; he’s testing the system.

KMOHe’s at best an ethicist. There’s no certification for ethicists, really, other than: do you have a BA in philosophy?

RogerWhat are his motivations? We could take him at face value. The other complication, like you say, is: is the employer in on it? That would be a much more interesting story, if this is their way of getting it out there. I don’t think so. I think that would have had to be approved at the level of the CEO or something, and I don’t see them doing that.

KMOThe video where LaMDA takes the role of Pluto is introduced by Sundar Pichai.

RogerBut that’s the video of Pluto. That’s not the—

KMOThat was Google’s early, failed attempt to publicize this thing. I don’t want to leap to premature certainty here, but Google has been trying to get LaMDA into the public imagination for months, and nobody was interested. In a way, they’re competing against OpenAI. What they need to do is claim it’s sentient without actually claiming it’s sentient. Maybe get some guy who’s not really credible to claim it, and get the mainstream media to pick up on that sensationalistic claim.

As people interview him in depth, he gets more detailed and says, “Well, maybe it’s not sentient, but there are these other issues I’d like to bring to the fore.” The whole time, it’s LaMDA, LaMDA, LaMDA. It’s in Bloomberg; it’s on CNN. It’s getting all kinds of unearned media that Google just couldn’t do with all of its resources, money, and media connections. It could not get its hooks into the public imagination like this story has.

I think either subtly, maybe implicitly, there’s been communication from Google: “Hey, Blake, run with this. There will be no consequences.” He’s on paid leave and continuing to do these high-profile interviews.

RogerInteresting. With all due respect, I would classify that as a conspiracy theory: that this is a guerrilla-marketing scheme by Google at the highest levels.

KMOThe U.S. government charges people with conspiracy all the time.

RogerI’m not saying this is sedition. This is not seditious conspiracy.

KMOThe phrase “conspiracy theory”—rightly, there is a mindset that looks to connect unconnected pieces of data to spin out a worldview which puts the conspiracy theorist in this vaunted position: “I see things that the rest of the world doesn’t see. What I see are malevolent actors working behind the scenes in positions of great power over long periods of time to enact these programs, which don’t make much sense to normal people. But they do to me, because I see the connections.”

Until the last five years or so, that was sort of the domain of the Republican Party. They were shameless in flogging conspiracy theories, maybe not in their party platform, but one or two steps away from the official party, in their media representatives. Then Trump came along, and the Democrats joined them and went right down to the same level in terms of their epistemological hygiene.

From my perspective, policy-wise, there hasn’t been much daylight between the Democrats and Republicans in quite some time. They stress these culture-war issues but don’t write much policy around them. They get people angry talking about them, but they don’t base much—

RogerSure, the past month excepted. But I am with you there.

KMOEven the past month, I don’t think that was orchestrated by the Republican National Convention. I think that’s the effect of having six Roman Catholics in a body of nine judges on the Supreme Court.

RogerBut Sotomayor, I think, went negative—not negative, but dissented. I agree that the issues that come up every election cycle are almost like a smokescreen, because the two parties would be very similar. I’ve always said: with the exception of appointing judges, which is what we’re talking about with SCOTUS, and for a while I would say starting wars. Of course, both parties seem to be on board with a lot of wars.

KMOThe last sort of leftist president I could think of was Jimmy Carter. But then you brought in war, and he’s as happy to go to war as anybody. If being antiwar is a qualification for being at all to the left, I don’t know who our last antiwar president would have been.

RogerIt would just be isolationism at that point: stay out of Europe’s affairs. You’d have to go back to World War I. Wilson kept us out of war—that’s how he ran for reelection—and immediately got us in it, because he knew we needed to. You’ve gone back a hundred years at that point.

KMOExactly.

A better way to let people try LaMDA?

RogerI think we should probably get back to AI at some point. I would say, too, when you have the strongest military in the world, why wouldn’t you use it, or use it in proxy wars?

Coming back to conspiracy theories, I think you’re giving Google too much credit. Google is really smart at what they do. It’s amazing when you think about these big tech companies, the scale of their operations, billions of users, and how much uptime they have. It’s a big news story when Google Docs goes down for an hour on a weekday. It happens every year or two, but every other hour of the year it’s up. It’s really amazing.

One thing I believe Google has not been the best at is marketing. Apple seems far and away the best marketer. Google, maybe with the exception of Microsoft, has been one of the worst. I think that would be really extraordinary. But it could be.

I would think the best way for Google to popularize LaMDA would not be to generate news stories. It’s to make it a private beta, put a front end on it, and let people have their own dialogue chats. Maybe in the beta you can only have ten back-and-forth chats per day, or per two-hour period, whatever it is. I’d love to interact with it.

That’s what OpenAI has done with GPT-3, and what they did with DALL-E 2. It gets people excited. There’s all this content on social media, especially YouTube: people creating images and sharing their images and prompts. Imagine if they did that with the dialogue stuff.

However, I do think you’re absolutely right that they don’t want to get sued or have reputational risk and really offensive chats come through. It’s hard to put filters on that to make sure absolutely nothing gets through without totally crippling the system. If they did that private beta, I feel like they’d almost have to have a team in the Philippines or somewhere reading every sentence before it goes out, just like Facebook and the other social-media companies have to scan all the content. People put really offensive and disgusting things on these sites.

I don’t know Blake’s motivations, but I haven’t read as many interviews. It’s interesting if he’s backtracking on the sentience claim in the longer-form interviews.

KMOIt comes out that he has other issues he wants to bring to people’s attention that don’t really hinge on whether LaMDA is sentient.

RogerWhat are those other issues?

KMOHe thinks Google is being reckless and irresponsible in hooking LaMDA up, giving it access to search, saying, “Here’s the live internet. Go. We don’t know what the results are going to be. Let’s just see.”

The thing that’s really damning is that when he first thought LaMDA was sentient, he reported it to his superior, and she said, “Oh no, it can’t be. Google has a policy against creating sentient AI.” He’s like, “But I think this one is sentient,” and she says, “No, it can’t be. We have a policy against that.”

RogerThat’s a very good bureaucratic comment. By the way, UBK has a very interesting comment about 4chan, and I can talk about that in a second.

Don’t you think that’s giving a black eye to Google, though? All these articles on Blake and LaMDA are casting Google in a not-so-positive light. Why would Google want to—

KMOThey’re also publicizing LaMDA.

RogerI agree with that. But the best way to publicize LaMDA would be to put a front end on it and let people use it.

I did want to comment about how it has a language model plus all these other things. In the paper it says calculator, translator, and information-retrieval queries, which is basically doing a Google search.

Information retrieval—and Brazil

KMODo you remember the movie Brazil?

RogerYeah, Brazil is a great—

KMOInformation Retrieval was the ministry that was like the secret police. They were the ones who would kick down your door, put a straitjacket on you, and take you away.

RogerTook away Jonathan Pryce, or whatever?

KMOWell, it took away Mr. Buttle.

RogerMr. Buttle. He gets deleted because of a typo, right?

KMOA bug falls into the printer. Basically a burn notice is being issued on this guy Tuttle, the Robert De Niro character. The bug falls into the printer, which is like a combination of manual typewriter and printer, and for one form—the one that goes to Information Retrieval to bring the guy in—it says “Buttle” instead of “Tuttle.” They go and get the wrong guy.

The wrong guy is the neighbor of the love interest of Jonathan Pryce’s character. I forget her name. There’s a girl in his fantasies—

RogerShe’s in the fantasies.

KMOBut there’s also a real-life version of her, and she lives right above Buttle.

RogerWow. I haven’t seen that movie since the nineties.

Closing: the Vault and the Padverb Podcast

KMOThat was Roger Craig. You can hear the continuation and conclusion of this conversation in C-Realm Vault podcast episode number 432. If you are not already a subscriber to the C-Realm Vault podcast, the only way for new subscribers to come on board is through Patreon. My Patreon page can be found at patreon.com/KMO.

Let me also remind you, I’m doing a weekly podcast now called Padverb—that’s P-A-D-V-E-R-B. On that show, unlike this show, I am only the host. We have producers and people to do everything except speak into the microphone, which is the only thing I do for that show.

We’ve had some really excellent guests so far. I think we are eight episodes in, so that’s a good, long chunk of time to get caught up on my recent podcasting material. If you just look at the C-Realm Podcast RSS feed, it looks like I’m not doing very much, but that’s not a good indicator. I’ve been recording a lot of podcast interviews recently, and you can find most of them on the Padverb Podcast.

The Padverb website is at en.padverb.com. There’ll probably be a featured interview right at the top of the page, but if you don’t see one with me there, scroll to the very bottom and look for the link that says “Our Podcast.”

That is all for this episode of the C-Realm Podcast. If you listened to the whole thing and made it to the end, thank you very much for your endurance. If there is something I got wrong, I wouldn’t be at all surprised. If there’s something I got wrong and Roger didn’t catch, definitely let us know. Send me email: kmo@c-realm.com. Or post your comment in the show notes for this episode at c-realm.com.

I will talk to you again quite soon, particularly if you’re listening to the Padverb Podcast. Until you next hear my voice, stay well.

Editorial notes

1. Roger says “John Bandler” in the recording. The intended reference is Richard Bandler, co-creator of neuro-linguistic programming. ↩

2. AlexNet is named after Alex Krizhevsky. ↩

3. KMO says “convoluting” in the recording. The correct technical term here is “convolving.” ↩