AI & Human Flourishing | The Machine

Stump:

Welcome to Language of God. I’m Jim Stump. 

Hoogerwerf:

And I’m Colin Hoogerwerf.

Stump: 

In our first episode we explored what AI is doing to people. We started with that question because we think it’s the one that is often at the heart of people’s anxiety about the technology and lies underneath a lot of the questions about ethics and how to use it. If you haven’t heard that one yet, don’t worry, we won’t spoil anything here and you can still listen to that one before we get to our final episode, which will take what we’ve learned both about the human and about the machine, and see what kind of theological principles might come about. 

Hoogerwerf:

Which means today, we’re going to look more closely at the machine. 

Stump: 

If we’re going to start somewhere, it probably makes sense to go back to the 1950’s. 

Chen:

So there’s a paper right that that I feel like everyone who uses computers should read.

Fain: 

I’ll actually read you a little bit from Turing’s  paper on computing, Machinery and Intelligence.

Hoogerwerf:

Before hearing that, let’s introduce a couple of guests. First, Sherol, who we met just briefly at the end of the last episode.

Chen: 

I’m Sherol Chen. I’ve been in research and development of AI for the last 20 years.

Stump: 

And also here to help us with understanding the technical side is Brandon. 

Fain: 

I am Brandon Fain, an assistant professor of the practice here at Duke University in computer science.

Stump: 

Sherol and Brandon are going to help us understand the technical side of AI, starting with a bit of history as to how we got where we are.

Hoogerwerf: 

And as they mentioned, Alan Turing’s paper, Computing machinery and intelligence, which was published in 1950, is a good place to start. Here’s Brandon reading from that Turing paper:

Fain: 

He wrote: In the process of trying to imitate an adult human mind, we are bound to think a good deal about the process which has brought it to the state that it is in. Instead of trying to produce a program to simulate the adult mind. Why not rather, try to produce one which simulates the child? The idea of a learning machine may appear paradoxical to some readers. How can the rules of operation of the machine change? An important feature of a learning machine is that its teacher will often be very largely ignorant of quite what is going on inside. The view that the machine can only do what we know how to order it to do appears strange in the face of this intelligent behavior presumably consists in a departure from the completely disciplined behavior involved in computation.

Stump: 

This idea that machines might learn in the way a human child learns was ahead of its time for a couple of reasons. One of them was due to the state of computing technology and hardware at the time. The other was because it competed with a different philosophical approach, which is that the way to build intelligent machines was to program them with rules. 

Schuurman: 

You could think of the early days in AI as sort of attempting to explain things symbolically or knowledge based systems where you would enter all the information or define it precisely.

Hoogerwerf: 

This is Derek Schuurman from Calvin University.

Schuurman: 

Knowledge based systems where you would enter all the information or define it precisely.

Chen: 

It felt like rules were the way to go. Of course, we need to map everything out for the machines to follow. 

Fain: 

many of the early researchers in AI, in the 60s through the 90s sort of largely assumed that whatever true artificial intelligence would look like, the medium of that intelligence needed to be logic, very often, first order logic, quite specifically symbolic logic

Stump: 

I used to teach Symbolic Logic. It was one of my favorite philosophy classes and part of the standard philosophy curriculum for students. It’s developed in the early 20th century by people like Bertrand Russeld. And the goal of symbolic logic was to take all of natural language and convert it into a perfectly precise artificial language that was governed by rules. It turned out not to work entirely. But was hugely influential for the philosophies known as logical empiricism and logical positivism. And those were still in vogue by the middle of the 20th century when this thinking on artificial intelligence was just beginning. 

Hoogerwerf: 

At the point when Turing wrote that paper in 1950, the term artificial intelligence wasn’t yet used. 

Fain: 

The British called it in the 1950s machine intelligence. And artificial intelligence was coined by John McCarthy in the 50s. And the usual, famous opening document is his proposal for funding for a Dartmouth summer workshop where this term is coined and where he sort of defines that. 

Hoogerwerf:

So John McCarthy essentially makes up this term in a grant proposal. Well he gets the funding and ends up gathering a bunch of leading scientists together for a many weeks long workshop. 

Stump: 

So out of this meeting comes this new term—Artificial intelligence. But the definition of that term isn’t completely clear, even today. 

Schuurman: 

There’s a lot of competing definitions actually.

Fain:

My favorite definition of AI, in the broadest sense today is a very fluid and under specified one, and that’s for a reason. So my favorite definition is anything that a computing machine does that if it was done by a human, would be considered intelligent, would be considered intelligent behavior, whatever that is, we’ll call that AI.

Schuurman:

The sort of seminal textbook in computer science on AI by Russell and Norvig has, I think, a matrix of like four different definitions that look at different angles, but it’s basically getting machines to do things that would require intelligence done by people.

Stump: 

Well, you might have some questions about this definition. Math seems like something that requires intelligence and yet we don’t call calculators AI today. 

Fain:

The point of this is that historically, the nature of artificial intelligence is a moving target.

Stump: 

In the early days of computer science, machines were really just starting to do things that before only people were able to do. And one of those things, that happened while World War II was going on, was code breaking. 

Chen: 

And before they would have like you know rooms of people and analysts and mathematicians trying to calculate and coordinate these things and these codes were cracked by mathematicians, human mathematicians. All of a sudden they had a machine that was able to break these codes, and that you know that meant that there was a more successful outcome for these people. And so you know, even when we think about like calculators and slide rules, abacuses, these were all tools and devices that helped make decisions that humans previously had to make.

Stump: 

From the 1950’s through the 1990’s a lot continued to happen in the research and development of artificial intelligence, but for the most part, it didn’t make it to the public. Instead this time is marked by the advent of other technologies that are related but not a direct output of machine learning research…the internet, email and search algorithms. 

Chen: 

So this is where we have the tech companies, the search engines, the internet created this whole brand new ecosystem for information and data and things that can be electronically transferred. Right, email was taking off in the 90s. And so what we were calling like the mention of like the combination of statistics and data being AI, that was actually happening, but it was called big data at the time.

Hoogerwerf:

In the 1990’s and into the early 2000’s a shift starts to happen in AI research, which is built on all that came before it and is spurred along by the increased computing power of machines, but there also were some shifts in the ideas behind how an intelligent machine could be built. 

Fain: 

There have been several paradigm shifts, but one of the most important is the move from a more symbolic—sometimes it’s called GOFI or good old fashioned artificial intelligence approach—to a more modern approach, which tends to be more statistical and algorithmic and orientation.

Hoogerwerf:

One way to make this shift clear is to talk about one of the early ideas for an application an intelligent machine might have. 

Fain: 

In 1960 if you went and asked whether boosters or knockers—skeptics, that is—what if a computer could do would you say, well, that’s clearly AI. And you know, one of the most common answers was if it could play chess. 

Stump: 

In the 1950’s people were starting to theorize a chess playing machine. Alan Turing even wrote a program that could play a whole game of chess, but there wasn’t any machine at the time capable of running the program. As the technology developed these programs got better and better, to the point where they could play full games, but they were still built to follow rules. Eventually as the computing power grew, the machines could play out vast amounts of scenarios and pick the best one. Finally, in 1997, a computer program beat the reigning world champion chess player, Garry Kasparov. 

Fain:

People today, like my students, have grown up their entire lives knowing that if they ever win a game of chess against the computer, it’s because it let them.

Hoogerwerf:

That 1997 program that best Kasparov was called Deep Blue and it sits at an interesting moment in our historical exploration, when computing power and data were powerful enough for a program to be able to sort through huge amounts of data very quickly. But Deep Blue was still a program that was built with a set of rules, and its main advantage over a human was its ability to search through future possible combinations. But what Deep Blue never did was learn from any move it ever made. 

Wenger:

There’s always been somewhat competing camps on do we teach the machine, or let the machines, you know, kind of learn, or do we tell them how to learn?

Hoogerwerf: 

This is Emily Wenger joining our chorus of technology experts. She’s a professor in the Department of Electrical and Computer Engineering at Duke University. 

Wenger:

And it seemed that perhaps in 2014 there was a confluence of well, this approach, if we let them learn, and now we have fast hardware, seems to work pretty well.

Chen: 

And then we get into 2015, which is I would say around 2014 was when the waves were breaking for machine learning and deep learning.

Schuurman:

Rather than completely defining a knowledge area, have the machine learn itself.

Hoogerwerf:

A lot of this development came about when it became clear that there might be some really practical uses.

Chen: 

So we were finding ways of making these calculations faster, designing the hardware, designing chips, so that the computations could happen at like this much more rapid rate. So we didn’t have to wait, you know, a month for a computation to finish. It could happen within a couple minutes. And so with these three advancements—the data, the hardware, and the and the algorithm advancements—we started to see like, oh, there are things that, you know, honestly, the early advancements were very practical things. So all of a sudden, you know, you didn’t have somebody having to interpret the handwriting of the postcard of the address of the recipient. 

Stump:

One of the milestones that happened in the early 2010s came out of the work of trying to get computers to be able to classify images. 

Schuurman:

Your listeners can’t see this, but I’m holding a cup. So how does a computer vision system recognize what a cup is? Right? Well, it has an ear. It’s kind of cylindrical, but sometimes it doesn’t have an ear, and it has a variety of different colors, and it can even be shaped. So how would you define all the possibilities for a cup? Well, the alternative is just to show, you, know, 1000s or millions of images of cups, and then it determines statistically what constitutes a cup. What is cupness is kind of what you’re trying to capture statistically. So machine learning sort of leans in that direction. 

Hoogerwerf: 

We all know what a cup is when we see one. But actually coming up with the rules that define a cup gets harder. My friends and I have a similar long running debate about what counts as a boat. A floating log? Probably not. But a floating log with a hole cut into it all of a sudden becomes pretty close to a canoe. The point is that there are always tricky edge cases and so telling a computer what a cup or a boat is isn’t as easy as it might seem. One of the early success stories in doing this was called Alexnet and was a program that attempted to classify lots of images into categories, and many of those categories happened to be dog breeds.

Chen:

Can a machine figure out if this is a puppy or a muffin? If you just Google puppy or muffin, you’ll see these cute Chihuahuas that kind of look like muffins, or like these mop dogs. And some of these mop dogs really do look like mops, and so it is kind of an obscure, fun little like emergent, like kind of a colloquial sort of like this little moment of like fascinating ourselves with that question of like what is what does it mean to look like a mop dog versus being an actual mop. So these things emerge, and what happened, I guess eventually was just that we saw the, you know, machines and some sometimes were just able to do this quite effectively and at a speed that humans would take maybe a little bit longer, and so that probably happened around 2016, 2017.

Stump:

In our story of chess, after the 1997 defeat and some even more powerful rule based programs, computer science pretty much lost interest in chess. It was a problem that had essentially been solved using the old rule based approach. But in 2017 a new program was created called AlphaZero. And instead of giving the program any guidance on what was a good move, the programmers gave the rules of chess and had it start playing games. No direction at all. And just like they found with image classification, it turned out this was a really effective way at getting a machine to play chess. Today, a program like AlphaZero completely crushes a program like Deep Blue, even though it has never been told what is a good move and what is a bad move. 

Hoogerwerf:

That brings us up to 2022. 

Wenger:

And then ChatGPT was released, and it took on a commercial form that people could really latch onto.

Chen: 

So the big you know LLM ChatGPT—What is it? ChatGPT is a large language model, and that is a type of model that is called a sequence model. 

Stump:

In 2022 chatGPT is launched to the world. For free by the way. And for the most part, this historical conversation we’ve just had is mostly unknown to the people that use it. 

Chen:

it did feel kind of like people saw opportunities to to persuade and to create new narratives to buy people’s attention, and that was worth something. And people did that to a great amount

Hoogerwerf: 

From Sherol’s perspective though, the advent of LLM’s is just another step along a history of development, that surely had big steps, but wasn’t the same kind of surprise that it might have been for many of the initial users. 

Chen:

Like it’s like, oh, what changed in 2022? You know, not very much changed from my point of view. Except all of a sudden, everyone started talking about large language models.

[musical interlude]

Stump: 

We’re making this episode in 2026, 4 years into the widespread use of large language models. That means that we’ve all experienced the phenomenon of LLM hallucinations and sycophancy and the period where there was a strange inability for LLMs to be able to accurately make pictures of human hands and fingers. And we’ve been at it long enough that many of us have built routines around using LLMs for work and in our daily lives. But I think it’s fair to say we do all of this without knowing how it really works. 

Hoogerwerf: 

I guess in some ways that’s true of a lot of technology. I can’t really tell you how my iphone works either. Or even really how a radio works, except that when I turn it on, sound comes through it and that somewhere inside , I think there is a big coil of wires. 

Stump: 

The difference, I think, is that those technologies aren’t doing anything that would have you think they aren’t just machines doing a task, whereas the output of an LLM can appear to be something much more than just a machine that works in ways that you don’t understand. In this case, it can sure feel like a human is on the other end. And if not a human, at least a mind. 

Hoogerwerf: 

Yeah, there seems something really mysterious about what an LLM does. But most of the technology experts we talked to seemed much less mystified by what is behind the technology. 

Chen: 

For me it had been like this very common day-to-day thing. Like I’ve just been around it; it’s not new to me. It’s not. It’s just I’ve seen it grow over the years. It’s not this big jump for me.

Fain: 

Modern LLMs are purely sort of numerical models trained on statistical data, fine tuned in various procedures that we can talk about if you’re interested in order to exhibit certain kinds of behavior.

Schuurman: 

Yeah, the machines of today, I’m certain, are not conscious 

Stump:

To be sure, there are experts out there who think differently and the media tends to really like the stories about technologists who are convinced they’ve found consciousness in their machine. I’m not sure I’m ready here to tell you once and for all what constitutes consciousness and how we’ll know it when we see it. I’m pretty confident that human consciousness is a function of being alive, and computers are not living creatures; they are machines. But I’m less confident that the only possible kind of consciousness is human consciousness, and even if these do a remarkable job of mimicking the outputs of human consciousness, I think we should be open to asking whether there is another kind of consciousness possible that does emerge from machines.

Hoogerwerf:

So we know from our historical exploration that LLMs came out of the shift away from rule-based methods when developers started having a lot better success using statistical techniques…

Fain: 

including very much all modern LLMs that involve no formal symbolic processing.

Stump: 

Brandon often notices that students are a little confused by this because the questions they ask show some assumptions they have about the technology. 

Fain:  

Implicitly, they’re assuming that somewhere deep down in there there’s like somebody diagramming sentence grammar, and there isn’t, nowhere in there, in the same way that a human child at age five learns to speak a language way before they know anything about the rules of how language work, with absolutely no conception of that. There’s nowhere deep down where there’s logical reasoning, nowhere where there’s structural diagramming of sentence with a computational model of linguistics in a formal sense, people tried that for decades in natural language processing. They tried very hard, extremely, intelligent people, world class linguists and computer science scientists working together, it did not result in empirically useful systems.

Hoogerwerf: 

Ok, so how then does it work? 

Stump: 

We’re going to get a little technical here but we think it will help us with some of our bigger philosophical and theological questions we want to ask later. 

Hoogerwerf: 

And the whole thing starts with a huge amount of data. 

Wenger:

So the AI trainers hoover up a lot of data, typically the whole internet.

Hoogerwerf: 

All of that data goes through a training process…

Wenger:

So it’s sort of this push-pull thing. You put the data through, it predicts. You say no, no, no, this way through this gradient process, and then through the push and pull of seeing many different types of data, the model is pushed down what we hope to be a slope towards the best set of parameters that approximate the distribution we care about.

Hoogerwerf: 

Essentially, through this training process, the model is just starting to see how language is used and it starts to build a giant database that shows how all the words it sees relate to each other. Which brings you to a point where you can give the model some text…

Fain:

Well, so you give some input, right to an LLM—

Stump:

Let’s—for example’s sake—say our input is “what does it mean to be human?”

Fain:

—your input is going to be processed in the following way. It’s first going to be tokenized.

Stump:

This is a fairly straightforward process.

Fain:

The best analogy would be to think of it’s a token is roughly a word.

Hoogerwerf:

It’s actually a little more complicated than that because it will break up prefixes or stems of words and punctuation, but thinking about a token as essentially a word will do for us for now. And each model will have, essentially, a dictionary of tokens that it can look up for your prompt.

Stump:

ChatGPT, as we’re making this episode, has about 200,000 tokens in its vocabulary. And so your prompt will be turned into a list of numbers.

Fain:

like the first token is number 3000 in my vocabulary, the second token is 2107 in my vocabulary, and so on and so forth. Nothing particularly deep there

Hoogerwerf:

So I was curious and wondered if I could actually find out the token numbers for our prompt. Using ChatGPT I was able to download the token library and find all the tokens for that prompt. So “What does it mean to be human” is turned into “4827, 2226, 480, 4774, 316, 413, 5396, 30.”

Stump:

Hmm. Got it. But this is still all just set up. It doesn’t tell us how it answers the question in a convincing way.

Hoogerwerf:

No, Not yet, there are still a few more steps. Hang with us. The next step is to feed those tokens into the model.

Fain:

Each token is going to be transformed into, well, into a numerical vector.

Hoogerwerf: 

A numerical vector is just a series of numbers. So initially the word “human” has a string of numbers associated with it that could be 100 or 1000 numbers long, depending on the model. The longer that string of numbers—the more dimensions the model has and the more expressive it can be.

Stump:

And those numbers in the vector mean something?

Hoogerwerf:

Well, to the model it means something, but nothing you could figure out.

Wenger:

You can try and map those features back to something the human eye sees or understands, but it’s typically quite difficult because the model is operating in mathematics world. It’s using you know its number and numerical representation of the world to process the information it sees, and the human brain just doesn’t work on that level

Hoogerwerf: 

So it’s not like the first number in the vector for human is “mammal” and the second is “bidepal”. The numbers are just a mathematical way the model has learned, through training, to relate this word to all the other words. 

Stump: 

So I guess you could think about this a little bit like GPS coordinates. Your house has coordinates, but the numbers themselves don’t mean anything on their own. They only start to mean something in relationship to other numbers on a two dimensional map. So these numbers attached to the word human help the model to know how that word relates to other words in this mathematical space. In our prompt, the word human, might show up near the word “person” with the map of the model. But to capture the complexity of how words are related to each other, add a thousand more dimensions. 

Hoogerwerf:

Right so you have that vector for each word in your prompt and so all together you have a matrix…a row for each word and a column for each of the numbers in the vector however long that is. Which ends of being an unbelievable amount of information about all these words, but without any context.

Fain: 

How should we think about the meaning of this word as it relates to the other tokens appearing in the sequence?

Hoogerwerf:

That first set of numbers is just what the model learned from training about those words, but it needs to know what the tokens should be in the context of my prompt.

Stump:

So “it” in “what does it mean to be human” doesn’t mean much without any context.

Hoogerwerf:

Right, and order matters too. “What human does it mean to be?” would be a really different question. So that first matrix of numbers gets transformed into a new matrix of numbers that takes context into account. One stage of this calculation is called a transformer block.

Fain: 

each block is going to be composed largely of an attention mechanism followed by a feed forward, fully connected multi layer perceptron component

Stump:

What?!

Hoogerwerf: 

[chuckles] Yeah, I just liked the sound of that. But it’s not as complicated as it sounds. Or actually maybe it is, but at least it’s not important that we understand it to this depth. What we can say is that attention is just the process of determining  which tokens are important in context and then the model calculates a new set of numbers. It does this a bunch of times. 

Stump: 

And we still don’t have any response to our prompt yet. 

Hoogerwerf: 

No but we’ve done everything we need now to predict what the next token should be. 

Fain: 

So that first word of the response after you’ve passed everything through the transformer, for scale, like a chat GPT level model might have a few 100 billion parameters you pass through that whole model to get the first word of the response. Just to generate the first word of a response when you query chatGPT, approximately four to $600,000 of hardware has to fire to run all of that.

Stump: 

So that gets us to, “what does it mean to be human?—People”

Hoogerwerf: 

Yeah, then it has to do the whole thing all over again. And this time it includes the “People” 

Fain:

That first word of the response token, really roughly a word is going to be appended to the end of your input and whatever else was in the system prompt or whatever, and that all of that together is going to be fed back into the model at the beginning with just the change that you’ve appended the first word of its generation to the end. All of that passes through the model again generates scores for likely next tokens you’ll sample from that distribution. That’ll be the second word of the response that gets appended to the end, and the whole thing gets passed through again, and on and on and on and on.

Stump:

So let me take a stab at summarizing this: We put a question into the prompt box. The LLM converts our words to its tokens. And it has a giant map of how all 200,000 of its tokens are related to each other based on the examples of human language it’s been trained on. And then for each token, it creates a vector   in this 1000-dimensional space to generate the probabilities of the next best word. And there is some randomness or variability in this, so it’s selecting a word from a set of possible tokens. Once that one is determined, it starts the process over again to pick the next word. And this continues until it has a complete sentence or group of sentences that are the length of answer it was expected to give. How’s that?

Hoogerwerf: 

For as complicated as that is, it’s probably an over simplified explanation and I’m sure a lot of our experts would pick at parts of it. I’m not sure we need to understand it even to the depth that we’ve gone to here, but we do this partly to make a point, which is that there isn’t anything magical happening here. Once we have this massive amount of data about how human language typically works, these machines with all their computing power are able to reproduce very plausible sounding sentences. Like the answer to our prompt, what does it mean to be human:—People have been asking this question for about as long as we have any records and there are no easy answers. We could answer the question biologically, exploring the long history of the species homo sapiens or we could answer the question more philosophically, 

Stump: 

Well now I’m starting to feel like the answer to “what does it mean to be this particular human” is that I need a nap. 

Hoogerwerf: 

I can see how it’s possible to get lost in the details of all this, but for me, after talking to Brandon and doing the work of trying to understand how this process works, I think I have had a really different perspective when using a large language model. Before I knew what was going on it felt like this mysterious being inside my computer that was talking back to me. But understanding some aspect of how it determines the next token helps me to understand better what it is—and what it isn’t. 

Stump: 

It also explains something about why it’s not very good at some things. Like I’ve always been surprised at why it has never been very good at word counts. But understanding what’s happening underneath clears that up a bit. 

Fain: 

So the whole process I just described when trying to count or solve an arithmetic puzzle or something the model’s doing it through the processing I just described, trying to predict token at our generation, token at a time through the attention mechanism, whereas Microsoft Word or Google Docs, when it’s counting the number of words, is literally just looking for white spaces and then running a for loop to count up by one after each of these breaks.

Hoogerwerf:

While that does help me better understand the limits of what a large language model is good at,  it still leaves me with a question: I understand this process that’s been described as simply picking the next best word (or token)…but it seems like ChatGPT is doing more than that when it answers my questions in a way that is true. We play the game Balderdash with my family where we all make up definitions to words. The winner is the one who put together the best sounding answer—or maybe the funniest—but not the right one. It seems like the process that’s been described about next token generation would be good at playing balderdash but wouldn’t be as good at winning a trivia contest…and yet it would probably beat any human at a trivia contest. 

Is there something baked into language that also leads to true answers?

Stump:

Normally, for us there’s the assumption that for our language to be true, it has to correspond to the way things are in the real world. But if an LLM is simply trying to produce the next best token, then it’s been trained on all of this language that is on the internet and presumably enough of that language on the internet itself is true, that when an LLM tries to produce the nexr best words it ends up saying true things, most of the time. 

Hoogerwerf:

But there is actually more to it than that. Because the modern tools like ChatGPT and Claude are using an LLM but they are also bringing in other computing tools.

Wenger: 

The next token prediction capability of these models is now so wrapped in all these other software pieces, like web search and content retrieval and things like that. So, all of those pieces are coming together to inform how the model produces tokens. 

Fain:

It’s not just running your prompt directly into a model. What it’s doing is running a fairly traditional Google search for your query, identifying top hit articles and things like this, extracting chunks of text that seem relevant to your query from those search results, and then feeding all of those chunks of text right into the prompt of the model along with your question.

Stump: 

And you can start to see where this could go, eventually ending at a system that we might call Artificial General Intelligence, or AGI, where AI is not only able to use language really well but can do lots of other tasks.

Fain: 

So classically, artificial intelligence tools, machine learning models that we have built, and things like that have generally been for a particular purpose or application. And when people talk about AGI, they mean some artificial intelligence system which does not have a particular narrow purpose, but rather is intended to do any number of things who knows what. 

Stump: 

There’s no committee or anything that would decide when AI would become AGI . Already a lot of things we used to call artificial intelligence we don’t anymore. Chess programs being an obvious example. We have other AI programs that have narrow uses. Something like facial recognition for a smart phone. But large language models that now wrap in a bunch of other computing tools start to get closer to something that seems like it could be AGI. Still a lot of people are thinking even beyond that.  

Fain:

They’re thinking like something that surpassed human intelligence in every sense and regard, all simultaneously, or something like that. Sometimes people also refer to this as like strong AI or something. If that’s what you have in mind for AGI, then it is much less clear whether we have or will arrive at anything like that, or whether that’s desirable.

Hoogerwerf: 

That probably starts to make people anxious because with that capability you start to wonder what could go wrong and how much control developers have. There’s this story I sometimes hear which is that the people making these things don’t actually know what’s happening and are completely surprised by the results. The surprise there might be true in some sense, but clearly, from listening to Brandon, the experts aren’t—like Dr. Frankenstein or something—totally unaware of what might result from their experiments. 

Fain:

Do experts have no idea what is going on? No, that’s not true. It’s kind of hyperbolic exaggeration. 

Schuurman:

As a computer scientist, you know, I have, I’ve been trained to be able to look through the covers of the machine and right down to the sort of bare metal and see how it works. And it’s demystified.

Stump: 

But while they understand how the system is built and the process it goes through, these are incredibly complex systems, so understanding is different than being able to trace everything exactly.

Fain:

Are you stepping through the reasoning of the model step by step? Well, no

Schuurman:

You know, once you have trillions of multiply accumulates, you know, swirling around in a machine. You know, trillions of nodes being used in a, in a in an artificial neural network, then kind of knowing what it’s going to do is sort of more than you can, you know, anticipate, you know, just because of our cognitive limitations, but that that isn’t an ontological thing. It’s simply a matter of the complexity.

Stump: 

And so the surprise comes more from how well these systems can work and what they are able to do, rather than the surprise of something emerging which was completely unanticipated. 

Fain:

Or a broader sense of what we don’t understand, of course, very fundamentally, is the societal impact of putting these models into the world. So we know how we train the given LLM, but Okay, now we’re hosting it, and people are using it for who knows what? What impact is that having? How socially beneficial or harmful is that? 

Hoogerwerf: 

That, is of course, the question that we started exploring in our last episode, which brings us full circle. But it leaves us with a final question, which is what are we supposed to do now? 

Schuurman:

AI is going to be capable of doing a lot of things, and we just have to decide if, if, if that’s how we want to use it. 

Stump:  

That’s what we’ll try to figure out in our next and final episode. 

Language of God is produced by BioLogos. BioLogos is supported by individual donors and listeners like you. If you’d like to help keep this conversation going on the podcast and elsewhere you can find ways to contribute at biologos.org. You’ll find lots of other great resources on science and faith there as well. 

Language of God is produced and mixed by Colin Hoogerwerf. That’s me. Our theme song is by Breakmaster Cylinder. 

If you’d like to get in touch we are always happy to hear your feedback, questions, or ideas for future episodes and conversations. Email us at podcast@biologos.org.

Thanks for listening.