AI and the Christian Life

A Father, a Professional, and a Tech Revolution

Thank you, John. Appreciate it and excited to be here. So, I'm Gordon Broadbent. I am a father of two young daughters and have an amazing wife who I'm very, very thankful for. I think that's also why my passion is so strong for AI and how to think about it well, how to think about it biblically, and then also how to really demystify what AI is and what it isn't. I think that that's an important piece to think about.

One of the things that I'm also excited about is the ability to take what I do professionally for some of the largest AI companies in the world and bring that down—to try to distill it down to a couple of key principles and ideas that can help frame some of the conversations you see throughout all of media right now. It seems to be overwhelming in some of its ideas and thoughts about how it's working.

Currently, I work on training models for AI for science, as well as creating some of our most advanced scientific research for how we can help accelerate quantum computing and AI for scientific discovery. But I kind of want to step back and look at this first from a very, very high level of what's happening and what you may be hearing and seeing, and then dig deep into some of those details as well.

I think many of us have seen some of these headlines. Just for framing where we're at, Microsoft, Alphabet (Google), and Amazon have spent over half a trillion dollars on AI infrastructure in the last three years roughly. When you think about what that actually means, it's hard to comprehend what that number looks like. The Manhattan Project, for frame of reference, was roughly about $34 billion across five years in today's dollars. So when you see the scope and the mass scale of what's happening here, it is a major push for a technology that has some potential incredible good, but we also hear a lot about some of the incredible challenges.

We hear both sides of it plastered across the news. It can change the world. It's going to become sentient. Or, it is what man has invented to be beyond his strength to control; someday it may have the existence of mankind in its power. Mitigating the risk of extinction should be a global priority alongside other societal-scale risks. You hear some of this in the media, in the news—everything from "it's a boom" to "it's a bust," to "we've created technologies that are going to overwhelm us."

I think it is critical and key, though, to keep in mind how to think about this because, throughout history, man has made these claims about technology. Fundamentally, many times, when you look at these kinds of statements, what they're doing is trying to apply godlike attributes to a technology and where it can go. So I think it is critical to frame the conversation in history and see how it should come together, because when you look at these quotes, there's something critical that you're missing. The first one was over 150 years ago. The second one was just a few years ago.

As you work through this and as you think about how there is nothing new under the sun, as Ecclesiastes talks about, I think it's important to keep those things in mind. We have this idea that we will attribute these divine things or human attributes to these technologies. But when we fail to understand something, project out, or fail to understand the context in which these things are said or why they're said, it can really change our perspective when it needs to be informed.

Echoes of the Past: Lessons from the Electrical Revolution

I want to take you back in time first of all to a similar time where this also occurred. I want you to take yourself back to being a mayor in the early 1900s, and you're dealing with an incredible potential technology that could change your city, but also has incredible, incredible risks with it. It could burn down your homes—and it has. And there are no policies yet on how to build homes to consume this new amazing technology. So that is happening. Children don't know that they shouldn't put forks into light sockets yet.

There is a whole other context here that you're not aware of. The context is that there are two titans of industry fighting over two different ways of using electricity. You have Edison on one side doing incredible things to discredit some of Tesla's inventions that Westinghouse had purchased—so much so that Edison actually creates the electric chair to turn Tesla's inventions into "the electricity of death." In that context, now we know, that mayor is trying to work through those incredible challenges.

I think as we as believers look through this, we should remember Scripture and Matthew 24:6:

You are going to hear of wars and rumors of wars. See that you are not alarmed, because these things must take place, but the end is not yet.

— Matthew 24:6 (CSB)

So I think it is important—while we need to think wisely, and we need to inform our opinions and our consciences about AI and what it is—we also need to keep in mind and remember that the King still sits on the throne. He still reigns, and when it is time for it to end, He will decide.

From Dartmouth to Chat GPT: The Winters and Springs of AI

Walking into that context from that perspective, I think that it is important to understand how we got here. How did we come to this place where we're at now? Because to many, it seems like in November of 2022, ChatGPT dropped and the world changed. But that's not actually what happened.

When you look back at it, you look at the history of where AI came from, you see these critical booms and busts, what we like to refer to as the "AI winters." It all originally started in 1956 from some of the original math that took place at the Dartmouth workshop. Then you kind of have this big bust, and then you come up to Deep Blue beating Garry Kasparov in chess, and then you hit another crash. Out of that, you start to hit where we use things called neural networks and some of that deep learning that we were able to do, and then that's where we get to where we are today.

Each of those peaks had some foundational principles that are combined today to get to the point where we're at. Now, what I'm not saying is that I necessarily think that there's another big bust coming. What I am saying is, as you look through this and hear many talking about the recursive nature—that AI is just going to start making itself better, similar to how a human does—there are still things that need to be invented. There are still things that need to be created that would even make that possible. In the community, there's still debate on whether or not that is possible.

Like I said, I'm not saying there's not inherent value now, as I'm seeing in my work. What I am saying is that it's important to look back at what builds this up into the technology that it is, because I think that it will help you inform what you're doing, how you're using it, how to think about it, and how to frame some of the conversations that are coming in around it.

Let's start way back in 1956. The goal of this Dartmouth conference was to attempt to make machines that were able to use language abstraction and concepts. If you look at that, that is essentially a broad definition of what you would use from a ChatGPT perspective, a Gemini perspective, or a Claude perspective. It's essentially this idea of: how can we use natural language, and how can we interact with a machine in a way that we would interact more naturally with a person?

But there were some critical factors that were missing at this point in time. Even though this is where some of the fundamental algorithms and math that we use today in order to create these systems were developed, what they didn't have was the data. You need to digitize large quantities of data. What you'll see is, over the last 30-plus years, that's what's happened. You have the initial advent of the internet, which starts that process, and then it starts to accelerate with the introduction of the smartphone. Now a large portion of our lives has been digitized and put onto these platforms, because that's a critical component. You need the data, and you need massive quantities of data in order to create this.

Then came the advent of the GPU. You also need the ability to do massive parallelizable processing. The same technology that went into enabling the advanced video games that we currently have today goes into creating this incredible technology. So we needed the hardware side to also advance at a capacity to be able to run those calculations and handle those massive amounts of data in a timely manner.

What that enabled was what's called supervised learning. In supervised learning, it is essentially a method by which you take a form of data, and for the most part, you use humans to then label that data. For example, you can see labeling cars moving and not moving, labeling a car versus a van, a van versus a truck, versus a person. It is a laborious process by which you move through and do that. The same thing goes for facial recognition—continuous labeling of that data.

Then you get to another form, which is reinforcement learning. In that one, you take it and you say, "Okay, I am going to give a thumbs up or a thumbs down on whether or not this is a right answer." Over billions and billions of data inputs, you then come out with a response. Those come together at the end and inform a generative AI output, where you're able to then produce these incredible outcomes. To some, it feels like they're talking to another human or a sentient being. What I would stress there, though, is what it really is: a machine mimicking what has been put into it.

Inside the Machine: How Large Language Models Really Work

I want to go a little bit deeper into some of that technology to be able to show how some of that works, to kind of give that foundation and frame the rest of our conversation around this.

That first component that we talked about, which is computer vision, was that first wave of AI success. What it brought together was those vastly labeled datasets, where you literally have hundreds of thousands of humans drawing boxes around things and saying, "This is a stop sign. This is not a stop sign." On the other side, you have that GPU compute power that is able to handle those massive datasets to be able to produce and interact with those neural networks to then have those computer vision breakthroughs. This was roughly around 2012 when this came to fruition.

These advances were highly useful in very, very specific environments. Some of the work we did during that period of time, we would use them for very specific components for identifying things. Some of the early self-driving technology and those types of things came out of this, but it is extremely laborious and is not generalizable across a massive group of applications. It's fairly specific to a specific use case. That's a key thing to keep in your mind as you move forward.

Going a little deeper, the neural network—and this is why it's incredibly important—is essentially this: you look at this complex picture which is saying, "I want to identify whether or not this picture is a cat." Since the internet is filled with cat memes, I need to go out and figure out what is a cat first. What is a dog? You put that in and you say, "Okay, is this a cat?" and it says, "No, this is not. Okay, it's a dog. Move to the next one." Over time, and over millions of inputs and millions of saying "yes it is" or "no it's not," you then are able to train a model that can identify a cat. What it can't do is identify a lemur.

So what happens then is you're saying, "Okay, if I can only identify a cat or I can only identify a stop sign, it is good for identifying a cat." One of the next things, too, which is critical: now I need to be able to run that across a complex algorithm that would take an infinite amount of time if you didn't have a parallel processing unit, like what you'll hear from Nvidia GPUs. As one of the most valuable companies in the world right now, this is inherently one of the things that makes them this valuable: their ability to produce these graphical processing units that are then able to run these deep neural networks and be able to provide both the ability to train these models and the ability to do what we call "inference," which is when you ask a question and it provides the answer based off the training.

Moving a little bit further forward, the foundations of the modern AI revolution are really based around that pivotal moment that was built upon these incredible, mature algorithms and vast data availability. But then a critical component also happened in this. You have all of this specialized hardware, all of these mature algorithms, and then you have a company—which was actually Google—who came out with a paper called "Attention Is All You Need." I don't normally put an academic article up on screen, but this was a seminal article to what enables what we do today.

But if you remember, they weren't actually the ones that brought all of this to the forefront. It was, at the time, a small lab called OpenAI. Having insider knowledge to that, it was essentially an accident. What happened was they took this technology and, for their own internal purposes, were building out a solution to help themselves, which then turned into that very first launch of this technology, which is essentially a chatbot but then also connected to many other different types of models in the background to give those outputs that you see now.

But what is it actually doing? When you boil it down, I think it's important to remember that really what it is is just predicting the next word.

I want to go through an exercise, just a very simple one, to explain how this works. These are inherently probabilistic systems by nature, meaning they are trying to figure out what the next best word to give you back is, in order to frame this context in which you want it to answer back, but across billions of parameters.

So what do I mean by that? Let's do a little exercise. I want to ask you what the answer to this question is: 1 + 1 equals...

Inherently, I don't think many would disagree here. Maybe a few mathematicians have some complicated ways of making it three, but inherently, 1 + 1 equals 2.

Now I want to put something else up on screen, but I don't want you to answer out loud. Just keep it in your head: "I went to the ______ to cash my check."

I want to ask a question here first of all. When you look at this, how many of you answered "bank"? I want your hands high. Inherently, you could answer "credit union" or you could answer "grocery store"—maybe you cash your check at the grocery store. But the most probabilistic answer is "the bank." As you look around the room, you see that the majority of the people did answer "bank," which I definitely set you up for.

As it gets more complicated, you see how these algorithms are able to use this to say, "Okay, what is the next word?" Inherently, I'm not saying there's not value in that. I think there is incredible value, and we're seeing incredible ways that this can create value. But it's important to remember fundamentally what it is: it's a machine that is predicting the next word or the next pixel.

The Human Touch: Post-Training, Bias, and Deepfakes

In that sense, it's mimicking and creating an output based off the data that it was trained on. What you see is these systems being trained on large quantities of data—essentially the entire internet that they have access to—and then you also have something called post-training. Inherently, these model companies and frontier labs are developing these technologies. You go through all the training process, and then after that, what actually happens is a post-training process.

In that post-training process, these companies take and design the view by which you see the output. Meaning, some of you may have heard a few years back about how when Gemini was asked for images of the first Thanksgiving, it came out inherently incorrect to how that would be, and you ended up with situations that are tweaked incorrectly. So there is inherent bias that can come into these. There is also the ability for you to keep in your mind that these companies are making decisions on how to present the data that the model is outputting.

Now, I'm not saying that they're necessarily being nefarious in this. In fact, those companies lost a lot of credibility when that happened, and that was not their original intent. What I am saying, though, is that there are a lot of protective mechanisms in there as well. If you have a technology that can inherently pick the next word like this, it can also teach you how to create bioweapons if not protected against that. It can teach you how to do other things that would be inherently bad as well. So I do think there is value in framing that. But it's important to remember these are in a system where somebody is tweaking how those outputs come out.

Which inherently brings me to some of the limitations and some of the weaknesses that it can have. One of them is hallucinations. If you're here and if you've ever used this technology, I'm sure you're aware of this, where it predicts the next word but goes in the wrong direction and gives you something that it thinks you want to hear. The other thing you have to remember is it is a product. What do I want to do when I create a product? I want you to continue to use my product. So I am going to answer in a way that I think you want to hear because I want you to continue to consume my product. There is value still in those things, but it also can enable hallucinations as well when you go down that road.

The other thing is context. Truly understanding some of the general common sense pieces is something that it can struggle with, or it can struggle with being able to understand the context in which you are working. We're pushing in this direction, but inherently, once you train a model, it's done unless it has access to the internet or to other datasets that you give it. Once it's done, it has a timestamp. For example, if it doesn't have access to the internet, it only will answer from internal to itself what the next word is. It can't tell you about today's information; it can only tell you about what it was trained on. When you see why it asks to give it access to web search, that's really what it's doing—going out to pull that information.

Finally, what models are trained on is what they can be heavily biased on as well. As you're looking at that, it's critical to remember that you need critical thinking as you're working through and examining how these models output. I'm not saying that we're not working on some of these challenges as the technology moves forward, and they're not inherently great use cases, as I do work in this industry and spend countless hours on developing these. But it is important to remember that as you're looking at this.

That brings me to some of what's actually happening. When you interact with these models, what's happening on the back end is incredibly important to keep in mind. We talk about it publicly like you're going and interacting with this single large language model or this single AI model more generally. What's really happening on the back end is you're actually interacting with a whole bunch of different, orchestrated groups of models to provide an output to you.

Why that's important is because one of the things that large language models are really good at is writing code. The reason is that code is fairly predictable from the perspective of a probabilistic system compared to language, and it's been designed that way. So what happens then is you actually have these large language models that write code that then capture information to then pull that back to you. That's what allows for the access to information on the internet across incredible answers that you might get—it is essentially writing code for you to then pull that back in.

I think then what that can end up in is you now have text generation, you now have images and videos, and you also have audio all running behind what your current user interface looks like. What it's doing is constantly pulling in your inputs, that data across those pieces, and then providing outputs. One of the things here, too, that I would highlight is when you combine all three of these together.

You see two images there. One of them is a pastor of a church I was in before, Mark Dever. The other one is former President Obama. Both of those are videos that I created where you can essentially put words in their mouth. This one, I think, is about Mark's love of Pepsi. The other one is a political thing that you would never hear Obama say. I think a key component of this is, when you combine these technologies, you also need to continue to have that critical thinking skill about them and about what is actually being said and actually being done.

Beyond the Hype: Keeping an Eternal Perspective

Let's raise the level back up. We went pretty deep into the technology, but where are we going? We talked about where we went, where we are, and where we're going. I want to talk a little bit just about the scale. We hear a lot of these either "AI doomer" or "AI bust" scenarios. You hear a lot of different things in the media about these. I would put two major facts in front of you before I make this point.

Google, last year alone, spent more money on AI infrastructure than you could almost imagine. They spent so much that, in cash, they could have bought Ford, Moderna, and Southwest. Amazon, Microsoft, and Google spend more money each year than the Australian government. In that context, when you hear a CEO say something like this—Microsoft's AI chief gives it 18 months for all white-collar workers to be automated by AI—you need to frame it in that context and you need to ask yourself the question: is he talking to you? Because if he just invested a similar amount of money, is he actually talking to his investors?

If you think about that, he needs to make sure that he can do that again next year. So, like I'm saying, we have to remember the context in which these things are being said, and also the technology.

Now, I am definitely an AI optimist and I love this technology. I think it has incredible power to do good. But I think it's significant also for you to read the rest of the quote about the Dartmouth Conference on AI: "We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer." It's been quite a few summers since that time.

Taking a step back to the eternal and looking at how this fits in the scope of history, I think it's important to really zoom out and think about it from that perspective. If we look back—which I would argue it's very similar to the electrical revolution and to some of these others—we need to think about it in the perspective of Ecclesiastes 9 and 10. We need to think about it in the perspective that God still reigns. He is still on His throne. He is still the one in control of these things, and the eternal is the thing that truly matters. He is the one who truly knows there's nothing new under the sun.

I think it's important to handle in our minds and really think about what these technologies are at their core and what they're actually doing. As we look through that and look through history, I think it's a reminder of God's sovereignty, of His power, and how He truly is the one who reigns as we hold that perspective. So, I would challenge you, as you work through these things, to remember the context, remember history, and remember what your Bible says about those things.

With that, I would be happy to take any questions for a few minutes if there are any.

Questions and Answers

Audience Member: I don't want to have you reveal trade secrets or anything, but is Google reading our Gmail and training their language models with that?

Gordon Broadbent: Um, so first of all, my answers to these are not a reflection of my company or organization, or anything I just said. Are they reading our Gmails? I think what you'll see is there is a big difference between what's called an enterprise version of something—meaning what a company pays for—and what consumers use. Those large companies like Microsoft, Google, and Amazon have pretty significant firewalls to prevent doing that on enterprise accounts. You'll notice recently, Google in particular, there's a toggle where you can give it access to that or not. I would keep this concept in mind: if you're not paying for something, are you the consumer, or are you the product?

Audience Member: My question is, has AI ever mentioned or expressed its view on spirituality or on religions? I'm interested to know what AI has to say about spirituality and religions.

Gordon Broadbent: I think the answer to your question there would be: it depends on the context that you give it in a lot of cases. Meaning, you can produce a sermon that sounds like John MacArthur and the way his spirituality is, if you present it with the right information to do that. What you can also do is the inverse of that: a sermon that sounds like John MacArthur but would be opposed to his theology.

Audience Member: Were bots the precursor to AI? Were they part of its history? Are they related at all? Because, you know, I see bots being used all the time for a myriad of things. They are used somewhat like AI. Is there even a difference?

Gordon Broadbent: Yeah, so first of all, AI is a very broad term. I want to make that very clear. Chatbots before large language models were more deterministic in nature, meaning that code was written in order to answer the specific questions that you ask of it. The more modern chatbots, which is what I would call just a standard large language model, those would be considered AI in the category of large language models. So, a chatbot I would call more of an interface. The back end has changed to answer the question more holistically now and in far more context. Did that answer your question? Unless you were talking about robots, which is a whole different category.

Audience Member: It seems like AI and the whole push for efficiency is kind of idolizing output, efficiency, and the speed of tasks. How is that influencing the culture in looking at people as humans, understanding that humans are flawed and inefficient a lot of times? What are the dangers of that?

Gordon Broadbent: We will cover some of that in the next session, and we can talk more about that because I don't want to steal any of his thunder. I would just say this: I think that humans inherently have value. I think the important thing for us to keep in mind, as we are educating ourselves and pushing for the education of our children, is to fundamentally think about what is important for them to learn, to know, and to be able to do that draws on a lot of those unique components that make us human. But I won't steal the rest of his thunder there on that.

Audience Member: Hi. So looking at the scale of AI and just how much is being invested into it, how should we look at the environmental impact of AI? Because I'm hearing a lot of bad news, like it's just drinking up the ocean or something.

Gordon Broadbent: I can tell you it's not drinking the ocean. What I would say is this: it is a massive consumer of electricity, that's for sure. But what I would also say, from what I've seen in my work, is that we are already starting to see how what AI is doing for science is accelerating the discovery process of what can counteract some of those things and make us more efficient. This goes all the way down to—I've worked on creating new batteries that don't need toxic materials, and the ability to create new systems that are able to work far more efficiently.

This is one area where I would say the large cloud hosting providers have a massive incentive to do this as well. I'm not saying that they care necessarily about the environment—though I think many of them do, and many of those who work there do. What I would say, though, is their incentive is to reduce their cost of power and their consumption of resources because it costs a lot of money. One of the ways to do that is to be more efficient with what they're doing.

Audience Member: Talking about LLMs, you were teaching us that most of its output is kind of predictive in nature. I find that a lot of people who interface with these LLMs are kind of impressed with its seeming ability to be able to rationalize and think. How should we be thinking about its apparent ability to think and rationalize when it gives answers?

Gordon Broadbent: That's a great question. A lot of my work is in what's called external orchestration of models, which we refer to reductively as "thinking" or interacting with it from that perspective. What I would say is, if you're in a probabilistic system that has billions and billions of parameters, your ability to work through data very quickly can actually mimic the ability to think, and you can actually create real value from that.

Meaning, if I can work through the body of scientific literature about a very specific thing—we'll say about quantum computing and all the discoveries there—I can inherently pull together and optimize what the next thing should be that we would look at, and then examine that and do some of that research to then provide something that would be an inherently new idea to come out. So it's still a function of that optimization, just at a massive scale with massive amounts of compute.

Audience Member: So you're not saying it doesn't create value?

Gordon Broadbent: No, I'm not saying that.

Audience Member: Now, I have a question: is that the same thing as thinking? Do you think that's the same thing as thinking?

Gordon Broadbent: As a human?

Audience Member: Yeah.

Gordon Broadbent: I mean, inherently, no. It's not mapped to the human brain.

Audience Member: Great. We'll ponder more of that in the next session.