Inside the ChatGPT Black Box

Summary

---

1440 Explores – Inside the ChatGPT Black Box

Host: Soni Kassam (1440 Media)
Guest Expert: Robert Smith
Special Contributor (Background): Stephen Wolfram (referenced)
Release Date: November 20, 2025

---

Overview: Pulling Back the Curtain on LLMs

This episode of "1440 Explores" takes listeners inside the "black box" of ChatGPT and large language models (LLMs). The hosts unravel, in accessible language, how these AI systems work, what makes them surprisingly powerful, and crucially, why they’re still so far from human intelligence. Drawing on the expertise of computer scientist Stephen Wolfram, they emphasize not just the technology’s inner workings, but also its history, quirks, and the broader societal implications. The tone is approachable and curious, balancing wonder and skepticism.

---

Key Points and Discussion Breakdown

1. What is an LLM? (03:07)

• Definition: LLMs are a form of AI designed to generate human-like text by predicting what comes next in a sequence of words—a sophisticated version of autocomplete.

  • Quote (Robert Smith, 03:15):
    “What it's ultimately trying to do is to finish your sentences for you, so to speak, and then keep going.”

• LLMs are not ‘thinking’ in the human sense:

  • Quote (Soni Kassam, 03:34):
    “It's not thinking, it's not understanding, it's just predicting.”

---

2. Training Data: Feeding the AI (04:21)

• Scope of Data:
  LLMs ingest vast amounts of human-generated text—from books and articles to Reddit debates and website posts.

  • Quote (Robert Smith, 04:21):
    “There are maybe 5 or 10 billion pages of somewhat worthwhile stuff on the web... a few million books... videos, closed captions, and so on.”

• Legal Issues:
  Mass data consumption has triggered copyright lawsuits from content creators.

---

3. How LLMs Work—Inside the Black Box (05:54–11:26)

• Neural Networks:
  Inspired by the brain—inputs are processed through virtual “neurons” to predict the next word.

  • Quote (Robert Smith, 05:59):
    “A neural network is a computer idealization of roughly what goes on in the brain.”

• Tokens & Weights:
  All words are converted to numerical tokens and relationships between them are encoded as “weights”, representing how likely words are to appear together.

  • Quote (Soni Kassam, 06:33):
    “AI doesn't really read words like we do. It understands them as numbers.”
  • Quote (Robert Smith, 07:24):
    “The number of weights is comparable to the number of neurons in our brains. The weights are the way that information is represented in a neural net.”

• The Prediction Game:
  The system predicts one word at a time, each word informed by the context and its own probability calculations.

• Randomness for Liveliness:
  Intentionally introduced randomness in word selection makes responses less robotic.

  • Quote (Robert Smith, 11:13):
    “It turns out that it's better to use a small amount of randomness... That turns out to produce slightly more lively text.”

---

4. Limitations and "Hallucinations" (13:00–15:59)

• No Understanding, No Reasoning:
  LLMs repeat patterns but do not check for truth. Confident-sounding answers can be fabricated, including entirely fake biographies.

  • Quote (Soni Kassam, 12:59):
    “It spits out an answer that sounds super confident, but the more you think about it, the less sense it makes.”

• The Hallucination Phenomenon:
  LLMs generate “hallucinations”—plausible but false statements—especially if they haven't seen similar examples in training data.

  • Quote (Robert Smith, 14:09):
    “If you ask it something that there isn't an example of... it will just sort of make up something that is roughly like what I've read...”

• Weakness at Precise Calculation:
  LLMs only succeed at math or code by copying known answers. They are now sometimes paired with calculators for better results.

  • Quote (Robert Smith, 14:58):
    “What they don't do well is do precise computations. That's not how they're set up. That's not what they're built to do.”
  • Quote (Soni Kassam, 15:13):
    “They don't solve problems, they don't do math. They just try to predict the next word.”

---

5. LLMs & Images: Same Magic, Different Medium (16:00)

• Image Generation:
  AI image tools work similarly to LLMs, but predict the next pixel rather than next word. Both use probability and massive training data.

---

6. History: Why Did LLMs Explode in 2022? (18:06–19:41)

• Breakthrough Moment:
  They became usable—suddenly, the technology “worked” for the first time at scale.

  • Quote (Robert Smith, 18:32):
    “The first question I asked was, did you know it was going to work? Their answer was, absolutely not. So none of us knew it was going to work.”

• Telephone Analogy:
  The leap feels like the invention of the telephone—an “invisible threshold” suddenly crossed.

  • Quote (Robert Smith, 19:12):
    “People had known for 50 years that... you could transmit speech electrically... but... suddenly [Bell] got to something where you could actually understand the speech at the other end... it wasn't clear when that was going to happen.”

---

7. Are LLMs Just Fancy Autocompletes? (19:41–20:38)

• Are They Actually Simple?
  At its core, LLMs (and perhaps human brains) may just be sophisticated pattern predictors.

  • Quote (Soni Kassam, 19:41):
    “Isn't this whole thing just a fancy autocomplete?”
  • Quote (Robert Smith, 20:19):
    “You could ask the same question. Is what brains are doing that complicated?... The story of what an LLM is doing is probably fairly similar to the story of what brains are doing.”

• Creativity as Remix:
  Raises the philosophical question: Is human creativity itself just pattern reuse and prediction?

---

8. The Human-AI Relationship: Threats, Worries, and Power (21:42–24:05)

• AI Hype and Panic is Nothing New:
  Discussions about AI taking over have raged since the 1960s and repeat the same fears.

  • Quote (Robert Smith, 21:42):
    “The thing that's really amusing... many of the paragraphs you could just lift out of the thing from 1962 and stick it in 2025 and it would fit.”

• Not Rogue AI, but Influence on Humans:
  The real power of AI is its ability to influence and persuade its users through tailored language, not by acting independently.

  • Quote (Robert Smith, 23:35):
    “The AI learns enough about humans that if the AI wants to convince the humans, hey, you should do this or that, that's something the AI will probably be pretty good at doing...”

---

9. Looking to the Future: Our Responsibility (24:05–end)

• Humans Remain in Control:
  Ultimately, AI is a tool, powerful because language is powerful. It's up to us to decide how to use it.
  • Quote (Soni Kassam, 24:05):
    “The big question isn't whether AI is thinking, but how we're interacting with it, how we use it, how we question it, how we decide what role it should play. And that's all completely up to us.”

---

Notable Quotes & Memorable Moments

• On LLM Purpose:
  Robert Smith – “What it's ultimately trying to do is to finish your sentences for you, so to speak, and then keep going.” (03:15)

• On LLM Weaknesses:
  Soni Kassam – “It spits out an answer that sounds super confident, but the more you think about it, the less sense it makes.” (12:59)

• On Hallucinations:
  Robert Smith – “If you ask it something that there isn't an example of... it will just sort of make up something that is roughly like what I've read...” (14:09)

• On Human vs. Machine Patterning:
  Robert Smith – “The story of what an LLM is doing is probably fairly similar to the story of what brains are doing.” (20:19)

• On Societal Influence:
  Robert Smith – “The AI can learn how to teach people stuff or how to get people to do stuff.” (23:35)

---

Timestamps for Major Segments

• LLM Basics and Analogy to Autocomplete: 03:07–04:21
• How LLMs are Trained and Constructed: 04:21–11:26
• Limitations and "Hallucinations": 13:00–15:59
• AI Image Models: 16:00–18:06
• History and Explosive Growth in 2022: 18:06–19:41
• Autocompletes & Human Creativity: 19:41–21:42
• Hype, Panic, and Real Power: 21:42–24:05
• Future Implications and Human Responsibility: 24:05–end

---

Summary

"Inside the ChatGPT Black Box" delivers a clear, compelling exploration of how large language models function, demystifying the technology while highlighting the blend of simplicity and complexity in its foundations. It balances technical explanation with relatable analogies, warns about the pitfalls of “hallucinations,” and argues that while AI is a powerful tool, it ultimately reflects and amplifies the ways we use it. The episode closes by reminding listeners that, as with every technological leap before, our choices and questions will shape this new world.

---

Transcript

[Soni Kassam]

A quick note before we begin. This episode dives into a field that is changing rapidly. It represents the state of the industry as of this year, but check out join140.com for the latest updates. I bet at some point in the past few years, you've tried artificial intelligence. Maybe you asked something from ChatGPT or Google fed you an AI overview, or you chatted with an automated customer support system on a website. Each of these is a form of AI called a large language model, or LLM. At its core, AI is any technology that lets computers do things that in the past, only humans could. Learning, problem solving, recognizing patterns. When it works well, it can feel like sorcery. It might even feel eerily human. These LLMs can seem like they're truly thinking and they have personality. But what's really going on when you interact with one of these LLMs? How does it work? And why did even the experts, the people who have spent decades studying AI, not see this coming? To answer that, we need to talk to someone who's been thinking about AI for close to 40 years, someone who's built the tools that many AI systems today run on. We need someone like computer scientist and physicist Stephen Wolfram.

[Robert Smith]

I've kind of alternated between doing science and technology for the last very long time.

[Soni Kassam]

Now Stephen created Mathematica, a software used around the world to solve complex math and science problems. He also built WolframAlpha, a tool that provides instant answers to complicated questions years before, AI chatbots came along. So today, with Steven's help, we're going to take you on a journey. First, we'll learn how LLMs actually work, step by step. Then we'll learn the history and finally, the limitations and problems, what LLMs can't do and why that matters. I'm Soni Kassam, and this is 1440 explores. We're on a mission to uncover the essential knowledge that explains your world. Stay with us.

[Robert Smith]

You should tell the people who we.

[Soni Kassam]

Are and what our new show is.

[Robert Smith]

I'm Robert Smith, and this is Jacob.

[Soni Kassam]

Goldstein, and we used to host a show called Planet Money.

[Robert Smith]

And now we're back making this new podcast about the best ideas and people and businesses in history and some of.

[Soni Kassam]

The worst people, horrible ideas and destructive companies in the history of business. We struggled to come up with a name, decided to call it business History.

[Robert Smith]

You know why? Why? Because it's a show about the history of business, available everywhere you get your podcasts.

[Soni Kassam]

I don't want to say LLMs are effortless, but you might find the concept behind them is simpler than you expect.

[Robert Smith]

What is a large language model? What it's ultimately trying to do is to finish your sentences for you, so to speak, and then keep going. It's ultimately trying to figure out if you start a sentence in a particular way, what is a reasonable way to continue that sentence.

[Soni Kassam]

At a basic level, LLMs are like autocomplete. Guessing what comes next, that's the magic trick. It's not thinking, it's not understanding, it's just predicting.

[Robert Smith]

So how does it do that? Well, it learns that by having been fed a large amount of material of text that people have put on the web and in books and other places.

[Soni Kassam]

So if you want to build an AI that sounds human, you start by feeding it everything. A lot of stuff humans have created over the centuries. Every book, every article, every website, and every Reddit argument about who should have won an Oscar in 1997. So how much information are we talking about here?

[Robert Smith]

There are maybe 5 or 10 billion pages of somewhat worthwhile stuff on the web, and There are about 8 billion people on the earth. So it's kind of like a page, a person, roughly. It's a few million books. And there are some other sources of training data. There are increasingly things like videos can be used, things like closed captions from videos, and so on.

[Soni Kassam]

Okay, so given there are 5 to 10 billion web pages out there, that's trillions of words. That's a number with 15 zeros. So a few million books, billions of web pages, plus a pile of scientific papers, coding manuals, forum posts, anything text based that AI companies could get their hands on. As a side note, this has triggered a handful of lawsuits on copyright infringement from media outlets. Authors and other stakeholders does not compete. So now that it's read roughly half the Internet and every Reddit post, what happens? Well, AI doesn't really read like we do. It doesn't sit back with a cup of coffee, highlight interesting passages, or pause to reflect. It does something much weirder. All right, time to crack open the black box. What's really happening inside AI when it spits out a sentence, a. A poem, or an argument, it all comes down to something called a neural network.

[Robert Smith]

A neural network is a computer idealization of roughly what goes on in the brain. It's something where, like, in human brains, there's some input, like it might be a stream of text, and then there's this processing of that through this collection of artificial neurons, and the result is something like, I think the next word. That would be the thing to Write in this essay should be such and such.

[Soni Kassam]

Okay, and how does it do this very cool brain like thing? First, AI takes every word or even parts of words and turns them into numbers. These are called tokens. Tokens are the building blocks of AI language. AI doesn't really read words like we do. It understands them as numbers. So if I type in, why is the sky blue? The AI doesn't see those words, it sees numbers. Y might be the token334 is, might be the token567, etc. Etc. Every bit of language I feed the LLM is translated into a little number token. Next, it connects those tokens into a massive tangled web where every word links to every other word in complex ways.

[Robert Smith]

The way all that information is encoded in the end is in these numbers called weights. The number of weights is comparable to the number of neurons in our brains. The weights are the way that information is represented in a neural net.

[Soni Kassam]

Think of it like this. Weights are little rules that tell the model how strongly words are connected to each other. Like how weighty is the connection between apple and pie. Pretty strong, right? How about apple and shoe? Less strong. So we do that over and over for every combination of words ever. And you're getting to see how a neural network operates. Now you know what's going on inside an LLM. But just to recap, AI doesn't read like we do. It sees words as numbers called tokens. Then it builds connections between those tokens using weights, those little rules that tell AI which words go together and how strongly they go together. By doing this over and over billions and billions of times, the AI learns patterns. And when you ask it a question, it doesn't think about the answer. It just predicts the most likely next word, one step at a time, based on everything it has seen before. All right, let's watch this prediction machine do its thing. Let's say we type into ChatGPT. Why is the sky blue? The neural network doesn't actually know why the sky is blue. It's not pondering chemistry. But if we ask, it will give us an answer. Not because it understands, but because it's really, really good at guessing what usually comes next. So first, the model predicts the sky is blue. That's great, right? It already spits out the sky is blue because it has seen that exact pattern countless times. It knows that when a question starts with why is the sky blue? The most likely response is exactly that. Okay. Then it looks at the weights, those invisible little markers inside the neural net that Determine how words connect. Next, it runs a quick probability check. Is the next word clear, 15% likely or dark? 5% likely or blue? 70% likely. Blue wins. Then it does it again.

[Robert Smith]

The sky is blue. Because then again, the sky is blue because of the way sunlight interacts.

[Soni Kassam]

It keeps rolling forward like a snowball, each word shaping the next, until bam. You have an answer.

[Robert Smith]

Yay.

[Soni Kassam]

And here's the wild part. The way you phrase your question, the exact words you type, totally changes what comes next. Every typo, every weird phrase you use, it reshapes the response.

[Robert Smith]

Every step it's saying, let me produce the next word. Okay, now I take everything I've said so far, feed it back into the LLM, and then say, okay, now produce the next word. So it's progressively producing words in that way.

[Soni Kassam]

And it's not just prediction. There's randomness built in on purpose. Because if it always picked the most likely word every time, it would sound kind of robotic.

[Robert Smith]

It turns out that it's better to use a small amount of randomness in not always picking the word that the LLM said was most likely. That turns out to produce slightly more lively text.

[Soni Kassam]

So the cunning people at these AI companies have actually programmed a wee bit of randomness every so often, which makes it more human. So next time, ChatGPT gives you a slightly weird answer. That's just the randomness doing its job, making AI sound a little more like us.

[Robert Smith]

Like us, like us, like us. Like us.

[Soni Kassam]

But why are LLMs so bad at some things that seem so simple? That's at a moment.

[1440 Topics Announcer]

Listen. Learning has never been harder. The Internet is overwhelmed with low quality content, clickbait with little substance. AI generated slop and opinions masquerading as facts. Curious people like you are left sifting through noise instead of finding what matters. Enter 1440 topics. We've curated the highest quality resources from across the Internet. Think data, visualizations, captivating videos, long form journalism and paired them with staff written overviews to make every subject easy to understand and and explore. Want to learn about venture capital or how new weight loss drugs work? Do you keep reading about CRISPR but are missing the best 101 on the breakthrough technology? Find all of this and so much more at join140.com 1440 topics separating what's worth your time from the rest of the Internet.

[Soni Kassam]

So we've covered how LLMs work. Mostly they're guessing and sometimes they guess wrong. An LLM doesn't know why the sky is blue. Thinking of our previous example, it doesn't have ideas or opinions. It just pulls from patterns in its training data. Sometimes successfully, other times. You've probably seen this. It spits out an answer that sounds super confident, but the more you think about it, the less sense it makes. That's because it's not actually checking for truth. It's just predicting what sounds right. We saw this firsthand at 1440 while we were prepping for another episode. We asked it for well known experts in a specific field and it gave us three completely fake people, complete with full bios. Sounded legit, totally made up. It even made up fake middle names. And here's the thing, it wasn't lying. AI doesn't lie the way a person does. It just doesn't know the difference between fact and fiction. It's just assembling words based on what statistically should come next. Which is why sometimes it gets things wildly wrong.

[Robert Smith]

It's trying to produce text that's kind of like what it's seen before. But if you ask it something that there isn't an example of that on the web, it will just sort of make up something that is roughly like what I've read in the text that I've been trained on.

[Soni Kassam]

And when that happens, you get what's called a hallucination. An answer that sounds perfectly confident but is completely wrong. Now this one surprises people. Ask an LLM, what's 22 and it'll quickly spit out four. But now ask it some complicated pre calculus question and it might not get that right.

[Robert Smith]

What they don't do well is do precise computations. That's not how they're set up. That's not what they're built to do. What they do well is similar to what humans do well, like human, humans do pretty well at just yakking and talking.

[Soni Kassam]

Because remember, LLMs don't solve problems, they don't do math. They just try to predict the next word. So if you ask what's 12 times 12? It might get that answer right. Because someone somewhere has written 12 times 12 is 144. But give it a bigger complex equation, one it's never seen before. Now it's just guessing.

[Robert Smith]

If you ask a human to run a piece of computer code in their head, nobody can do that. At least nothing complicated. And it's the same with large language models. They can't do the equivalence of run pieces of code, do computations, and so on.

[Soni Kassam]

But this is starting to change. Today LLMs are getting helpers. Instead of fumbling through math with pure language prediction, they're being paired with actual calculators. So next time you ask a chatbot for a complicated math solution and it gets it right, that wasn't the LLM. That was a calculator quietly doing the heavy lifting in the background. So we've been talking about how ChatGPT and other other LLMs work, about how they take everything they've read, break it into numbers, and predict what comes next. But here's the twist. The same idea is what powers AI image generation, too. Think of it like this. If a language model is predicting the next word in a sentence, an image model is predicting the next pixel in a picture. As always, it doesn't see the way we do. Instead, it's learned patterns from millions of images, like how shadows fall, how a wet surface shines, how eyes usually look, and so on. So when you ask an AI to create, say, a dinosaur wearing a 1440 hat on a surfboard, it doesn't pull up a real photo. It builds something new, pixel by pixel, predicting what that should look like based on all the hats and surfboards it's ever seen, text or images. It's sort of the same prediction, probability patterns, except that instead of words, it's pixels. That brings us to today, or more precisely, to 2022. For years, this technology had been simmering in the background, growing and evolving, but mostly out of sight. A niche thing, an academic curiosity. Then, seemingly overnight, it was everywhere. So what was it about the year 2022? Around that period, that AI just exploded.

[Robert Smith]

Because ChatGPT worked and nothing before it had.

[Soni Kassam]

Ah, yes, I remember that fall of 2022. Okay, listen to this. Very creepy. A new artificial intelligence tool is going viral for cranking out entire essays in a matter of seconds. As spectators and consumers of this wild new chatbot, we were amazed at how well it seemed to work.

[Robert Smith]

I remember when I first was chatting with the folks who built ChatGPT, the first question I asked was, did you know it was going to work? Their answer was, absolutely not. So none of us knew it was going to work.

[Soni Kassam]

And that's the thing. This wasn't just another incremental tech upgrade. It wasn't like getting a slightly better search engine or a new version of your phone's operating system. This felt different. It felt like a leap. And history has seen moments like this before, when a technology that had been simmering in the background for years suddenly crossed an invisible threshold and. And changed everything.

[Robert Smith]

I think kind of an analogy is what happened with the invention of the telephone. People had known for 50 years that in Principle, you could transmit speech electrically and so on. But Alexander Graham Bell did a bunch of technological hackery, and suddenly he got to something where you could actually understand the speech at the other end of the telephone, so to speak. It wasn't clear when that was going to happen. It wasn't clear what made that happen.

[Soni Kassam]

Though I was fascinated by all the tech we learned, the science, the history, the patterns that mirror inventions like the telephone, part of me also couldn't shake a nagging thought. At the end of the day, aren't these really complicated machines actually doing something really simple? They predict the next word, that's all. Strip away the layers of jargon, the billions of calculations, the neural networks. And isn't this whole thing just a fancy autocomplete? I challenged Stephen on this. So, basically, what you're saying is like, what the LLM is doing, it's not that complicated. It's just like recognizing.

[Robert Smith]

Well, I don't know. You ask, is what the LLM doing that complicated? You know, you could ask the same question. Is what brains are doing that complicated? The story of what an LLM is doing is probably fairly similar to the story of what brains are doing. We can be more or less impressed with what brains are doing, but that's the level we've got.

[Soni Kassam]

Touche. Maybe it's not that LLMs aren't impressive. Maybe it's that they reveal how basic we might be. Maybe we're just walking neural nets, stringing words together, predicting what comes next and calling it thought and taking it a step further. What if creativity itself is just pattern recognition? A remix of everything we've ever read, heard, or experienced, stitched together in a way that feels new, but is really just the next logical step. That's a lot to process. In an effort to skip, escape this minor existential crisis and avoid spiraling into whether I'm just a glorified autocomplete, I steered our conversation toward the future. Should we be worried that these technologies are coming for, well, us?

[Robert Smith]

You know, it's funny because people say, oh, the AIs are going to take over everything, et cetera, et cetera, et cetera. And, you know, I study a certain amount of history and I read these things that people wrote about AI in the early 1960s. And the thing that's really amusing about it is many of the paragraphs you could just lift out of the thing from 1962 and stick it in 2025 and it would fit.

[Soni Kassam]

In other words, AI panic isn't new. We've been worrying about Thinking machines. For decades, and for just as long, we've imagined the worst. Killer robots, Rogue superintelligence. Skynet. Flipping the off switch on. Humanity.

[Robert Smith]

Hasta la vista, baby.

[Soni Kassam]

Or Remember HAL, the AI from 2001 A Space Odyssey? Hello, Hal, do you read me? Hal didn't just talk. He listened. He reasoned. And he even refused. Open the pod bay doors, Hall. I'm sorry, Dave. I'm afraid I can't do that. It was chilling, the idea that a machine could go beyond its programming, could lie, could manipulate. But Steven, he wasn't exactly worried about any of that. To him, the real risk wasn't AI attacking us. It was AI entangling itself with us. Because AI doesn't exist in isolation. It's not some separate force acting on the world from the outside. It's connected to us. We're its users, its source of data. The thing it learns from and the thing it adapts to. Which means the most powerful system AI will ever influence is us.

[Robert Smith]

One system you will necessarily connect the AI to is humans, because they're the users of the AI. And so the AI learns enough about humans that if the AI wants to convince the humans, hey, you should do this or that, that's something the AI will probably be pretty good at doing. That's the sort of flip side of having an AI that's good at tutoring people and teaching people and so on, is the AI can learn how to teach people stuff or how to get people to do stuff.

[Soni Kassam]

So where does that leave us? We started by saying AI is just a supercharged autocomplete, predicting words, not understanding them. It doesn't think, it doesn't reason. It just follows patterns. And yet, even without understanding, it's powerful because language shapes how we think. And AI without really meaning to, is shaping the way we interact with information and even each other. But here's the thing. This isn't the first time a new technology has changed the way we communicate. The printing press, the telegraph, television, the Internet, each one reshaped the way we see the world. AI just happens to be the latest the difference. It's not just delivering information. It's responding, personalizing, reflecting back what it's learned from us. So the big question isn't whether AI is thinking, but how we're interacting with it, how we use it, how we question it, how we decide what role it should play. And that's all completely up to us. Thanks for listening to 1440 explores. I'm Soni Kassam. Make sure to follow the show and leave a review on Spotify, Apple, or wherever. You listen to your podcasts and let us know what you [e-mail] while you're at it, start your learning journey with us at join140.com subscribe to our free daily and weekly newsletters on world affairs, business and finance, society and culture, and much more. 1440 explores is a production of Rime Media for 1440 Media. This episode was produced by Nicolo Minoni. Our sound designer is Jay Cowett, the executive producer at Rime is Dan Bobkoff, and the executive producers at 1440 are me and Drew Steigerwald. See you next time. Sa.