Our director, computer engineering prof <a href="https://www.ecs.baylor.edu/person/dr-robert-j-marks-ii” rel=”nofollow noopener” target=”_blank”>Robert J. Marks, was a guest last Monday on the Humanize Today podcast, hosted by Wesley J. Smith, talking about the benefits and perils of AI.
As Smith notes, he has worked on artificial intelligence for more than three decades, including work with NASA, the Jet Propulsion Laboratory, the National Institutes of Health, the Army Research Lab, and the Office of Naval Research, as well as consulting for Microsoft and Boeing.
Here’s the audio (August 10, 2026/1:08:28). And here are just a few points to ponder that came up in the discussion:
● What do generative AI models like Anthropic’s Claude and X’s Grok actually do?
Marks: [13:56] Large language models are basically word completions. For example, when you’re typing, it suggests the next word, or the next couple of words. So if I typed in “How are”, it would type in “you?”
But it’s much more sophisticated than that, because it takes the context of previous conversations, something called a context window. That context window places the next word, not only in relationship to what you’re saying now, but everything you’ve said in the past. Current models go to a million words…
Smith: [13:57] Is this what they mean when they say AI learns? That is, it takes everything that it has experienced — I’m using that term very loosely — in the past or that has passed through its software in the past, and then uses that to apply what should come next?
Marks: [13:57] Yes, it does, but it does that in the following sense: The generative AI has been trained on the corpus of probably all the written material that exists in the world, in trillions and trillions of documents. But most of the important works of the world have already been used to train AI, and there was concern for a while that these generative AI models were running out of fuel.
● Can AI limitlessly generate new knowledge?
Marks: [13:58] This was proposed: Why don’t we use AI to generate more training data for AI?
Well, people that have done that, it doesn’t work. There’s something called model collapse , which says that if you use AI to train AI to train AI to train AI, that it eventually becomes kind of a blubbering idiot. It doesn’t get better.
Note: A problem that has raised alarms among book lovers — that AI companies are buying rare old books and destroying them to feed data to the machines — stems from this very issue: The companies are running out of other sources of new information. See, for example, “A ‘bananas’ order for 5,000 obscure titles from a bookshop in Galway fuels suspicion. Independent retailers are receiving vast orders for obscure titles, fuelling suspicions that tech companies are are using them to train AI” (Irish Times, August 10, 2026 )
● What about the development of AI superintelligence?
Marks: [14:00] First of all, I would maintain it’s a religion. … The data processing inequality that Claude Shannon came up with in information theory really prohibits this. AI cannot generate AI. There’s lots of people like Ray Kurzweil in his book, The The Singularity Is Near (2005) who talk about these exponential explosions and the ability to do things. But exponential explosions are never sustainable.
Note: Shannon’s theory is sometimes expressed as “Processing cannot create information.”
● What are the benefits that you hope to come for humanity from the technology?
Marks: [14;08] … Another one is healthcare. I think there’s thousands of papers which are published every year. You have all of these papers published every day, and the normal physician can’t read all of these papers, but AI can. And what AI can do is it can put together all of this information and present it. Hopefully in a verifiable way. Once you get a result from AI, of course, you have to verify it. And so we get on the backs of giants very easily now, as opposed to Newton, who did it through access to his few thousand books.
● What about AI putting people out of work?
Marks: [14:26] Historically, there were the Luddites, of course, that were doing weaving. They came up with an automated way to do that. By the way, that was the first computer. They actually had hole punch cards that they used for the weaving.
And the Luddites came in and they said, you know, this is terrible that they’re taking away our jobs. They began to destroy a lot of these places where they did the automatic weaving. And so, yeah, the technological is going to be disruptive. But I would argue, on the other hand, that it also creates jobs.
And also I think, with what we understand about AI, we can also look at jobs that it won’t disrupt, and those are the ones that use the unique human attributes of understanding and creativity.
Listen to the rest here.
Note: Historically, new technologies have only eliminated jobs in the short term. For example, when farming was mechanized in North America, people moved to the cities. Most got jobs in factories and offices, producing newly available goods and services. In turn, their jobs created other jobs. For example, people who worked in offices, as opposed to barns, needed tailors and barbers/hairdressers. The businesses themselves needed typewriters and stationery. Later, the businesses needed photocopiers and computers.
Every successful new technology became a business for someone. Over time, standards of living rose.
Dr. Marks’s most recent book is Non-Computable You: What You Do That Artificial Intelligence Never Will (Discovery Institute Press, 2022). He is working on a revised edition.
