AI: Bubbling Questions about NPR Limitations

Scott Detrow, host:

I asked ChatGpt to write an introduction to the radio segment on artificial intelligence. My Prompt – Write a 30-second introduction to the Radio News segment. Segment Topics – After years of promise and sky expectations, there is a way to suddenly wonder whether technology will hit the ceiling. This is part of what we got.

(Reading) For years it was welcomed as a future – a game changer aimed at rebuilding industry, redefine everyday life, and breaking boundaries that we never imagined. But now, this groundbreaking technology once endless promise is faced with new scrutiny. Experts ask, did we hit the ceiling?

It was ChatGpt. Pulling the Wheel back to human – MIT has reported on the past week throwing cold water at the value of AI at work. Consumers were disappointed with the latest version of CHATGPT, which was released earlier this month. Openai CEO Sam Altman has brought the idea for AI Bubble to the forefront, while tech stocks have taken DIP.

I'll talk about all of this with Cal Newport, a New Yorker contributor writer and professor of computer science at Georgetown. welcome.

Cal Newport: Thank you for welcoming me.

DETROW: Start with the latest version of ChatGpt. Was it really a shame?

Newport: It's a great technology, but it wasn't a transformative technology. That's what we've been promised since the GPT-4 came out. So the next major model was the next major leap, and the GPT-5 wasn't.

DETROW: One thing I pointed out in a recent article is that voices are being generated. It has always been an exponential leap and has truly owned in recent years. And a kind of general idea would, of course, always be a leap and boundary until we gain superhuman intelligence.

Newport: And the reason they owned is because we first made those leaps. So there was a real curve. It appeared on paper in 2020. This is how quickly these models get faster as they grow, and GPT-3 and GBT-4 fell correctly on their curves. That's why I have a lot of confidence in the AI ​​industry. However, after GPT-4, its progression fell off that curve and became quite flat.

DETROW: ChatGpt is the leader. Obviously, this is a big data point, as this is the most famous of all these models. But what are you looking for to get the feeling, is this just one blip, or what is the whole picture here?

Newport: This is a problem with all large language models. Essentially, the idea of ​​just training the model bigger and longer and making it much smarter has stopped it from working all over the place. We first began to notice this in the second half of 2023, early 2024. They focus on what I call post-training improvements. This is more focused and more progressive, with all major models of all major AI companies focusing on this progressive approach to improving now.

DETROW: I want to talk about it right away. First of all, I would like to get your ideas on this other big headline these days. This MIT report – ubiquitous headlines were 95% of corporate generative AI pilots fail – 95%. Do you think that number is surprising?

Newport: The number is not surprising at all. What we wanted was to happen with AI in the workplace. This was an agent revolution. But the model isn't enough for that. They hallucinate. They are very unreliable. They make mistakes and take strange actions. Therefore, as soon as you leave these tools that have built on top of the language model – very narrow applications where language models are very good, these more common business tools are still less reliable.

DETROW: You're talking about hope, and many of these companies have hope. Many investors have hope. But there are many people who are really surprised about all of this. Whether that means work safety, some of what happens in the line with AI, whether you know, you know, you know, you know, you know, or not it means a sci-fi type view. Do you think slowing down is inevitably good news for people who are worried, or do you think this continues to focus on so many industries?

Newport: I think it's good news for those worried. I think it will be the next five years.

DETROW: I understand.

Newport: I think the idea is that, like Dario Amody, you can have an unemployment rate of up to 20%, and you can have up to 50% of all new white-collar jobs that will be automated in the near future. The technology is not there and there is no route to get there in the near future. The distant future is another question, but I don't think about the Doom scenario we've been asking for over the past six months or so.

DETROW: mentioned before after training. You had a great ratio phor for it, it relates to the car. Do you walk us through it?

Newport: Well, there are two ways to improve your language model. The first method is to make it bigger and train longer. This is what is called pre-training. This is what provides the basic functionality of the model. Then there are other ways to improve them. This can be considered after training. So, before training, if the soup of the car is on top of the car, like the car. And what happened is attracting attention in the industry ahead and after training, so they are less likely to try and build a much better car, focusing on trying to get more performance from the cars they already have.

DETROW: Does this lead to a massive rethinking of what comes next? Or do you just fine-tune your current approach to how these models improve?

Newport: I think it's almost a moment of crisis for AI companies, as the capital expenditure required to build these large models is incredibly large. And to make huge amounts of money from these technologies, you need a very lucrative application. How can you make enough revenue to justify the hundreds of billions of dollars of capital expenditure needed to train these models?

DETROW: What does this mean for people who are already starting to use AI in their daily lives at work or at home in the near future? Will that continue? Do you think we'll hit a bubble there? For example, do you think it's coming next on a small consumer scale?

Newport: I think more effort will be gained from the fit of the product market. So, rather than focusing on making this model bigger and bigger, we need to pay more attention to not only accessing the model through a chat interface, but also building bespoke tools on top of these basic models for specific use cases. So, in reality, I think the footprint of a normal user's life will be more useful as you may get custom tools that are suitable for your particular job.

There's still a lot to worry about. Language models can do all sorts of things that are pain, like today. Generating Internet slops. It's much easier to have persuasive misinformation. The possibility of fraud is explosive. All of these are negative. But I'm probably just going to get better tools as an average user in the near future. That's not necessarily so bad.

DETROW: It's Cal Newport, author and professor of computer science at Georgetown University. Thank you for coming in.

Newport: Thank you.

(Music Sound Bite)

Copyright©2025 NPR. Unauthorized reproduction is prohibited. For more information, please see the www.npr.org website Terms of Use and Permissions page.

The accuracy and availability of NPR transcripts may vary. You can modify the transcript text to fix the errors and match the update to the audio. Audio on npr.org may be edited after the original broadcast or publication. The authoritative record of NPR programming is audio records.