How DeepSeek’s AI Compares To Established Models

IRA FLATOW: This is Science Friday. I’m Ira Flatow. Later in the hour, an investigation into fraud and misconduct within Alzheimer’s research. We’ll talk with the author of a new book who makes the case that image doctoring might be part of the reason scientists haven’t yet come up with an effective treatment for the disease.

But first, last week, if you recall, we briefly talked about new advances in AI, especially this offering from a Chinese company called Deep Seek, which supposedly needs a lot less computing power to run than many of the other AI models on the market, and it costs lots less money to use. It’s been described as so revolutionary that I really wanted to take a deeper dive into Deep Seek.

What’s all the buzz? What makes one model smarter than another, less power hungry? Joining me to help dive into that is Will Douglas Heaven, senior editor for AI coverage at MIT Technology Review. He’s based in the UK. Welcome back to the program, Will.

WILL DOUGLAS HEAVEN: Hi. Thanks a lot for having me.

IRA FLATOW: Happy to have you. Is it really as good as people are saying?

WILL DOUGLAS HEAVEN: Yeah the thing is, I think it’s really, really good. I think the thing that has got people really shocked is that it is as good as the best that the US has made. We’re at a stage now where the margins between the best new models are pretty slim, you know?

Is OpenAI’s best better than Google’s best? Is that better than Anthropic’s best? Just the fact that a Chinese company has matched what the best US labs can do is itself a shocking thing. I don’t think people thought that China had caught up so fast.

IRA FLATOW: So what is its competitive advantage here? From what I’ve been reading, it seems that Deep Seek computer geeks figured out a much simpler way to program the less powerful, cheaper NVidia chips that the US government allowed to be exported to China, basically.

WILL DOUGLAS HEAVEN: They’ve done a lot of interesting things. There’s also a lot of things that aren’t quite clear. So we don’t know exactly what computer chips Deep Seek has, and it’s also unclear how much of this work they did before the export controls kicked in.

But from the several papers that they’ve released– and the very cool thing about them is that they are sharing all their information, which we’re not seeing from the US companies. It looks like they have squeezed a lot more juice out of the NVidia chips that they do have. They’ve done some very clever engineering work to sort of reprogram them down at very low levels to kind of get more power out of the box than NVidia gives you by default.

So that’s one cool thing they’ve done. These are also sort of got innovative techniques in how they gather data to train the models. But one key thing in their approach is they’ve sort of found ways to sidestep the use of human data labelers, which, you know, if you think about how you have to build one of these large language models, the first stage is you basically scrape as much information as you can from the internet and millions of books, et cetera. And you know, we’re probably familiar with that part of the story.

IRA FLATOW: Stealing other people’s data, in other words.

WILL DOUGLAS HEAVEN: Yeah.

IRA FLATOW: Yeah.

WILL DOUGLAS HEAVEN: Yeah, pretty much. And as a side, as you know, you’ve got to laugh when OpenAI is upset it’s claiming now that Deep Seek maybe stole some of the output from its models. I mean, the schadenfreude is sweet.

But all you get from training a large language model on the internet is a model that’s really good at kind of like mimicking internet documents. It’s not something that’s very useful. The chatbots that we’ve sort of come to know, where you can ask them questions and make them do all sorts of different tasks, to make them do those things, you need to do this extra layer of training. And that’s typically been done by getting a lot of people to come up with ideal question-answer scenarios and training the model to sort of act more like that.

IRA FLATOW: So you need you need a lot of people involved is basically what you’re saying.

WILL DOUGLAS HEAVEN: Yeah, exactly. That’s time consuming and costly. Deep Seek’s found a way to do without that. Probably the coolest trick that Deep Seek used is this thing called reinforcement learning, which essentially– and AI models sort of learn by trial and error.

Listeners might recall Deepmind back in 2016. They built this board game-playing AI called AlphaGo. And sort of the amazing thing that they showed was if you get an AI to start just trying things at random, and then if it gets it slightly right, you nudge it more in that direction.

If you do that many, many, many, many times, then you end up incrementally getting better and better and better. So you ended up in Deepmind’s case, with an AI that could, starting from scratch, went on to beat a human grandmaster at Go. What deep seek has done is applied that technique to language models.

You know, they didn’t want it to play a game. Obviously, they wanted it to get better at giving thought-through answers to questions that you asked the language model. And again, to start off with, it did a pretty poor job, but they nudged it bit by bit in the right direction. And you let that run enough times, and it sort of figures out itself how to get better, sort of improving bit by bit as it goes.

IRA FLATOW: So what you’re basically saying is that it’s teaching itself how to get better.

WILL DOUGLAS HEAVEN: Yeah, I hesitate to sort of phrase it like that because it always gives the eye some sense of agency, and it’s, you know, going to do its own thing. Yeah, there is a term called self-play. It sort of learns to play itself and get better as it goes. So you can think of it in that way.

IRA FLATOW: You know, aside from the human involvement, one of the problems with AI, as we know, is that the computers use a tremendous amount of energy, even more than crypto mining, which is shockingly high. I mean, is Deep Seek less energy-hungry, then, for all its advantages across the board?

WILL DOUGLAS HEAVEN: Yet again, this is something that we’ve heard a lot about in the in the last week or so. And the answer to that as well is not as clear as it was initially made out. It does seem that they trained the model. They built the model using less energy and more cheaply.

Running it may be cheaper as well, but the thing is, with the latest type of model that they’ve built, they’re known as sort of chain of thought models rather than, if you’re familiar with using something like ChatGPT and you ask it a question, and it pretty much gives the first response it comes up with back at you. But there’s a brand new sort of paradigm in chatbots now where you ask it a question, and it sort of takes its time and steps through, sort of shows its answers, shows its reasoning as it steps through its response.

And each one of those steps is like a whole separate call to the language model. So although Deep Seek’s new model R1 may be more efficient, the fact that it is one of these sort of chain of thought reasoning models may end up using more energy than the vanilla type of language models we’ve actually seen. And another complicating factor is that now they’ve shown everybody how they did it and essentially given away the model for free. I think we can expect so many other companies and startups and research groups sort of picking it up and rolling their own based on this technique.

IRA FLATOW: One of the criticisms of AI is that sometimes, it’s going to make up the answers if it doesn’t know it, right?

WILL DOUGLAS HEAVEN: Right.

IRA FLATOW: There are two layers here. One, how does it stack up on reliability or this issue, as they call it, hallucinations? And second, because it’s a Chinese model, is there censorship going on here?

WILL DOUGLAS HEAVEN: Yeah, so a lot of stuff happening there as well. All models hallucinate, and they will continue to do so as long as they’re sort of built in this way. You know, there’s statistical slot machines.

You can polish them up as much as you like, but you’re still going to have the chance that it’ll make stuff up. And I have seen examples that Deep Seek’s model actually isn’t great in this respect. If it can’t answer a question, it will still have a go at answering it and give you a bunch of nonsense.

Anecdotally, based on a bunch of examples that people are posting online, having played around with it, it looks like it can make some howlers. But yeah, the question of censorship is interesting. I mean, I guess it’s not surprising at all that, you know, a model built in China, it can’t tell you anything about Tiananmen Square.

It won’t answer questions about Chinese politics at all. It just says, you know, I’m sorry. Let’s talk about something else.

IRA FLATOW: If they’re innovating like this but making their code available– as open source, as you say– are we likely to see the other competitors saying we’re going to use this because why not?

WILL DOUGLAS HEAVEN: Yeah, I mean, you can download the deep sig app from the app store or Google Play and have a go with it right now.

IRA FLATOW: For free?

WILL DOUGLAS HEAVEN: For free. Yeah, I’ve been playing with it. I mean, I quite like it.

In many ways, it’s sort of– it’s more friendly than ChatGPT’s or Google’s Gemini. But there are also lots and lots of companies that sort of offer services that sort of provide a wrapper to all these different chatbots that are now on the market, and you sort of just– you go to those companies, and you can pick and choose whichever one you want within days of it being released. Many of these companies were offering Deep Seek as one of the alternatives. Of course, not just companies offering, you know, Deep Seek’s model as is to people, but because it’s open source, you can adapt it.

So there’s a company called Huggy Face that sort of reverse engineered it and made their own version called Open R1. There’s also a technique called distillation, where you can take a really powerful language model and sort of use it to teach a smaller, less powerful one, but give it most of the abilities that the better one has. And Deep Seek’s R1 has already been distilled into a bunch of different models. Yeah, so I think we’re going to see adaptations of it and people copying it for some time to come.

IRA FLATOW: So what’s your take on artificial general intelligence? We’ve talked about this before on the show. Will we be getting there and when, in your opinion?

WILL DOUGLAS HEAVEN: I am super skeptical about everything about AGI.

IRA FLATOW: You are?

WILL DOUGLAS HEAVEN: Partly, it’s just a term that means very little. It means different things to different people who use it. I think it’s become a marketing term more than anything else.

I mean, I roll my eyes when people like Sam Altman tell us that AGI is coming. He was telling us that two or three years ago, and when I spoke to him then, you know, he’d say, you know, the reason OpenAI is releasing these models is to show people what’s possible because society needs to know what’s coming, and there’s going to be such a big societal adjustment to this new technology that we all need to sort of educate ourselves and get prepared.

Now he’s talking about AGI is still coming, but he means something, I don’t know, like a sort of a workplace productivity tool that we’re all going to use. And that’s now what he means. And I’m picking Sam Altman as the example here, but like, most of the big tech CEOs all write blog posts talking about, you know, this is what they’re building.

I think AGI has been this term that essentially means, you know, AI but better than what we have today. The definition that’s most usually used is, you know, an AI that can match humans on a wide range of cognitive tasks. But how would you really test that, and how would you know when we’ve got there?

And it’s not clear at all that we’ll get there on the current path, even with these large language models. Maybe they’ll plateau soon. Maybe they’ll just be very, very good language mimics and, you know, we’ll stop there, and ther’ell have to be a whole other breakthrough in a different type of AI technology to take us further. But I just– AGI is my least favorite term.

IRA FLATOW: Well, Will, I want to thank you for taking us really into the weeds on this. I hope we still have a few listeners left who appreciate how deeply we’ve taken a dive here, but I really enjoyed it. Thank you for taking time to be with us today.

WILL DOUGLAS HEAVEN: Thanks a lot.

IRA FLATOW: Will Douglas Heaven, senior editor for AI coverage at MIT Technology Review.

Copyright © 2025 Science Friday Initiative. All rights reserved. Science Friday transcripts are produced on a tight deadline by 3Play Media. Fidelity to the original aired/published audio or video file might vary, and text might be updated or amended in the future. For the authoritative record of Science Friday’s programming, please visit the original aired/published recording. For terms of use and more information, visit our policies pages at http://www.sciencefriday.com/about/policies/