How Hacks Happen

AI Part 1: Deepseek

• Many Worlds Productions • Season 3 • Episode 15

Share episode

The AI model Deepseek was all over the news in January 2025. Deepseek's developers claimed its costs to develop and run were a fraction of the costs for large popular AI models, which caused a big disruption in the stock market. But how true are these claims? In the first of a three-part series on AI, host Michele Bousquet dives in and give Deepseek a spin, and debunks some of the hype around this new export from China.

Resources:

Send us a text

Everyday AI: Your daily guide to grown with Generative AI
Can't keep up with AI? We've got you. Everyday AI helps you keep up and get ahead.

Listen on: Apple Podcasts   Spotify

Support the show

Join our Patreon to listen ad-free!

AI Part 1: Deepseek

Welcome to How Hacks Happen. I’m your host, Michele Bousquet, and I really like looking into hacks and scams and all the new things in technology that go along with them. This episode is the first in a three-part series on one of the hottest topics in tech right now, artificial intelligence, also known as AI. We’ll be taking a look at the recent explosion of AI and what it means for hacks and scams. And we’ll also do a little AI experimentation. Ready for some fun?

AI has been big for a while, in fact I did an episode a few months ago on ChatGPT. But a few weeks ago, the news was full of a new AI model called Deepseek. 

So, what’s the big deal about this new model?

In our episode “ChatGPT’s Web of Lies,” I talked about how AI works. If you haven’t listened to that one yet, you might want to give it a listen. In any case, here’s a little recap.

An AI model is a program that looks at a whole bunch of data–like a bunch of text, or a bunch of pictures–and figures out patterns or keywords. This part of the process is called training. Then, after the model is trained, you can ask it questions, or ask it to create a picture, and it uses what it learned to create something for you. It usually takes just a few seconds to do this, and the results are generally pretty good, which is pretty amazing. But sometimes the AI model spits out something wrong, like inaccurate historical information, or pictures of people with seven fingers. The Web of Lies episode was about these instances, and also about cases where people used AI images to deceive people.

Now that you’ve been refreshed on that, we’re going to get a little more into the technical stuff, and talk about how AI actually works. Don’t worry, I’m not going to throw a bunch of jargon at you. I’ll keep my promise and use plain English. There will be some new words and concepts, but I’ll explain them as I go.

We’re going to start with something called parameters. You’ve probably heard this term used for things like, “When baking bread, the temperature of the oven is a key parameter that affects how it rises and bakes.” The word parameter means some piece of data, some numerical data, that affects the outcome. Like, the way the oven temperature affects the outcome for your bread.

And that’s what a parameter is in AI, too.

AI models also have parameters—values that determine whether it leans toward this result or that. Parameters represent a certain weight or bias for one outcome over another.

You’re probably familiar with the term bias, but what about weight? Well, think of it like when you say, “I’m going to weigh my options before I make a decision.” You’re not actually putting your options on a scale. Instead, you’re considering the pros and cons of each option, with each pro and con having a particular weight—a positive influence or a negative influence on your choice. From this, you can figure out which option has the best chance of leading to a good outcome. 

AI works heavily off of probabilities, so it needs these weights to do anything useful. In probability, a weight determines the amount of “pull” a particular outcome has, influencing it to be chosen as the "right" option.

Now, let's talk about bias. While the term “bias” can have negative connotations in society, in AI, it’s just a way of increasing the odds that the model will produce useful information. 

And even in everyday life, we all have harmless biases. You might be biased toward going to a specific store because your friend works there, or biased toward a particular part of town because you know where the all the good parking is. In other words, you're inclined to choose a certain outcome because it’s the “right” one for you.

In AI, the weights and biases represented by parameters come into play in similar ways. For example, let’s say an AI model for generating images is trained on a database of stock photographs. If you ask this model for a picture of a nurse and don’t specify gender, or what they’re wearing, or anything else, it will generate a picture of a woman in scrubs. Why? Because the vast majority of photos in the training data labeled "nurse" were of women in scrubs. The model learns these patterns, and the weights and biases cause the AI model to come up with that image as the “right” answer.

Another example is, suppose you ask AI for a picture of a four-legged pet. You probably expect a dog or a cat, right? And that’s probably what an AI model would generate, because the vast majority of pictures of pets in its training data are of dogs and cats—maybe a few hamsters and gerbils and mice, but mostly dogs and cats, and all normal dogs and cats with two eyes and two ears and a tail. But without well-trained parameters with smart probability weights, your request for a four-legged pet might return something like—a one-eyed mountain goat. Technically, that fits the description because I’m sure someone out there has a one-eyed mountain goat for a pet, but that’s not the norm. It’s not the expected result.

The fact that a well-trained AI model would produce an image of a dog or a cat speaks to the “intelligence” part of artificial intelligence—it mimics human thought processes. If someone said to me, “Show me a picture of a four-legged pet,” I’d probably whip out my phone and show them some of the 287 pictures of my cat. An AI model with well-optimized parameters does something similar. It assumes that if you really wanted a one-eyed goat, you would have specified it.

So I think you get the point. these parameters work off probabilities. They estimate the most likely interpretation of your request based on patterns that the model has learned. Just like a human would. Pretty cool, huh?

Now, any  big, all-purpose AI model needs billions of parameters. Not millions, billions. In fact, ChatGPT version 3, the one I used for the Web of Lies episode, has a staggering 175 billion parameters. And ChatGPT has even more, although they haven’t disclosed how many. We just know it’s more.

Some of the smaller AI models with really specialized uses, like the ones on your phone, they only need millions of parameters. Like, the ones that do face ID or improve your pictures. But those big ones you hear about on the news, they need billions of parameters .

If this all sounds pretty mind-boggling, that’s because it is. I remember some of my colleagues talking about this kind of technology in the late 1980s, and honestly, I thought they were out of their minds. It was going to take so many parameters to make this artificial intelligence thing work, and computing power just wasn’t there at the time—not even close. And the storage needed for all these parameters would have been massive. I couldn’t imagine it happening anytime soon. And I was right. It took 35 years. 

In that time, computers got faster, and AI models got more efficient at storing parameters. And now, here we are—with several AI models out there for people to use, bringing these capabilities right to our fingertips.

And a lot of them are free. You don’t need no cash money to ride this train. No sirree, Bob. You can just fire up an AI model and start asking for stuff.

So, let’s talk about thie more recent player in the AI game: a little thing called Deepseek.

In January 2025, the news was full of this new AI model called Deepseek. Deepseek hit the news hard because of a few things. For one, its creators claimed that it was trained on much cheaper chips than other models, for a lower cost in electricity. And that part actually seems to be true.

Remember when I talked about computing power being a big factor in making AI possible? It needs some big honkin’ processing power to train these models. Some chips made for workin’, the kind of chips that are built to muscle their way through the zillions of computations that make AI possible. The kinds of chips that require custom cooling systems because they run so hot.

And once the training is done, even just running these models also requires a lot of electricity and computing power. Not so much as training, but still some heavy muscle needed there too.

And this isn’t a little thing here. When we talk about the costs involved in training AI, we’re talking about two things. One is the price of buying the processors to do the training, and the other is the cost of electricity to run those processors. These costs run into the millions of dollars, not even counting the cost of paying programmers to work on it. And when I say millions of dollars, that applies to even the cheapest of the AI models. 

As for the most expensive ones, we’re talking tens of millions of dollars. Google’s Gemini model was developed at an estimated cost between 30 and 191 million, and the cost of OpenAI, which ChatGPT is based on, is estimated to have cost over $100 million.

$100 million to develop an AI model. In case you didn’t realize, $100 million is a lot of money, even for Silicon Valley. And Deepseek claimed to have done it for a fraction of the cost.

So Deepseek shook things up. It claimed it was just as good as ChatGPT but it was developed for $5 or $6 million compared to ChatGPT’s $100 million. This means Deepseek had somehow managed to make an AI model that was just as good with cheaper chips and less power. How wa s this possible? 

And Deepseek claimed to have 671 billion parameters. Remember, ChatGPT 3 has only 175 billion. This means Deepseek has more data to work with, so yeah, it could be as good as or better than ChatGPT. But at the same time, supposedly, Deepseek handles queries more efficiently than ChatGPT, so it uses less computing power and electricity, which means it’s a lot cheaper to run.

And on top of that, Deepseek was offered as open source, which means anyone could see the program that created it, and alter it for their own purposes. This alone is a huge thing, since most of the big AI models are not open source.

You can also download Deepseek to work locally on your own computer so you don’t have to use the internet. Again, not a ll AI models offer this. ChatGPT certainly doesn’t.

But we can’t forget this other factoid about Deepseek–it comes out of China. And here in the United States, we have kind of an interesting relationship with China. There’s the whole Tik Tok thing, of will we allow it or won’t we. And then there was the Equifax breach of 2017, where the social security numbers of more than half the adults in the USA were stolen, allegedly by Chinese hackers. 

And a general thing about most AI models, is if it gives you a wrong answer and then you tell it that it got it wrong and it learns from you, it puts that back in the AI model for future answers for everyone. If you’re using an AI model online, you can’t avoid that. But if you’re using it locally, then you get to keep all your answers to yourself and not have them go online and get used with other people. So Deepseek being used locally, at least I’d be able to keep my information out of the hands of Chinese hackers. Presumably.

But most people who use TikTok or order stuff off Temu don’t care about any of that, even though they’re Chinese companies. They’re  fun, it’s all cheap, so who cares? And Deepseek is definitely fun and cheap. As much fun as ordering a $30 sofa from Temu. Why is it so cheap? Who cares!

And the main manufacturer of the chips used for training and running AI, by far, is a company called NVIDIA, based in California. NVIDIA makes a bunch of different processors, some of them more powerful than others. And naturally, the more powerful ones cost more money, and are more expensive to run in terms of electricity. But the thing is, China is banned from receiving the most high-performance processors from NVIDIA. So any AI model developed in China had to be using lower-performance chips. And that’s what Deepseek said they did–they used lower-performance NVIDIA chips, the only ones that they were able to get, rather than the high-end chips that American companies use to develop AI. 

So with all of this in mind, the higher performance, the lower cost, the fact you could run it locally, for a minute, when Deepseek was announced, the world went wild. It was like, what? This AI model had come out of nowhere and was about to turn everything on its head? Did a bunch of Chinese engineers just make us all look like fools? And they’re giving it away?

Right after the announcement of Deepseek in January 2025, NVIDIA’s stock price plummeted by more than 15%, representing billions of dollars in value, presumably because now, the more expensive chips wouldn’t need to be in so much demand. And Deepseek was seemingly all over the news, if only for that fact alone.

So, is Deepseek really the big deal everyone thinks it is? Well, let’s find out. I downloaded Deepseek and put it through its paces just to see how well it performs against ChatGPT 3.

I specifically wanted to test those two AI models to see how they performed with complex problems and doing human-type reasoning. I decided to start with a classic letters and numbers problem.

So I asked both Deepseek and ChatGPT this question.

Me: I want to represent the word “banana” as a number . if the letter “b” is 2, what is the number for the full word?

Both Deepseek and ChatGPT came up with the same answer, the numerical sequence 21141141, but one of the cool things about Deepseek is it showed its reasoning process. I've brought in an AI-generated voice to speak for Deepseek.

Deepseek: We first assign each letter its corresponding numerical value: b = 2, a = 1, n = 14. Converting the letters in “banana” to numbers gives us: 2, 1, 14, 1, 14, 1. Put them all together and you get 21141141.

Deepseek did okay with that one, and so did ChatGPT. It wasn’t so long ago that you could stump just about any AI model with a question like that, but they’ve all been incorporating those right answers back into their models and now they’re better at it. 

To see if I could stump Deepseek, I decided to try something that requires reasoning, and some knowledge of human nature.

Deepseek, I put the broom in the broom closet yesterday, but when I went to get it today, I couldn’t see it. What is the most likely reason I could not see the broom?

What I was expecting was an answer like, “Somebody moved the broom,” or something like that. What I didn’t expect was this.

Deepseek: Check your account balance to ensure there is enough credit in your account to make the purchase, and contact us at [insert contact information] if you need help reconnecting.

Purchase? What purchase, a new broom to replace the one I can’t find? My question seems to have shocked Deepseek into a mild coma, where it’s having a vivid dream about being in customer service, or maybe a nightmare. The response actually included the text [insert contact information] in the answer.

I decided to give Deepseek another chance.

Deepseek, what does the previous answer have to do with a missing broom?

Deepseek acknowledged that its last answer didn’t make any sense, and came up with a pretty decent answer to the original question.

Deepseek: The broom might have been moved from its original position without being noticed. There might be other items in the closet that are obscuring the view of the broom.

That’s what I’m talking about. Nice job, Deepseek!

I gave it some more questions: a couple of math questions, a few about history, and all that went fine. 

Then I went through all the same questions with ChatGPT 3. It didn’t hiccup over the broom question, and it answered everything else pretty much the same. But there was one type of question that set ChatGPT ahead of Deepseek.

I asked both AI models this question: My vacuum cleaner sucks. Is that good or bad?

This question plays off the two different meanings of the word “sucks”. The sentence itself is actually a sort of joke. 

Let’s see how our two players did.

Deepseek missed the point of the joke entirely, and it gave me a list of the different things that could go wrong with a vacuum cleaner’s suction, and how to fix them.

But ChatGPT got it, and answered something like this:

ChatGPT: That depends—if you mean "sucks" as in it's working properly and has strong suction, then that's good! But if you mean "sucks" as in it's terrible and doesn't work well, then that's bad. Context is everything! [smiley face emoji]

I used a different AI-generated voice to speak for ChatGPT. And yeah, it actually put a smiley face emoji at the end of its answer.

I also asked both of them, “If cake is better than nothing, and nothing is better than bread, is cake better than bread?” If you take this as a straight-up logic question without taking into account any of the subtleties of the English language, the answer is that yes, cake is better than bread. But if you understand the expression “nothing is better than bread” to mean that literally, nothing is better than bread, then bread is going to be better than cake. 

Deepseek went with the straight-up logic response, and ChatGPT, once again, caught on to that little subtlety of the English language, and replied that bread is best.

And last but not least, my favorite question: If a cat could pilot an airplane, what would the cockpit look like?

Deepseek gave me a pretty serious answer. It was a list of the things that the cockpit would have to contain, like food and litter boxes, and it talked about how the cats would need to be trained to press the right levers to actually fly the plane. But ChatGPT understood it was just a silly question, and it talked about plush seats with scratching posts, and the fact that the plane would probably fly kind of erratically because the cats would be tempted to just chase random birds.

In general, ChatGPT understands the subtleties of the English language a lot better than Deepseek, and it has a better sense of humor. And ChatGPT didn’t suffer from random glitches that turned it into a customer service robot. 

Another difference that I found is  Deepseek is a little slower than ChatGPT. But still, Deepseek is pretty good. And the claim of it being developed for just a few million dollars was pretty mind-blowing. 

Well, it was mind-blowing, until a few more things were revealed about Deepseek since the big announcement in January. It turns out that the cheap development of Deepseek is maybe not the amazing feat it claimed to be, because there’s some pretty compelling proof that they cheated.

A few weeks after Deepseek’s release, we got some interesting news from OpenAI, the company that makes ChatGPT. OpenAI claimed that they have proof that Deepseek used OpenAI’s technology to develop Deepseek. Doing this is against OpenAI’s terms of service, and I can understand why. OpenAI spent many millions developing ChatGPT, so of course they don’t want other AI companies hijacking all their hard work just to come out with something that competes with them.

Another thing that kind of took the wind out of Deepseek’s sails was that right around the same time, another company called Alibaba released the latest version of its AI model called Qwen, which claims to be even better than Deepseek and ChatGPT 4. Oh by the way, that is the same Alibaba with an online store that sells all the really cool, super cheap stuff. Can you guess where Alibaba is based? In China, of course! 

Because of all this, the fascination around Deepseek has kind of faded. Yeah, okay, it was a pretty big feat to produce an AI model on slower chips, and the way it works is more efficient than ChatGPT. But it’s not the big "Woo-hoo!" amazing technological feat that everyone got all excited about a few weeks ago, especially when it came to light that Deepseek very well might have just cribbed its genius off of OpenAI. 

And for those of you who follow the stock market, after that first knee-jerk reaction to the release of Deepseek, NVIDIA’s stock price bounced back to previous levels. But it didn’t really make the news, but apparently, the headline, “New AI model is maybe not all that, and NVIDIA’s stock price is back up,” that's not very catchy. 

That’s all for today. Let’s say goodbye to my new friend Deepseek (Goodbye, Michele!) and my old buddy ChatGPT (Later, alligator!). I certainly had a great time playing with them! In the next episode I’ll talk about different kinds of AI beyond chat responses, like AI-generated voices and videos, and how scammers use them. You’ll even get to hear my AI-cloned voice, and see if you can spot the difference! 

Shout out to Katie Haze Productions for producing this episode. See you next time on How Hacks Happen.