We're sunsetting PodQuest on 2025-07-28. Thank you for your support!
Export Podcast Subscriptions
cover of episode Researchers Create a Reasoning Model for Less Than $50 - DTNSB 4951

Researchers Create a Reasoning Model for Less Than $50 - DTNSB 4951

2025/2/6
logo of podcast Daily Tech News Show

Daily Tech News Show

AI Chapters Transcript
Chapters
This chapter explores the creation of an AI reasoning model, S1, trained for under $50 using a distillation method. The researchers leveraged Google's Gemini 2.0 model and implemented a clever waiting mechanism to enhance accuracy. Ethical considerations regarding intellectual property and Google's terms of service are also discussed.
  • AI model S1 trained for under $50 using distillation
  • Leveraged Google's Gemini 2.0 model
  • Implemented a waiting mechanism to improve accuracy
  • Ethical concerns about intellectual property and Google's terms of service

Shownotes Transcript

Enjoy a brilliant sleep experience with Soundcore from Anker. Stressed out by your partner's snoring? Having trouble falling asleep? Waking up too easily? Suffering from poor quality sleep? Now, put on Soundcore Sleep A20 earbuds. Experience unparalleled pressure-free comfort perfect for side sleepers. Choose your favorite sound in your curated playlist. Feel your body getting lighter and lighter and enjoy a full night of peaceful sleep with the A20's long-lasting battery.

Then wake up feeling fresh with a personal built-in alarm. Get the sleep you deserve with Soundcore Sleep A20 earbuds. Discover more on soundcore.com. S-O-U-N-D-C-O-R-E. Soundcore. Use code SLEEP at checkout to get $30 off. S-L-E-E-P in all caps.

Getting engaged can be stressful. Getting the right ring won't be at BlueNile.com. The jewelers at BlueNile.com have sparkled down to a science with beautiful lab-grown diamonds worthy of your most brilliant moments.

Their lab-grown diamonds are independently graded and guaranteed identical to natural diamonds and ready to ship to your door. Get $50 off your purchase of $500 or more with code LISTEN at BlueNile.com. That's BlueNile.com, code LISTEN for $50 off.

Ryan Reynolds here from Mint Mobile. I don't know if you knew this, but anyone can get the same premium wireless for $15 a month plan that I've been enjoying. It's not just for celebrities. So do like I did and have one of your assistants assistants switch you to Mint Mobile today. I'm

I'm told it's super easy to do at mintmobile.com slash switch. Upfront payment of $45 for three-month plan, equivalent to $15 per month required. Intro rate first three months only, then full price plan options available. Taxes and fees extra. See full terms at mintmobile.com.

Picture this: you're in the garage, hands covered in grease, just finished up tuning your engine with a part you found on eBay. And you realize, "You know what? I could also use new brakes." So where do you go next? Back to eBay. And you've got eBay Guaranteed Fit. You order a part, and if it doesn't fit, send it back. Simple as that. So when you dive into your next car project, start with eBay. All the parts you need at prices you'll love.

This is the Daily Tech News for Thursday, February 6, 2025. We tell you what you need to know, follow up on the context of those stories, and help each other understand. Today, James McGuff tells us all about IFA and more from your emails. I'm Jason Howell. I'm Shannon Morse. Let's start with what you need to know with a big story. ♪

A research paper released late last week demonstrates how Stanford and University of Washington AI researchers created and trained an AI model that's capable of reasoning. Reasoning is all the rage right now. In a similar way to OpenAI's recently announced O1 and then, of course, DeepSeek's R1 models.

In this case, the model was trained for less than $50 in cloud compute credits through the process of distillation. Now, distillation is where you essentially are sharpening the reasoning abilities through analyzing the answers of Google's, in this case, Google's Gemini 2.0 Flash Thinking Experimental Model model.

So it's really using the input and the output of another model to train your own. That's essentially what distillation is. Using 1,000 carefully curated questions and answers, the researchers were able to train S1, which is the name of their model, in 30 minutes, also using 16 H100 GPUs. So not a whole lot of hardware and definitely not a whole lot of time or cost.

Researchers also implemented a clever trick to enhance the reasoning capabilities of the model by asking each process to wait in order to afford it more time for it to arrive to improved responses and increased accuracy, hopefully at the end of the day, giving it the time to kind of think through these things and really not rush through the processing of them. Now, one thing this points to is

is the commoditization of AI models and really the impact that that is having very quickly on the industry. We've got high performing models coupled with low investments,

large-scale operations are already starting to feel the pressure of this rapidly changing landscape that doesn't automatically mean, as we've kind of learned over the last couple of years, that larger is better or more expensive is better. Now it's really kind of coming down. It's democratizing. And the flip side of this, of course, is the fact that models like this weren't created from the ground up with their own resources, right? Through that distillation process, they actually rely upon

the expensive training and resources of other models to get to where they are. And that has its own ethical challenges that really toe the line of things like intellectual property rights and reverse engineering practices that might actually be prohibited. And in this case, with Google, their terms of service forbid this sort of direct training from its models, although Google has not commented on

on this research yet. So democratization of AI models is really what this kind of screams out to me. Suddenly you are in a position, Shannon, where you can create a snubs AI model if you want, and you don't have to have multiple billions of dollars of these massive, you know, this massive kind of army of GPUs in order to do it. There are other methods. There are other solutions for you. What do you think about that?

You know, I'm kind of excited about that because like even from just a privacy perspective, like being able to build my own AI models sounds so exciting to me. And the fact that it's getting cheaper too is really interesting because I really hope that that helps from like a competitive market. Yeah.

In terms of drawing down the prices of different AI models and subscriptions for consumers. So hopefully this will be a positive thing going forward for all the consumers out there who are interested in AI modeling.

Yeah. And I mean, you know, of course, we also have the open source aspect. Yes. You know, which isn't necessarily, I think, a part of this story in particular, but it also speaks to kind of the wider kind of movement in AI. You know, the last couple of years have been

So much progress so quickly. It's taken the company OpenAI and made it a household name in a short amount of time. And they've, to many people, become the kind of the poster board example of

what it takes to succeed in AI. It takes all that money and that power and that fame and everything. And now we've got open source. Now we've got this process of distillation that's showing how models like DeepSeek, like S1, can suddenly, you know, kind of,

lower the barrier of entry for people. But there are those ethical considerations, right? Like Google's terms of service, I think, from my understanding, kind of forbid this direct training from its model. And that's a huge barrier that I don't know that we know the answer to that yet as far as how they're going to respond here. Yeah, that can be a major concern. And I'm just hoping that it goes the way of...

offering a lot more options. One of the things I really like about the way that these researchers are doing their modeling is they're increasing that wait time. So you're giving the AI model more of a chance to understand context of a situation or context of a prompt, which is kind of like something that we do as humans and is missing from a lot of the

more expensive AI models that I've been using as a consumer and as a content creator. So it will be really cool to see like, will this introduce more patients for consumers? Yeah, crank up the patience meter. Yeah. No, it is interesting because I know that you, you know, we were talking before recording this show and I do the same thing. You know, we're both

content creators. And as part of what we do, there are aspects of what we do that are made at least a little bit easier, a little bit quicker, whatever you want to put it, by leveraging the power and the accessibility of AI models like this. And yeah, and that's a really useful tool. And so this just kind of has my mind going as far as like,

When I use it, when I use those tools, I do notice how sometimes some of the tools are so fast to spit out the answer. And then now what I'm seeing with reasoning tools, and I'm starting to use them a little bit more in some of my research work,

is that it does kind of slow things down. And from, I don't know if it's right or wrong or what, but when I see that it's taking its time, I kind of feel better about it. Like I feel like, oh, well, at least it's not just giving me the first piece of junk that it comes back with. Like, you know, and some of them are actually showing you the thought process, thought in air quotes, because it's not actually thinking, but you know what I mean? There's something about that from a human perspective when I'm using it, where I'm like, okay, you're taking your time and that's all right.

I'm okay with that. I'm okay with that. So anyways, we'll see what this turns out to be. I am very curious to hear what Google has to say about this, considering it's just research. It's not just a bunch of kids going rogue out on the internet. Well, DTNS is made possible by you, the listener. Special thanks to Carmine Bailey, Vince Power, and Chris Beneteau. Thank you.

Picture this: you're in the garage, hands covered in grease, just finished up tuning your engine with a part you found on eBay, and you realize, you know what?

I can also use new brakes. So where do you go next? Back to eBay. You can find anything there. It's unreal. Wipers, headlights, even cold air intakes. It's all there. And you've got eBay guaranteed fit. You order a part, and if it doesn't fit, send it back. Simple as that. Look, DIY fixes can be major. Doesn't matter if it's just maintenance or a major mod. You got it. A

especially when things are guaranteed to fit. So when you dive into your next car project, start with eBay. All the parts you need at prices you'll love. Guaranteed to fit every time. eBay. Things people love.

Enjoy a brilliant sleep experience with Soundcore from Anker. Stressed out by your partner's snoring? Having trouble falling asleep? Waking up too easily? Suffering from poor quality sleep? Now, put on Soundcore Sleep A20 earbuds. Experience unparalleled pressure-free comfort perfect for side sleepers. Choose your favorite sound in your curated playlist. Feel your body getting lighter and lighter and enjoy a full night of peaceful sleep with the A20's long-lasting battery.

Then wake up feeling fresh with a personal built-in alarm. Get the sleep you deserve with Soundcore Sleep A20 earbuds. Discover more on soundcore.com. S-O-U-N-D-C-O-R-E. Soundcore. Use code SLEEP at checkout to get $30 off. S-L-E-E-P in all caps.

Ryan Reynolds here from Mint Mobile with a message for everyone paying big wireless way too much. Please, for the love of everything good in this world, stop. With Mint, you can get premium wireless for just $15 a month. Of course, if you enjoy overpaying, no judgments, but that's weird. Okay, one judgment.

Anyway, give it a try at mintmobile.com slash switch. Upfront payment of $45 for three-month plan equivalent to $15 per month required. Intro rate first three months only, then full price plan options available. Taxes and fees extra. See full terms at mintmobile.com. Worried about what ingredients are hiding in your groceries? Let us take the guesswork out. We're Thrive Market, the online grocery store with the highest quality standards in the industry. We restrict 1,000 plus ingredients.

so you can trust that you'll only find the best high-quality organic and sustainable brands all free of the junk. With savings up to 30% off and fast carbon-neutral shipping, you get top-trusted groceries at your door, and you can stop worrying about what your kids get their hands on. Start shopping at thrivemarket.com slash podcast for 30% off your first order and a free gift.

Yep.

Yep, many Warby Parker locations also offer eye exams. So the next time you need glasses, sunglasses, contact lenses, or a new prescription, you know where to look. To find a Warby Parker store near you or to book an eye exam, head over to warbyparker.com slash retail. All right, there is more we need to know today. So why don't we get right to the briefs. Cool.

Qualcomm reported stronger than expected first quarter sales of $11.67 billion, driven by healthy demand for smartphone chips with AI features, as well as expansion into PCs through partnerships with Samsung and Microsoft.

However, shares dropped 4.8% after the announcement due to concerns over flat 2025 revenue forecast for its patent licensing business, which faced some uncertainty after a key Huawei agreement expired.

On a positive note, the CEO, Cristiano Amon, announced that Arm Holdings has withdrawn its October 2022 threat to terminate Qualcomm's license agreement after Qualcomm had scored a partial victory in a December trial regarding PC chip licensing. Arm is seeking a new trial for unresolved issues. So then you're saying it's still going on. It's not done yet. Yes, it is.

DeepSeek said it would limit access to its API due to server capacity shortages stemming from overwhelming demand and interest. After last month's sudden rise in usage, a posting on the site explains the suspension of the ability for users to recharge API credits was necessary to avoid negative impacts to the service.

The company did assure that any existing values, if you happen to have any values in there, those would continue to be usable. The company also announced the end of discounts for model access slated for February 8th. That's two days from now. Hi, it's me. I'm the problem. I'm one of those people that downloaded it. I downloaded it too. I used it like a couple of times. Oh, we both did. Yeah, you got to check it out.

Open AI announced on X that it has removed the login requirement for users looking to use its search functionality after expanding the product for use by all online viewers back in December. Visitors to chatgpt.com will soon see the option to search in the query box by default, though the change will not apply to the mobile app.

Yeah, search. Coming for you, Google. And I use Perplexity for the AI search thing, and I'm super curious now that it's kind of opened up to play around with it, too, and see if you're not using it. It's a pretty compelling kind of feature set. Oh, okay. Worth checking out.

Google announced yesterday. Wow, this is a very AI-heavy day today. So if you don't like AI, I apologize. Google announced yesterday that it is releasing its big next-gen chatbot effort, the Gemini 2.0 family of models. Gemini 2.0 Pro with proficiency in coding and prompt complexity gets an experimental release, while 2.0 Flash and Flash Lite—get it—

are now available for developers inside Google AI Studio and Vertex AI. Google says 2.0 Flash Lite is its most cost-efficient model to date. And then finally, and we talked about this a little bit in our big story, Gemini 2.0 Flash Thinking Experimental is now available inside the Gemini app.

So

YouTube. Okay, here's some non-AI news. YouTube is pursuing TikTok creators through targeted advertising campaigns on that rival platform, capitalizing on TikTok's uncertain future here in the United States. YouTube has been spotted running sponsored posts on TikTok featuring testimonials from creators like, and I quote, get on YouTube, find your community, live your best life. Yeah.

While showcasing merchandise received from YouTube. And I have to ask YouTube, I've been on the platform for 16 years. Where's my merch? I haven't gotten any merch. Totally. I read that too. I was like, wait a minute. Where is this coming from? How do you get on that? I want merch. That sounds nice. Interesting stuff there. Wow.

Huawei's earnings reveal a 22% annual revenue increase, the fastest growth the company has seen since 2016. That's according to CNBC's calculations. All of this despite major U.S. restrictions, you may remember that, back in 2019 that halted the company from accessing U.S. tech services.

I wondered back then if that was going to signal the start of the end for Huawei, but I'm obviously wrong. Huawei credits a return to growth for its consumer business and, quote, rapid development in its car business.

Oh, that's interesting. That's going to be fun. Boston Dynamics announced a partnership with the Robotics and AI Institute, which was formerly known as the AI Institute, to bring enhanced reinforcement learning to its

Electric Atlas Humanoid Robot. So this partnership will aim to improve Atlas's dynamics and generalized mobility. Boston Dynamics and RAI have collaborated in the past, including the Spots Reinforcement Learning Researcher Kit released last year that increased Spots locomotion speed to a record of 11.5 miles per hour.

Go Boston Dynamics, go. Always interested to see what they come up with. It's always very fascinating to watch. The Asus Zenfone 12 Ultra launches today. It's a powerful alternative to the company's gaming-focused ROG, R-O-G, I still never know how to pronounce it, Phone 9 Pro, that will not see a release in the U.S. So you will not be able to get the Asus Zenfone 12 Ultra in the U.S., unfortunately.

But you might want it. The phone is powered by the Snapdragon 8 Elite chipset, has a massive 5,500 milliamp hour battery and a 6.78 inch LTPO display with variable refresh to 120 hertz, of course. The camera system has a really neat hardware based image stabilization system. It's not the first time that they've included this. They've been iterating it over time, but it's really, really capable.

And what premium smartphone in 2025 would be complete without AI features for things like removing background noise from videos, voice memo transcription, and summarization of documents? The Zenfone 12 Ultra starts at 1,099 euros. And you're right there with ROG. It's Republic of Gamers. Although technically, I've hung out with a lot of Asus PR folks, and sometimes they say ROG too, so so do I. So I think it's fine. I mean, it just feels right. The ROG. Yes.

This is my ROG. But yeah, ROG, Republic of Gamers. I will continue to not say it correctly. It's no matter how many times I'm correct, I never get it right. With the Super Bowl scheduled for this Sunday, Google had planned to run an ad for its Gemini AI tool that features a cheese merchant in Wisconsin using the model to write a product description for Gouda.

Now, in it, the model informs him of Gouda's popularity, claiming that it accounts for 50 to 60 percent of global cheese consumption. And a blogger on X pointed out the falsity of that, quote, fact. Google ultimately removed the error from the ad, but they made sure to point out that in this case, the false data was not the result of a hallucination on Gemini's part. Rather, it was an example of a fact that false data exists.

does in fact exist on the web. This is not the first cheese-related snafu coming from Google AI, as it once told users, and I saw the screenshots of this one, to use non-toxic glue to improve pizza. Yep, that was the thing. Cheese is the undoing of Google. At the end of the book, if Google ever goes out of business, the last chapter will just be called cheese, the final one here. Cheese.

I would read that book. These are the essentials for today. I hope you enjoyed them. Let's dive a little bit deeper in the ongoing stories and follow up.

So the Consumer Electronics Show, we were both at CES last month. It's but a fading memory at this point. But we've got some important stuff from CES. Tom noticed another tech convention had a booth there. So he decided to talk to the folks at Europe's biggest tech conference, IFA, about what it is and why folks at CES or even not at CES might be interested in it. I'm here with James McGuff, executive director of IFA.

I know a lot of people in my audience know the Consumer Electronics Show. They know CES. But I know I have a lot of Europeans in my audience who get mad that we're so American-centric. Absolutely. So I wanted to talk to you about IFA and give a lot of the non-Europeans in our audience a chance to get familiar with it if they're not already. So tell me what IFA is to begin with. So IFA is...

It's similar to CES, but probably one of the oldest standing consumer electronics, but home appliances shows in the world as well. So we're dating back to 100 years old as well. It's a show that brings together about 215,000 attendees. But unlike CES, it's not just about the industry and the trade. We also bring in the consumer audience as well. So it's a chance for...

for manufacturers and the latest brands to be able to showcase their products. We'll also talk about the future of where their brands are going, not just to the trade, but also to consumers who might want to touch and feel their products. Oh, fantastic. And I know a lot of people in the audience would love to come to CES if they could. Yeah.

if they wanted to attend, but they could. Anybody can go. Yeah, absolutely. I mean, the price point isn't high. It's like 20 euros for a consumer ticket, 40 euros for a trade ticket. So it gives people a chance to be able to come in and choose their own experience, whether you want to just experience the exhibition. We have a music festival that takes place as well. A load of activities that happen throughout the city of Berlin. So yeah, like being in Vegas, but

I guess another Sin City being in Berlin in Europe. Interesting. Interesting way to put it. So, yeah, you get an excuse to go to Berlin. Yeah, absolutely. You also have so many exhibitors. What kinds of different things would people see there than they would see at CES? Yeah, good question. So I guess a big core of IFA starts with home and lifestyle and kitchenware. So we have a lot more white goods within IFA than you would see at CES.

But increasingly and over the years, that's expanded into having a lot more consumer electronics. So moving into computing and gaming and audio, content creation is massive at the moment. We see that here as well, but it's the same as where we are. So, yeah, I would just say that it's probably –

bit more on the white goods, home side, and I guess sort of laterally into consumer electronics. And you have big keynotes and big announcements and things like that. Can you give me a few examples of big announcements that were made there in the past? Yeah, I mean, well, Shark Ninja launched a load of their new product ranges last year with an EFA. We had the Chancellor of Germany talking about how

you know, sort of IFA had been a bedrock of innovation over 100 years. We had the ALA Flying Cars, you know, so some of these more conceptual futuristic things. So everything from your sort of maybe some more of your big German brands like Miele and Bosch, and they're talking about the announcements that they're doing, all the way through to something that's still a conceptual stage like ALA Flying Cars. We'll have it all under one roof. Fantastic. If people were to want to come,

Whether they're on the trade side or consumer side, where would they go? When is it? How would they find out? Yeah, the show is taking place at the first week of September. So just after the summer. And all you need to go is to e4berlin.com and you'll be able to register your tickets. Registration will probably go live in around about March for international attendees. Some people need visas. But largely the campaign starts around April, May time. Fantastic. And e4...

is IFA. IFA. Innovation rule. Yeah, it's important for people to know that. Thank you so much for chatting. Nice to meet you. Nice to meet you too. All the best. Cheers. Now, if you had feedback about anything that gets brought up on the show, you can always reach out to us on the socials. We are DTNS Show on X, on Instagram, on Threads, Blue Sky, Mastodon. You can find us all on all the places.

Enjoy a brilliant sleep experience with Soundcore from Anker. Stressed out by your partner's snoring? Having trouble falling asleep? Waking up too easily? Suffering from poor quality sleep? Now, put on Soundcore Sleep A20 earbuds. Experience unparalleled pressure-free comfort perfect for side sleepers.

Choose your favorite sound in your curated playlist. Feel your body getting lighter and lighter and enjoy a full night of peaceful sleep with the A20's long-lasting battery. Then wake up feeling fresh with a personal built-in alarm. Get the sleep you deserve with Soundcore Sleep A20 Earbuds. Discover more on Soundcore.com. S-O-U-N-D-C-O-R-E. Soundcore. Use code SLEEP at checkout to get $30 off. S-L-E-E-P in all caps.

Ryan Reynolds here from Mint Mobile. I don't know if you knew this, but anyone can get the same premium wireless for $15 a month plan that I've been enjoying. It's not just for celebrities. So do like I did and have one of your assistant's assistants switch you to Mint Mobile today.

I'm told it's super easy to do at mintmobile.com slash switch. Upfront payment of $45 for three-month plan equivalent to $15 per month required. Intro rate first three months only, then full price plan options available. Taxes and fees extra. See full terms at mintmobile.com.

Now, we end every episode of DTNS with some shared wisdom. We had an overwhelming number of responses to the big story yesterday explaining why we weren't covering some of the stories coming out of the U.S. government. This was the first response right here that we're going to read that we got. And we think it's very representative of many of the responses that came in. We'll keep them anonymous just in case. And Shannon, tell us what they wrote.

They wrote, just finished listening to today's DTNS. I think you are perfectly justified and right on target with your reporting of recent events. Keep it up. It's your show. You are the boss and you can take it any direction you feel. Of course, you can't make everyone happy all of the time. Justin is on a lot and helps with the political stuff. I trust your judgment and your reporting. Still worth every penny.

penny personal anecdote I work for VA and I got chastised when I marked the first email from that server as spam or phishing great job so

Thank you for sending that in. Yeah, I think, yeah, what can I say? It's a real challenge. I know Tom probably, you know, definitely talked about this at length and everything. I think from my perspective, in the world of technology and politics and everything, it can be a little challenging to determine where that fine line is between the two. So we'll just continue doing our best. Yes. We hope you understand.

Thanks to James McGuff and everyone, including today's anonymous feedback who wrote in about yesterday's big story. We appreciate you for contributing to today's show and the show at large. And thank you for being along for Daily Tech News Show. The show is made possible by our patrons at patreon.com slash DTNS. DTNS does have a live version. Don't want to miss it. Called DTNS Live. That's on YouTube and Twitch.com.

You can find details on that and more on DailyTechNewsShow.com. We will talk to you tomorrow. The DTNS family of podcasts, helping each other understand. Diamond Club hopes you have enjoyed this program.

Enjoy a brilliant sleep experience with Soundcore from Anker. Stressed out by your partner's snoring? Having trouble falling asleep? Waking up too easily? Suffering from poor quality sleep? Now, put on Soundcore Sleep A20 earbuds. Experience unparalleled pressure-free comfort perfect for side sleepers. Choose your favorite sound in your curated playlist. Feel your body getting lighter and lighter and enjoy a full night of peaceful sleep with the A20's long-lasting battery.

Then wake up feeling fresh with a personal built-in alarm. Get the sleep you deserve with Soundcore Sleep A20 earbuds. Discover more on soundcore.com. S-O-U-N-D-C-O-R-E. Soundcore. Use code SLEEP at checkout to get $30 off. S-L-E-E-P in all caps.

Worried about what ingredients are hiding in your groceries? Let us take the guesswork out. We're Thrive Market, the online grocery store with the highest quality standards in the industry. We restrict 1,000 plus ingredients so you can trust that you'll only find the best high quality organic and sustainable brands all free of the junk. With savings up to 30% off and fast carbon neutral shipping, you get top trusted groceries at your door and you can stop worrying about what your kids get their hands on.

Start shopping at thrivemarket.com slash podcast for 30% off your first order and a free gift.