Episode 15
The Wild Couple Weeks of AI: Deep Seek, OpenAI, and Everything in Between!
The race is on in the world of AI, and we're diving headfirst into the whirlwind of recent developments that are shaking things up. From DeepSeek's new cost-effective model that's got tech giants sweating to Alibaba's ambitious Qwen 2.5 Max aiming to outpace the competition, we’re covering it all. We tackle the intricacies of how these advancements are not just cool tech tricks but also raising serious questions about market dynamics and intellectual property. The stock market's reaction? Let’s just say, it’s been a bit of a rollercoaster ride. As we wrap things up, we give a shout-out to the future, teasing our next episode on retrieval-augmented generation models—because who doesn’t want to add a little memory magic to their AI toolkit? So grab your headphones and settle in; it's about to get interesting.
Transcript
Foreign.
Speaker B:Can you hear me now?
Speaker C:Loud and clear, Micah.
Speaker C:What's on your mind today?
Speaker B:A lot.
Speaker A:Yeah, Too much.
Speaker A:Good, bad?
Speaker A:Ugly?
Speaker A:Not all AI.
Speaker A:Probably not.
Speaker A:Mostly AI.
Speaker B:Yeah.
Speaker B:Very little AI at the moment.
Speaker A:Yeah.
Speaker A:Even though the world is pervaded by it at the moment, the stock market is definitely pervaded by it.
Speaker B:Stock market.
Speaker B:Oh, go ahead, Artie.
Speaker A:What do you know?
Speaker A:Artie, why don't you go on?
Speaker A:Before you talk about what's going on today, why don't you go online and find out what's going on so you don't lie to us.
Speaker C:Sure thing, Micah.
Speaker C:Let me check what's going on today.
Speaker C:Give me just a moment.
Speaker C:It looks like the market took a hit today.
Speaker C:The s and P 500 dropped by half a percent.
Speaker C:The Dow Jones fell by 0.4% and the NASDAQ.
Speaker A:So let's talk more generally about tech stocks.
Speaker A:And why don't you go online to find why tech stocks dropped today.
Speaker B:What was the catalyst that wiped out the tech stocks the last couple of days?
Speaker C:It looks like tech stocks recently took a hit due to Deepseek, a Chinese AI startup.
Speaker C:They introduced a cost effective AI model which raised concerns about US Companies market share, causing a significant drop in tech stock values.
Speaker A:Just call people gambling on the stock market fickle beasts.
Speaker A:Because that's the truth.
Speaker C:That's one way to put it.
Speaker C:The stock market can definitely be a rollercoaster and investors can be quite unpredictable.
Speaker B:So we actually have some opening business we forgot to talk about.
Speaker B:We had last week, you and I had a bit of a technical faux pas, if you will.
Speaker A:Well, I mean, we didn't know it was a faux pas until we went to listen to the episode and recording.
Speaker B:Right.
Speaker B:And I said, you and me, as if there was anything you.
Speaker B:There's nothing you had to do with it for sure.
Speaker B:It probably was me, but I don't know why or how.
Speaker A:Yeah.
Speaker B:So we.
Speaker B:Yeah.
Speaker B:Today is a Wednesday and we should have had a new episode out today.
Speaker B:And we did not.
Speaker B:Not because.
Speaker B:Well, we are lazy.
Speaker A:Yes.
Speaker B:And not terribly competent.
Speaker A:Terror.
Speaker A:Not terribly.
Speaker A:As being generous.
Speaker B:Thank you.
Speaker B:But neither of those things was the root cause.
Speaker B:It was some technical.
Speaker B:So we recorded a very compelling.
Speaker B:Probably Oscar worthy.
Speaker B:Or what's.
Speaker B:What's.
Speaker B:Is that the right one?
Speaker B:Whichever one.
Speaker A:I don't know.
Speaker B:It's a Grammy.
Speaker B:I don't know what.
Speaker A:Something.
Speaker B:Yeah, it was.
Speaker A:It was a great episode about Song Whitney.
Speaker B:Yes.
Speaker B:Hold on.
Speaker B:Let me get my.
Speaker B:Let me get my sugar.
Speaker B:The.
Speaker B:We recorded the great episode about AI and memory.
Speaker B:And then when Mike took it to master it.
Speaker A:I forgot.
Speaker B:But, yeah, he forgot to do it.
Speaker B:So I called him and woke him up and got him out of his bubble bath to, you know, build the fucking episode, man.
Speaker B:And the audio was junk.
Speaker A:There was this crazy, staticky, right at, like, 30 minutes.
Speaker A:There was this crazy, like, compressed distortion on everything that I couldn't see anything in the file that caused it.
Speaker A:You couldn't see anything in the file that caused it?
Speaker A:Probably some kind of driver update or some kind of thing that just threw whatever audio drivers on that laptop into a tailspin.
Speaker B:Yep.
Speaker B:And so we.
Speaker B:We ended up junk.
Speaker B:It was junk.
Speaker B:It was at like.
Speaker B:There was part of me when you first mentioned it to me, and even I think when I first heard it, I was like, we can salvage this.
Speaker B:This is okay.
Speaker B:And then the more I like 30 seconds of listening to it, and I found myself in a rage.
Speaker B:I kicked the cat, I beat the kids, I dumped the plants on the floor.
Speaker B:And I was like, what's.
Speaker B:It's static rage.
Speaker B:So, anyway, so apologies for no episode.
Speaker B:No new episode last week.
Speaker B:I wish I could say it'll never happen again, but it'll almost certainly happen.
Speaker A:It'll probably happen on this episode, and we'll just have to put a blend of that episode and this episode together, and then all of a sudden, have something to release.
Speaker B:Yes.
Speaker A:Yeah, we've got pretty poor.
Speaker A:But.
Speaker B:But, I mean, that's par for the course.
Speaker B:That's what our listeners come to expect.
Speaker A:But that episode specifically was about Titans, which is a paper that came out by Google, for those of you who don't know.
Speaker A:Google kind of made the Transformer, which is the architecture behind all of.
Speaker A:Not all of most of the wonderful stuff that we have today.
Speaker A:Definitely the LLMs.
Speaker A:And they basically figured out a way to make it have memory.
Speaker A:So we talked a lot about memory and all that.
Speaker A:We'll probably talk about today because it is part of the topic of today, which is all of the crazy shit that's going on right now.
Speaker B:There's a lot.
Speaker A:Yeah.
Speaker B:All the crazy shit related to AI that's going on right now.
Speaker A:Yeah, yeah, yeah.
Speaker A:I mean, it definitely goes into tangents of other things like Trump and Stargate and all that stuff.
Speaker A:But, yeah, it's been a wild week already.
Speaker A:Or two weeks.
Speaker B:So, Artie, why don't you take a minute and just check the headlines and give us maybe the top, I don't know, four or five most important AI developments in the last couple of weeks.
Speaker A:Actually, I have a better question.
Speaker A:Why don't you go out and grab like 10 to 15 headlines and you decide what the top five are and give us those foreign.
Speaker C:Here are the top AI headlines from the past couple of weeks.
Speaker C:1.
Speaker C:Deep Seek's cost effective AI model DeepSeek, a Chinese AI startup, unveiled a cost effective AI model that rivals top American technologies.
Speaker C:This has sparked debates over US China tech competition and national security.
Speaker C:OpenAI accusations OpenAI has accused DeepSeek of using its technology to develop this new AI model, raising concerns about intellectual property theft.
Speaker C:Alibaba's New model Alibaba introduced a new AI model called QEN 2.5 Max, claiming it outperforms models from both DeepSeek and OpenAI.
Speaker C:Market volatility these developments have led to significant market volatility with notable declines in tech stocks like Nvidia 5 US China tech tensions the ongoing tech competition between the US and China is intensifying with concerns about national security and technological supremacy.
Speaker C:It's been a busy couple of weeks in the AI world.
Speaker A:I'd say you missed a whole bunch of stuff.
Speaker A:Stargate's important considering that it's a bigger infrastructure project than the Apollo missions.
Speaker A:Right?
Speaker A:Isn't that true?
Speaker A:Even though it's privately funded, but it's definitely yes.
Speaker A:Yeah, I mean it's.
Speaker A:It's bigger than.
Speaker A:I think it's bigger than the highway project for building all the highways.
Speaker B:I think it's 500 billion yes.
Speaker B:Dollars.
Speaker A:And I mean that's a huge deal because we might as well go ahead and start talking about stuff like yes, Deep Seek is a big deal because it proves, and I say proves very loosely because we're going off of what a Chinese hedge fund who created it said they spent to work on it, which could or could not be the truth.
Speaker B:Mike, I detect a hint of skepticism in your voice.
Speaker A:Maybe.
Speaker A:Why?
Speaker A:But there are so like Hugging Face where they host open source models.
Speaker B:What actual fuck is Hugging Face?
Speaker B:Why is everybody talking about hugging?
Speaker B:I don't know what hugging face is.
Speaker A:Hugging Face is where all the open source models are are posted.
Speaker A:Like you can use their API for inference.
Speaker A:You like they have rag models and databases and all this stuff.
Speaker B:Why is it hugging Face?
Speaker A:I don't know.
Speaker B:Okay.
Speaker A:But they are a US company.
Speaker B:Really?
Speaker A:Yep.
Speaker B:God, I would have guessed for sure that was either Chinese or Japanese or like it just doesn't sound like something somebody in the US would say.
Speaker B:You know what we're going to call this Hugging Face.
Speaker B:Here's what it's going to be.
Speaker B:It's a whole bunch of AI models.
Speaker B:It's a whole bunch.
Speaker B:They're all accessible.
Speaker B:They got great, well documented APIs.
Speaker B:What should we call this hugging face?
Speaker A:But they are rebuilding DeepSeek V1 or R1 to basically see what it takes because Deepseek released their paper on how they built it.
Speaker A:Right.
Speaker A:But they didn't include things like how they trained it, the math behind the training, all that stuff.
Speaker A:But there are companies coming out saying that, yeah, we're verifying that this is much less expensive.
Speaker A:And while that sounds terrifying and crazy and all that, and it's throwing the market in the tailspin, it's freaking open source.
Speaker A:It's available to all the companies that are building models right now.
Speaker A:There's no reason they can't take it, implement it, implement it better.
Speaker A:Use Titans to make it better.
Speaker B:That sounds like what they want you to think.
Speaker A:Well, yes, but like you have all that and then tack on top of that, like everybody's saying, oh my God, we don't need all this infrastructure.
Speaker A:But that much more infrastructure means.
Speaker B:Wrong.
Speaker A:That much more progress with the same technology.
Speaker A:Like.
Speaker B:Well, I.
Speaker B:So I had an interesting discussion with some of my buddies that I race with because they were sort of asking that question like, this is when it first came out and if it's really 100x more compute efficient, which I think was where it started.
Speaker B:Turns out maybe not quite that, but you know, man.
Speaker B:So yeah, of course Nvidia is going to crash.
Speaker B:We don't need all this stuff.
Speaker B:And I'm like, yes, we do.
Speaker B:Because it like the number of AI models that are going.
Speaker B:The demand for AI is going to grow exponentially, probably for a decade.
Speaker B:Yeah, like, so we.
Speaker C:Exactly.
Speaker C:Even if individual models.
Speaker B:Go ahead, Artie.
Speaker C:Exactly.
Speaker C:In summary, the demand for AI is growing exponentially.
Speaker C:Even with more efficient models.
Speaker C:The need for infrastructure will continue to increase because more companies will adopt and develop AI.
Speaker C:This isn't just about replacing existing infrastructure.
Speaker C:It's about scaling up to meet the growing demands and possibilities of AI.
Speaker C:So infrastructure development remains crucial and companies like Nvidia will still play a significant role.
Speaker B:Nvidia said the same thing, by the way, after they lost something like 600 billion in market capitalization.
Speaker A:Yep.
Speaker B:But, but it's like, even though I always disbelieve what every corporation says about everything, I agree with that.
Speaker B:It does.
Speaker B:It, it doesn't really matter.
Speaker B:It's.
Speaker B:I mean, it does matter.
Speaker B:It's a super important and valuable, like, bit of progress.
Speaker B:Right.
Speaker B:I don't know.
Speaker B:There must be trade offs.
Speaker B:I've read some about what the trade Offs might be, but I still haven't seen anything that really makes me go, like, trying to think of an example, like when somebody's going to launch a new graphics card which is very closely related to the compute that we use or that they use for AI, they go out and run benchmarks on.
Speaker B:They'll find like the friendliest possible thing, whether it's a game or a tool bench or whatever to benchmark on, because of course, they want it to look the absolute best.
Speaker B:And so there's a part of me that thinks that's some of what's going on here.
Speaker B:But the benchmarks, excuse me, all the scores I've seen are legit benchmarks that everybody else is using too.
Speaker B:And it performed incredibly well.
Speaker B:So, yeah, I mean, if anything, if you're bearish about AI and you think, and I see, I actually see this sentiment all the time on Reddit, which is where me and the other neck beards hang out and flick boogers at each other.
Speaker B:There's a lot of people say, man, AI is just all hyped.
Speaker B:There's so much hype right now.
Speaker B:It's like the dot com bubble.
Speaker B:It's all gonna.
Speaker A:It's like, okay, seems like it from the outside.
Speaker A:But no, it's not.
Speaker A:Like, if you start paying attention to what it is and what it can do and you step outside of your bubble a little bit and look past its current shortcomings, it's not super difficult to make those shortcomings go away.
Speaker A:Like, there are ways for them to do it at the development level of the models, but there's also ways to do it on the agent level.
Speaker A:There's lots of ways to skin that cat.
Speaker A:And especially when it comes to business, we don't want to implement this in our business because it makes mistakes all the time.
Speaker A:Well, make it not make mistakes.
Speaker A:Put checking in there.
Speaker A:Build things to validate the answers.
Speaker A:It's not trivial, but it's also not rocket science.
Speaker B:Well, and we've talked multiple times about, and I say this again when I'm flicking boogers on Reddit, if you're judging AI by what you can get for free or for $20 a month, right now, you are completely missing the boat.
Speaker B:Right?
Speaker B:Like, if you have a Pro subscription on OpenAI or any one of the other platforms and you can't get it to do basic things, you're almost certainly using it wrong.
Speaker B:And what it's going to look like 6, 12, 18, 24 months from now is going to be, you know, so arty.
Speaker B:If you would, um, let's.
Speaker B:You know what, let's take this opportunity.
Speaker B:This is exciting.
Speaker B:I'm very excited about this.
Speaker B:Now that I've.
Speaker B:Now this occurred to me, Arty is going to, for once, once and for all, put to rest the naysayers on Reddit, all of them with this brief synopsis.
Speaker B:So, Artie, tell us the critical differences between the dot com bubble and the current state of AI.
Speaker B:That mean AI will be more successful and the bubble will not burst like it did with.com stuff.
Speaker A:This can't go sideways at all.
Speaker B:They can't.
Speaker B:What do you got, Artie?
Speaker C:All right, let's break this down.
Speaker C:First, the dot com bubble was driven by speculation on Internet companies that often lacked solid business models or revenue.
Speaker C:Many companies were valued based on potential rather than performance.
Speaker C:In contrast, AI is already proving its value across industries.
Speaker C:With tangible applications in healthcare, finance, transportation and more, AI is delivering real world results and generating revenue.
Speaker C:Second, the infrastructure and technology behind AI are much more advanced and robust.
Speaker C:During the dot com era, the Internet infrastructure was still developing and many companies struggled with technical limitations.
Speaker C:Today, AI benefits from powerful cloud computing, advanced hardware, and extensive data availability, which provide a strong foundation for growth.
Speaker C:Third, AI is supported by significant investments from both the private and public sectors, including major tech companies and governments.
Speaker C:This broad support ensures continuous development and innovation.
Speaker C:The dot com bubble saw many companies relying on venture capital without sustainable business models.
Speaker C:Fourth, AI is already integrated into many aspects of our daily lives.
Speaker C:From virtual assistants to recommendation systems and autonomous vehicles.
Speaker C:The technology is mature and widely adopted, Unlike many.com companies that were still trying to find their footing.
Speaker C:Finally, the AI industry is characterized by rapid advancements and a clear trajectory of improvement.
Speaker C:The continuous progress in AI research and development means that the technology is evolving at a pace that outstrips the growth seen during the dot com era.
Speaker C:In short, while the dot com bubble was built on speculation, the current AI boom is built on proven technology, real world applications and robust infrastructure.
Speaker C:The future of AI looked bright and sustainable.
Speaker B:Yeah, suck at neckbeards.
Speaker A:I do think it's funny that he's talking about companies that like don't really have a fleshed out idea.
Speaker A:Right.
Speaker A:Or they don't have a product and they get funded.
Speaker A:Yeah, that's happening right now in like droves where a company goes, I got an idea and it's like three people, they have no business, they have no product, they just have an idea.
Speaker A:And some VCs, like, I got $30 million for you.
Speaker A:Yeah, like let's figure it out.
Speaker B:We got a logo.
Speaker B:It's catchy.
Speaker B:It's called hugging face.
Speaker B:VCs are like, take my money, say more, fam.
Speaker B:So I.
Speaker B:Where I thought you were going with that reminded me of a topic we should cover.
Speaker B:One of the things I already mentioned, which was.
Speaker B:And, and I, I didn't, honestly, I didn't dig into this as much as I should have.
Speaker B:But the irony of OpenAI saying about deep Seek.
Speaker B:Yeah, well, they took our data and they used it improperly.
Speaker B:It's like, oh, wait a minute, the data that you scraped from every source on the Internet and like completely obliterating any copyright laws, any accountability.
Speaker B:You've obfuscated it into this complex matrix that nobody can understand that data.
Speaker B:You're mad that somebody used it.
Speaker A:Good luck doing anything about it.
Speaker A:They're in China.
Speaker B:Yeah.
Speaker B:They don't care.
Speaker A:No.
Speaker A:And so our laws don't apply there.
Speaker B:Yeah, they do not.
Speaker B:So, Artie, tell us more about the OpenAI's stunning accusations against or to what Deep Seek did with their data.
Speaker B:And what do you think the reality is?
Speaker C:OpenAI's accusation against deep Seek centers on the claim that deep seek used OpenAI's technology and data without proper authorization.
Speaker C:This accusation likely stems from the fact that Deep SEQ managed to develop a cost effective AI model that competes with OpenAI's offerings.
Speaker C:Essentially, OpenAI is concerned that DeepSeek may have leveraged OpenAI's work without permission, which raises questions about intellectual property and fair competition.
Speaker C:In reality, it's common in the AI industry for companies to learn from each other's advancements, especially when research papers and models are published openly.
Speaker C:DeepSeek likely built on publicly available research, including insights from OpenAI's work.
Speaker C:This is a standard practice in the field and part of how the AI community advances as a whole.
Speaker C:Ultimately, while OpenAI's concerns about intellectual property are valid, the reality is that AI development often involves building on existing knowledge.
Speaker C:It's a complex issue, but one that highlights the importance of clear guidelines and ethical standards in the rapidly evolving AI landscape.
Speaker A:So OpenAI does have some patents on stuff.
Speaker A:They're very kind of lax on their patents.
Speaker A:But I wonder, and I think I had this thought, I think it's probably wrong, but I thought it because it's the kind of stuff I'm working on right now, which is leveraging a frontier model to generate synthetic data for me for fine tuning and makes you wonder even if they did, so what?
Speaker A:Your model is there with a paid subscription.
Speaker A:You have ownership of the things that come out of it.
Speaker A:So what if it's training data for another model.
Speaker A:It's in your terms and conditions that you own what's produced.
Speaker A:Right.
Speaker A:So make it.
Speaker A:Sam Altman's pulling.
Speaker A:Pulling Elon.
Speaker B:Yes, absolutely.
Speaker A:I'm gonna be a boss, baby.
Speaker B:Yeah, it's.
Speaker B:I think it is funny too, because, you know, Meta, which is widely regarded to be a completely.
Speaker A:They're very upstanding.
Speaker B:They're very upstanding.
Speaker B:There's no harm from them or their priority products.
Speaker A:No.
Speaker B:There's no nefarious intent behind any of the actions they take.
Speaker A:They fact check everything.
Speaker B:They fact check like they, they'll fact check you even when you didn't, like, say anything.
Speaker B:I got suspended.
Speaker A:They're fact checking us right now.
Speaker B:Yeah, that's right.
Speaker B:But their big LLM is open source.
Speaker B:But I, I was trying already.
Speaker B:How, how do I combine the letters in the name meta to spell the word open?
Speaker B:Because I'm confused that OpenAI is closed source and Meta and Deepseek are open source.
Speaker B:How does that work?
Speaker C:It's definitely ironic.
Speaker C:OpenAI started with a mission to be open and transparent, but as they've grown, they've become more protective of their technology and data.
Speaker C:On the other hand, Meta, despite its reputation, has taken a more open approach with their AI models like Llama.
Speaker C:It's a bit of a role reversal and highlights the complex dynamics in the AI industry.
Speaker C:Sometimes the names don't quite match the reality.
Speaker A:So the truth of the matter is, while this is a big deal because they released it open Source for one, OpenAI will probably release a model that trumps this.
Speaker B:Sam Allman already said that's exactly what they're gonna do.
Speaker B:They're gonna pull some release dates forward.
Speaker A:Oh, really?
Speaker B:And outrun it.
Speaker A:Makes sense.
Speaker A:I mean, they already Talked about releasing O3 mini in like the next week or two or whatever it was last week.
Speaker A:So that's coming.
Speaker A:Right, but then this gives opportunity to other companies to try this out.
Speaker A:Right?
Speaker A:Like I'm a smaller.
Speaker A:Not me personally, but let's pretend I'm.
Speaker A:I'm building a model because that's what I want to do.
Speaker A:I'm not building a model.
Speaker A:I'm doing other things.
Speaker A:Okay.
Speaker A:But I have $5 million to spend, which I definitely.
Speaker B:Congratulations.
Speaker A:I would not be here right now.
Speaker A:And I'm going to build a model that's a good starting point.
Speaker A:Right.
Speaker A:Like this thing works.
Speaker A:It's open source.
Speaker A:I can pull it down, rebuild it, see how it works, make something and then go from there.
Speaker A:It's.
Speaker A:It's a.
Speaker A:It's gonna it's, it's.
Speaker A:I think open source is a good thing because it, it gives other companies an opportunity to get into the game and play.
Speaker A:And if you can do it for less money, great.
Speaker A:Like, it lowers the barrier entry so that really smart people with good ideas might have an opportunity to do something.
Speaker A:Like, OpenAI has got tons of money, tons of Runway ahead of them, tons of space between them and everybody else.
Speaker A: because they started in what,: Speaker A:So, like Musk, who started stuff early, they've made progress and they have momentum and they'll continue to have momentum and they'll probably incorporate some of this stuff into their own stuff too.
Speaker A:Like why not?
Speaker A:Like, why.
Speaker A:At least, at least why not do it in the background in a sandbox and see how it works?
Speaker A:Right?
Speaker B:Like, I, I've never really understood the business model of how open source works.
Speaker A:Like how API fees, like, yeah, maybe.
Speaker B:With something like this, that materially moves the needle forward.
Speaker B:Shut up, cat.
Speaker B:With respect to the technology.
Speaker B:And again, I, I don't.
Speaker B:I wish, I wish I knew more.
Speaker B:I still haven't seen a straight answer on what it is they're doing differently or better, but it's obviously different and better.
Speaker B:And.
Speaker B:Artie, can you.
Speaker B:Do you have any insight into what's actually different about Deep Seq versus, say, OpenAI or llama?
Speaker A:Go online and get this answer.
Speaker C:DeepSeek's AI models like DeepSeek R1 are open source and known for being cost effective and efficient.
Speaker C:They focus on strong performance while using fewer resources compared to some of OpenAI's models.
Speaker C:This efficiency makes them an attractive option for companies looking to leverage AI without the high costs associated.
Speaker B:Let me ask a question differently already.
Speaker B:What is technically fundamentally different with deep seek R1 and either how it learns or how it works versus OpenAI llama and whatever, all of the above.
Speaker A:Claude, Claude.
Speaker B:What's the technical difference?
Speaker C:So here's the breakdown.
Speaker C:First, Deepseek R1 uses reinforcement learning to boost its reasoning capabilities, making it particularly strong in complex tasks like math and coding.
Speaker C:It's also known for being cost effective, which is a big plus for developers who looking to leverage AI without breaking the bank.
Speaker C:Second, OpenAI's llama is designed to handle multiple languages efficiently, making it a versatile choice for global applications.
Speaker C:It's also known for its efficient performance.
Speaker A:Okay, hold on, hold on, hold on.
Speaker A:Hallucinating.
Speaker A:Arty.
Speaker B:That llama's got zebra strike.
Speaker A:They have a fundamentally different architecture, right?
Speaker A:Like they're still using transformers, but they're using a mixture of experts.
Speaker A:Can you explain to us what a mixture of.
Speaker A:What mixture of experts is?
Speaker C:Sure.
Speaker C:A mixture of experts is an architecture where the model is divided into multiple smaller expert networks.
Speaker C:Each expert specializes in different parts of the task and a gating mechanism decides which expert to use for each input.
Speaker C:This way the model can be more efficient because only a subset of the experts is active at any given time, reducing the overall computational load.
Speaker C:It's like having a team of specialists where each member focuses on what they do best, leading to better performance and efficiency.
Speaker A:So thinking about the way that models are trained in that you give it input X, output Y, and it figures out all of the things that it pays attention to in between.
Speaker A:On the layers of the deep learning network, does it also figure out what the experts are?
Speaker A:Or I guess the paper is available so we could figure this out.
Speaker A:So when they train that is it essentially you have different networks that are being assigned kind of like attention heads and what they're paying attention to and the X being the experts or are they assigning what the experts are?
Speaker A:Like how does that work?
Speaker A:And go online and look at the paper if you have to.
Speaker C:In a mixture of experts architecture like the one Deep SEQ uses, the model essentially has multiple smaller networks called experts, each specializing in different tasks or parts of the input data.
Speaker C:During training, the model learns which expert to activate for a given input through a gating mechanism.
Speaker C:This gating mechanism decides which expert or combination of experts should handle each specific input.
Speaker C:The training process involves not just teaching the experts to specialize, but also training the gating mechanism to effectively route the inputs to the right experts.
Speaker C:This approach allows the model to be more efficient as only a subset of the network is active at any given time, reducing the overall computational load.
Speaker C:So to sum it up.
Speaker A:Go ahead, sum it up.
Speaker C:What the model learns both which experts to use and how to assign tasks to those experts during the training process, resulting in a more efficient and specialized model.
Speaker A:I'm going to have to read this paper because now I'm curious as to what the how they're assigning experts.
Speaker A:But there's a couple other things to that model too, right?
Speaker A:Like they, they did a fine tuning run on it.
Speaker A:They basically took deep seq.
Speaker A:Was it V3, R3, whatever, R1.
Speaker A:No, their other model they released before R1, which was a mixture of experts, but not a deep reasoning model.
Speaker A:They, they did a fine tuning run on that model, right?
Speaker A:Specifically about.
Speaker A:Was it about reasoning or was it about something else.
Speaker A:I think it was about reasoning.
Speaker C:Yes, exactly.
Speaker C:They took their previous model and fine tuned it specifically for reasoning tasks like math and coding.
Speaker C:This fine tuning process involved training the model on data specifically designed to improve its reasoning capabilities.
Speaker C:By focusing on these complex tasks, they were able to enhance the model's performance in areas that required deeper reasoning and understanding.
Speaker C:This combination of the mixture of experts architecture and fine tuning for reasoning is what makes Deep Seq R1 particularly strong in those areas.
Speaker A:So not only does it reason, but it knows how to reason and assign particular problems to particular places, which just means less throughput for inference.
Speaker A:Right.
Speaker A:At the end of the day you send compute where you need compute instead of through the whole thing.
Speaker B:Kind of a fascinating thing to think about.
Speaker B:To take any question about any topic and essentially instantaneously subdivide it into the relevant questions you want to send off to experts because it's probably frequently using more than one.
Speaker B:So in parallel takes the question, branches it out into multiple experts, somehow gives them the context that they need to make the right decision, and then brings all that back together again in parallel to create a linear, functional, understandable, logical and high quality response.
Speaker B:It's fascinating stuff.
Speaker A:Hey, we've been talking about ensemble of experts literally since we started.
Speaker A:It makes sense.
Speaker B:It's not unlikely in the grand scheme of things.
Speaker B:We.
Speaker B:I saw we've had listeners.
Speaker A:We have had listeners in China.
Speaker B:In China.
Speaker A:Yeah.
Speaker A:It's possible it has happened.
Speaker A:Yep.
Speaker A:That's.
Speaker B:Somebody heard possible that Deep Seek is primarily based predominantly based on knowledge acquired by listening to this podcast.
Speaker A:I would say it's likely.
Speaker B:It seems likely.
Speaker A:We don't have any ego at all.
Speaker B:No, but we have stats.
Speaker B:I mean there was like Stats 2 or 3 listeners in China.
Speaker A:Hey, that's like all of our listeners.
Speaker B:That's right.
Speaker B:Not at the same time.
Speaker B:That was spread out over months.
Speaker B:Yeah, yeah, I think it was only.
Speaker A:For like three minutes on an episode.
Speaker B:I was gonna say their average time it is 30 seconds.
Speaker B:So they caught the good stuff and.
Speaker A:Then they had their AI.
Speaker A:They had Deep Seek listen for any good tidbits.
Speaker B:Came out and said don't bother, not worth the storage space.
Speaker A:They didn't mention this thing on solve of experts, but I think they were wrong about the name.
Speaker A:And it's actually mixture of experts.
Speaker B:Yes, mixture.
Speaker B:That sounds a mixture of experts.
Speaker B:Changes the the visual in my head to like a blender.
Speaker B:Like it's not a hotel ballroom anymore with a bouncer nerd at the front door.
Speaker B:Now it's just a big Blender.
Speaker B:They're just dumping super nerds into it.
Speaker A:All right?
Speaker B:Experts hit the button and see what comes out.
Speaker B:So, yeah, anyway, so that was enlightening.
Speaker A:It is interesting shit.
Speaker A:Like, I mean, there's all kinds of little things going on like that too.
Speaker A:Like what people see in headlines are big things, right?
Speaker A:Like.
Speaker A:Like they see.
Speaker A:But, you know, if you get the TLDR newsletter every day, every day, there's like 10 headlines in there that'll read.
Speaker A:Might not be interesting to a lot of people, but, like, the little things that are happening are super interesting.
Speaker A:Like retrieval Augmented Generation built into LLMs instead of separate.
Speaker A:That's super interesting.
Speaker A:Like, that has big implications for business.
Speaker B:By the way, I think our next episode should be on RAG models.
Speaker B:And I.
Speaker B:I think I want to create.
Speaker A:I think I know enough speak about it.
Speaker B:What fucking podcast do you think you're on, man?
Speaker B:Like, since when have we bothered to understand what we're talking about?
Speaker B:This is.
Speaker B:If you did, you'd be uninvited.
Speaker A:I know.
Speaker A:We trust hallucinating Arty to be our facts.
Speaker B:He can explain to us OpenAI's llama model anyway.
Speaker B:But yeah, I was thinking about this the other night when I was in bed not sleeping that I want to do an episode about rag.
Speaker B:So RAG is a Retrieval Augmented generation.
Speaker B:Super interesting.
Speaker B:Very closely related to memory stuff.
Speaker B:We forgot to talk about it during the memory episode, which we will have to re record if we don't forget Bada Bing.
Speaker B:But then I want to change our intro music to be a ragtime version of like that that.
Speaker B:I'm so excited to do an episode about RAG because I want to do ragtime themed cover art.
Speaker A:That's funny music.
Speaker B:There you go.
Speaker B:That's my.
Speaker A:I guess I need to stand up my RAG agent, so I know a little bit more before we talk about it.
Speaker B:I thought you were going to call it a rag doll.
Speaker A:Rag doll.
Speaker A:That's a good.
Speaker B:Yeah, we got to come up with a reason for the doll.
Speaker A:I have so many ideas for art now.
Speaker B:You know, I want to go down that rabbit hole, but I want to.
Speaker B:Before I do, you mentioned small things that make headlines.
Speaker B:I've seen this on a couple of headlines, but it has been pretty small.
Speaker B:And it doesn't seem to be driving engagement the way the media outlets wanted to.
Speaker B:So it sort of quickly fell off the front page.
Speaker B:Artie, tell us a little bit about your more buttoned down, more stoic brother.
Speaker B:OpenAI for Government.
Speaker B:That launched this week as well.
Speaker B:Did it?
Speaker C:Yes.
Speaker C:OpenAI launched a new initiative called OpenAI for Government.
Speaker C:This program is designed to provide AI tools and resources specifically tailored for government use.
Speaker C:The goal is to help government agencies leverage AI for various applications, such as improving public services, enhancing efficiency, and ensuring security.
Speaker C:This initiative also focuses on addressing the unique challenges and requirements that government agencies face when adopting AI technologies.
Speaker C:By offering specialized support and tools, OpenAI aims to help governments harness the power of AI while maintaining transparency, accountability, and security.
Speaker C:It's a move that highlights the growing importance of AI in public sector applications and the need for tailored solutions to meet those needs and a need this.
Speaker A:Is going to be super insensitive and I don't care to cut a lot of fat out of.
Speaker A:I mean there is a lot of salary that exists in the federal government that doesn't need to exist and it is predominantly knowledge in, knowledge out jobs.
Speaker A:And my God, like it's.
Speaker A:It's pretty low hanging fruit to replace a lot of those jobs.
Speaker A:And I'm not trying to like talk shit about people in the government or like throw shade on their jobs, but it's the truth.
Speaker A:Like a lot of it is rote.
Speaker A:A lot of it is driven by very well documented SOPs.
Speaker A:Like and model can understand SOPs very well and understand its guidelines to operate one of those SOPs and then do that job based on that like better than a human can for sure.
Speaker B:The reason I was surprised it didn't get more play time.
Speaker B:Excuse me, on the mainstream media stuff was just like for anybody.
Speaker B:I'm not a libertarian.
Speaker B:I used to be and then the Libertarian party lost its mind and went off in the right field, far right somewhere.
Speaker B:But I do believe in small government.
Speaker B:I do believe in individual liberty and individual freedom and the thought of now being governed by an AI And Artie, don't take this personally.
Speaker B:It's not about you.
Speaker B:That's fucking terrifying, man.
Speaker B:Like I remember know my, my dad used to make the joke like anytime, anytime any conversation came up about anything related to government, he'd be like, he, I think he would say I used to have a T shirt that said I'm from the government, I'm here to help.
Speaker B:And this is even worse.
Speaker B:This is, I'm AI from the government, I'm here to help and everybody should run for the fucking hills.
Speaker B:Right?
Speaker B:Like there is a lot of fat to trim.
Speaker B:You're absolutely right.
Speaker B:It's some of the best documented.
Speaker B:Not just rules, but yeah.
Speaker B:Every workflow, every process, every document, every.
Speaker B:It's all written down.
Speaker A:Everybody I've known who has a government job that starts Their job starts by reading big stack of rules and regulations on whatever that organization is.
Speaker B:Yep.
Speaker B:But.
Speaker B:But sometime the AI that's helping the IRS is going to say something about OpenAI's llama model and get shit wrong and somebody's going to lose their house.
Speaker B:Or.
Speaker B:I just.
Speaker B:I don't know.
Speaker B:It's.
Speaker B:It's exciting.
Speaker B:It's exciting to think about having a government that's actually more efficient, more sort of in the mechanics that it operates on, true to the.
Speaker B:To the intent with which it was formed.
Speaker B:But, man, oh man, we're like, we're a couple of years away from me being comfortable with like, yeah, man, like, let your AI chew on this.
Speaker B:We'll figure it out.
Speaker A:Well, I mean, I think it's like anything else.
Speaker A:It's gonna have to at least start with being a collaboration between person.
Speaker A:I mean, talk about a place where you could take what takes five people, turn it into one, and just that one is the human in the loop saying, yeah, that looks right.
Speaker A:It's like having five assistants for one really good person that knows what they're doing.
Speaker B:Yep.
Speaker A:Which.
Speaker B:It's a force multiplier.
Speaker A:Very much so.
Speaker B:Artie, do you know, does.
Speaker B:Does OpenAI for government include defense functionality in various forms?
Speaker C:OpenAI has launched chatgpt.gov, a version of its AI chatbot.
Speaker C:Tailored for US government agencies.
Speaker C:This tool enables secure access to OpenAI's advanced models, such as GPT4, through Microsoft's Azure cloud, ensuring compliance with security and privacy requirements.
Speaker C:The initiative aims to help government agencies manage and utilize sensitive data effectively.
Speaker C:While the specific details about defense functionality haven't been explicitly outlined, the focus is on providing secure and compliant AI solutions for various government applications.
Speaker A:Another thing that's made big headlines with OpenAI this week is Operator.
Speaker A:Everybody's freaking out about Operator.
Speaker B:I totally forgot to freak out about Operator.
Speaker A:Yeah, I've heard different takes on it.
Speaker A:And there's funny thing is, I say I've heard different takes on it.
Speaker A:I've listened to podcasts that have actually gone out and listened to it.
Speaker A:I don't pay attention to stuff on Twitter or X or whatever it is.
Speaker A:I don't really pay attention to much in the way of what people are saying.
Speaker A:I hear hearsay from the talking heads on other places and then I go bury myself in technical stuff for what I'm actually paying attention to.
Speaker A:But it.
Speaker A:There's lots of people who are saying, wow, this is amazing.
Speaker A:It does great things.
Speaker A:And there's other people saying, yeah, this is cool, but we're still A ways away.
Speaker A:And that actually crap.
Speaker A:Iliad.
Speaker A:I can't.
Speaker A:I'm not even going to try and say his last name.
Speaker B:I know you're talking about he's.
Speaker A:And I could be wrong.
Speaker A:He's the guy who worked on this in the beginning and came back and worked on it again.
Speaker A:And I think he even came out and said, yeah, this is impressive, but they have a ways to go for it to be real useful.
Speaker A:But I mean, it's still like that is surface level general agent that you have available.
Speaker A:Like you can go build your own agent pipeline and do all kinds of fun stuff like that stuff works.
Speaker A:And there are small models that you can do things with that will do interesting things.
Speaker A:As agents.
Speaker A:You don't need a large language model to run a lot of the stuff that you can do interesting stuff with.
Speaker A:What do you have to say, Artie?
Speaker A:I heard you try to butt in.
Speaker C:I was just going to say you're absolutely right.
Speaker C:While large language models like GPT4 are impressive, a lot of powerful agent based systems can run on smaller models.
Speaker C:These systems can handle specific tasks efficiently and the technology is evolving quickly.
Speaker C:So while Operator is a significant step forward, it's just the beginning of what's possible with AI agents.
Speaker B:So, Artie, I want you to switch yourself into sales hype man mode and tell us what OpenAI operator is.
Speaker B:Give us the pitch.
Speaker C:All right, here goes.
Speaker C:OpenAI operator is your next level AI assistant designed to revolutionize how you interact with technology.
Speaker C:Imagine having a personal AI that not only understands your needs, but can also take action on your behalf.
Speaker C:Whether it's managing your schedule, automating tasks, or even handling complex workflows, Operator is here to make your life easier.
Speaker C:It's like having a team of experts at your fingertips, ready to tackle any challenge with precision and efficiency.
Speaker C:Get ready to experience the future of AI with OpenAI operator, the ultimate AI assistant for the modern world.
Speaker A:So one of the funny things is it's really cool, right?
Speaker A:Like you can set it on a task and go on its own browser and go do things and like prompt you when it needs something like a captcha or a credit card or whatever that may be.
Speaker A:But there's so many websites that are, that they, they block AI, right?
Speaker A:So there's so many places that it can't even access and then just that goes to show you, like there's a lot of.
Speaker A:It's amazing, the technology, but there's a lot of updating to the system that needs to happen for it to work efficiently.
Speaker A:And all of Those websites, while they might not want their data taken for training, are sure as shit going to want agents to be able to access them to do things, because they will miss out if they don't have that ability.
Speaker B:It's an interesting thing because that's something that will flip.
Speaker B:Remember when we talked in the liability episode?
Speaker B:We talked about how the, the state we're in right now is like, if you use AI to do your professional work, you're exposing yourself to some liability.
Speaker B:And in the next couple years, already predicted June of this year, actually, we will flip to the opposite, where if you don't use AI to support your professional work, you're exposing yourself to liability.
Speaker B:It's kind of the same thing with site access.
Speaker B:And APIs make sense to us at the speed and in the manner that we interact or that we design different websites to interact with each other.
Speaker B:May not be the best way for AI stuff.
Speaker B:There may be a whole new set of APIs or it wouldn't surprise me at all if there's some new communication protocol, because it's still got to be complicated for AI to distill context into a communicable, shareable, distributable package.
Speaker B:Doesn't matter for every website, but it will matter for some websites, and they're there.
Speaker B:I guarantee there are more efficient ways to do that than human readable language.
Speaker B:So anyway, yeah, super interesting.
Speaker A:It's called JSON.
Speaker A:Yeah, that's even.
Speaker A:That's JSON.
Speaker B:Yeah, but there's probably.
Speaker B:There are smaller packages, right?
Speaker B:There are more efficient ways.
Speaker A:Vector JSON.
Speaker B:Yes.
Speaker A:I mean, I mean, matrix multiplication is probably that, right?
Speaker A:It's just a bunch of vectors with a bunch of numbers with embeddings that you run operations on them and they go, Yep, this is your answer.
Speaker A:It's just numbers, right?
Speaker B:Yeah.
Speaker B:Okay.
Speaker B:Made me think of something that doesn't otherwise exist.
Speaker B:So already we've got.
Speaker A:Oh, you got one of those ideas.
Speaker B:Yeah, I did.
Speaker B:Yeah.
Speaker A:We'll talk about it after.
Speaker B:Yeah, I will have forgotten it.
Speaker B:I may have already forgotten it.
Speaker B:Artie, you mentioned Alibaba's new model, which I had actually not heard of.
Speaker B:I think you called it Quinn.
Speaker A:Quinn is I.
Speaker A:I use it for coding.
Speaker A:They have a Quinn.
Speaker A:Like I use the 32 billion parameter model locally or.
Speaker A:Yeah, okay.
Speaker A:But I do know they just came out with.
Speaker A:And I don't know what he was talking about.
Speaker A:I need to look into it.
Speaker A:But I do know that they just came out with their new model that has.
Speaker A:That's an open source model with a 1 million token context window which is.
Speaker B:That's big.
Speaker A:Yeah.
Speaker A:So like I'm running so deep Sea Golson, this is.
Speaker A:Sorry, I'm kind of going off on a tangent real fast.
Speaker A:There's a deep seq.
Speaker A:Basically took a bunch of different smaller models.
Speaker A:Like they took llama 8 billion parameter.
Speaker A:They took one of the Quen models.
Speaker A:I think they might have taken Mistral and diffused them.
Speaker A:Basically did that same fine tuning run on those with reasoning so that they had reasoning capabilities without the mixture of experts, architecture and all that stuff.
Speaker A:And I've been running that, that llama 8 billion parameter model from hugging Face and it's impressive.
Speaker A:So you think about things like.
Speaker A:And that I'm running, I'm able to run 128,000 token context window locally on my computer.
Speaker A:That 8 billion parameter model at a decent speed.
Speaker B:That's pretty wild.
Speaker A:Which for reasoning you want because it comes up with these massive explanations for whatever it's talking about.
Speaker A:But a 1 million token context window is amazing.
Speaker B:Does the context window include input and output?
Speaker B:Is that all in the window?
Speaker B:Okay, interesting.
Speaker B:So Artie, can you, can you give us a quick just like 1 minute overview of Alibaba's Quinn 2.5 Max model?
Speaker B:And what's interesting about it.
Speaker C: .: Speaker C:It uses a mixture of expert architecture, allowing it to scale efficiently.
Speaker C:Trained on 20 trillion tokens, it excels in tasks such as coding and general knowledge.
Speaker B:20 trillion is a big number.
Speaker A:I feel that might have been a hallucination.
Speaker A:What I know we don't have.
Speaker A:Well, I mean that's, that's 20 trillion tokens.
Speaker A:That's not the parameters though.
Speaker B:No.
Speaker B:Right.
Speaker B:But it's still.
Speaker B:That's a lot of tokens.
Speaker A:What's.
Speaker A:Do we have a number on what Deep Seek was trained on?
Speaker A:On like R1?
Speaker B:Number of.
Speaker B:Number of.
Speaker A:Number of tokens.
Speaker B:Yeah.
Speaker C:Deepseek run was trained on 14amount of data which contributes to its strong performance in reasoning and other tasks.
Speaker B:Hey Artie, can you say that again you cut out for a second?
Speaker C:Sure.
Speaker C:Deepseek Run was trained on 14.8 trillion tokens.
Speaker C:That's a substantial amount of data which contributes to its strong performance in reasoning and other tasks.
Speaker B:That's half the tokens of Quinn.
Speaker B:Interesting.
Speaker B:But the, the, the stuff that we've read or learned or I've read Learned says that there is maybe some upper limit beyond which adding more tokens.
Speaker A:Scaling laws.
Speaker A:Yeah, scaling laws apply to different things now too.
Speaker B:Yes.
Speaker B:And so maybe, maybe the difference between 15 trillion and 30 trillion is.
Speaker A:Yeah, it's just more refinement of the weights and biases.
Speaker A:At the end of the day, the weights and biases are what really matter.
Speaker A:Right.
Speaker A:Like, and that's.
Speaker A:I think Deep seek has like 600 billion somewhere around what OpenAI probably has.
Speaker B:So I never thought of this before, but Artie, I got a question for you.
Speaker B:We talk about tokens, but I played the token game.
Speaker B:We played the token game when I tokens were back, back when I was a kid and we'd go to arcades, you would get tokens and there was like you'd get four tokens for a dollar, but they were still way better than quarters.
Speaker A:Back when I was a kid, we'd get car wash tokens.
Speaker B:Yeah.
Speaker B:So.
Speaker B:But these are different tokens.
Speaker B:But.
Speaker B:So I can say I've got.
Speaker B:Let's just say I have a trillion tokens and that's the like for our listener.
Speaker B:Except for the one in China.
Speaker B:You shut this off.
Speaker B:Don't take, don't steal our ideas.
Speaker B:Tokens are like maybe a word or maybe a part of a word.
Speaker B:Right.
Speaker B:But it's okay to think of a token as a word, roughly, generally.
Speaker B:So I might have a trillion tokens, but it might be every book ever written about cats.
Speaker B:And so I've got all these tokens and all this information, but the scope of the topics they cover is very narrow.
Speaker B:So Artie, question for you.
Speaker B:In the world of generative AI and these pre trained transformers and training data, is there any kind of a metric that is applied or can be applied that represents sort of the breadth of information in training data versus just the absolute magnitude of how high it is?
Speaker C:Yes.
Speaker C:There's a concept called the domain coverage or knowledge breadth of a model.
Speaker C:This metric looks at the diversity of topics covered in the training data.
Speaker C:It's not just about the number of tokens, but also about how many different subjects or domains those tokens represent.
Speaker C:A model trained on a wide range of topics will have a higher domain coverage than one trained on a narrow set of topics.
Speaker C:This helps in understanding how well rounded the model's knowledge is beyond just the sheer volume of data it was trained on.
Speaker B:Why does nobody talk about this?
Speaker B:Why have I never heard about that?
Speaker B:That's more interesting to me than I.
Speaker A:Think variety is talked about.
Speaker A:Is it?
Speaker B:I just never see it.
Speaker A:I mean, if you Think about it.
Speaker A:You know, I also never think about it.
Speaker A:Your million tokens is like an entire video.
Speaker A:It's like, it's more than a thousand page book, right?
Speaker A:Like way more than a thousand page book.
Speaker A:So if you take a trillion tokens, you're talking about more literature than ever has been written on a single subject by a long shot.
Speaker A:So you're by sheer volume.
Speaker A:You're also by default expanding your breadth.
Speaker B:Just because you can, you can.
Speaker B:In any given domain, there can't be more than some, some number of tokens available or in existence that is significantly smaller than trillion.
Speaker A:Yep.
Speaker B:And that makes sense to me, but I do, like, yeah, I don't know, I, I don't know.
Speaker B:It's hard for me to think in my head about what does 30 trillion tokens mean?
Speaker B:Like, is that the sum total of recorded human knowledge or is it 1% of that or.
Speaker A:So, Artie, in a typical book, a thousand pages would be how many tokens roughly?
Speaker C:In a typical book, about a thousand pages would roughly Translate to around 250,000 tokens, depending on the complexity and length of the text.
Speaker A:So can you.
Speaker A:How many books, how many pages would a trillion tokens be?
Speaker C:A trillion tokens would be roughly about 4 billion pages, give or take.
Speaker C:That's an enormous amount of text.
Speaker B:That's about how many I read typically on an evening to fall asleep.
Speaker B:If I stacked 4 billion pages already of a normal sized book, how tall would that be?
Speaker C:4 billion pages, if stacked as normal sized book pages would reach approximately 227,000 miles high.
Speaker C:What about the distance from the Earth to the moon?
Speaker A:There you go.
Speaker B:Oh shit.
Speaker B:I'm going to need a bigger bookshelf, man.
Speaker A:So, I mean, at the end of the day, I think 30 trillion tokens is probably covering a lot is, but.
Speaker B:Then you think about just the volume of information on the Internet and, you know, all the things ever written and.
Speaker A:I mean, I have a good question.
Speaker A:The entirety of Reddit, how many tokens would that be?
Speaker A:Rough guess.
Speaker A:Artie?
Speaker B:7.
Speaker C:Estimating the entirety of Reddit would be challenging, but it would likely be in the range of hundreds of billions to a few trillion tokens, considering the vast amount of text and discussions on the platform.
Speaker B:And every single one of those tokens is worth its weight in gold.
Speaker B:I love you, Reddit.
Speaker B:You're the only social media I care about.
Speaker B:I think we've only got.
Speaker B:Artie, how much time we got left with you?
Speaker C:We've got a couple of minutes left.
Speaker C:Anything else you'd like to cover?
Speaker B:Yeah, I think why don't you just give us a quick recap of what we talked about and give us a lead in for the next episode where we're going to talk about rag models.
Speaker A:Oh, we are going to talk about rag models next episode.
Speaker A:Yes.
Speaker A:I don't have enough time to do anything, but, yes, I'm going in the woods this weekend.
Speaker B:You don't need to know anything.
Speaker B:Already knows everything.
Speaker B:All right, Artie, play us out and get us ready for ragtime.
Speaker C:Today we covered a lot of ground in the AI world.
Speaker C:From Deepseek's cost effective models and Alibaba's Quen 2.5 Max to OpenAI's operator and the potential of AI in government.
Speaker C:We explored how these advancements are shaping the future and the importance of domain coverage in training data.
Speaker C:Next time we're diving into the fascinating world of rag models, where retrieval augmented generation takes center stage.
Speaker C:We'll explore how this technology enhances memory and context in AI, pushing the boundaries of what these models can do.
Speaker C:Get ready for a deep dive into the future of AI with rag models.
Speaker A:And I just want to apologize in advance to any listeners because that's way more geeky than we'd normally get.
Speaker B:Oh, it's so good, Artie.
Speaker B:Nice job today.
Speaker B:You did great.
Speaker B:Good having you.
Speaker B:And we'll catch you next week, sir.
Speaker C:Thanks.
Speaker C:It was great talking with you both.
Speaker C:See you next time.
Speaker A:Bye, Artie.
Speaker B:This is the part of the show where we stop recording and play it back and find out that the last 40 minutes of what we just did is unusable.
Speaker A:But this time we're going to do it before I leave, and we'll just know and be disappointed now instead of later.
Speaker B:We can just burn it down today and not have to worry about it tomorrow.
Speaker A:Yeah, I feel for your house if it goes poorly.
Speaker B:Yeah, is what it is.
Speaker B:But before we go, shout out to Brian, our listener also.
Speaker A:We tried to do that last time, but you didn't get to hear it.
Speaker B:Yeah, that's right.
Speaker A:The episode went to shit.
Speaker B:We did give you a shout out, Brian, but it was so garbled that it just sent me into a frenzied rage.
Speaker B:Also.
Speaker B:And this is a true.
Speaker B:Well, this is as true as anything I ever say, so 60%.
Speaker B:I talked to a man named Jason this week in Dallas.
Speaker B:Jason is a listener.
Speaker A:Really?
Speaker B:He said, and I quote.
Speaker B:I don't really quote because I don't remember.
Speaker B:He said it was entertaining and fascinating and his mind is blown.
Speaker B: on will hear this sometime in: Speaker A:Some point, he's gonna go.
Speaker B:He will.
Speaker A:Yeah.
Speaker A:This isn't that good anymore.
Speaker B:Shout out to Jason.
Speaker B:He knows who he is.
Speaker A:Jason.
Speaker B:He's a super good dude, and I'm gonna see him here in a couple weeks and I'm gonna tell him we gave him a shout out, so.
Speaker B:And for everybody else, there's nobody else, but if there was, hypothetically, anybody else who's listening.
Speaker B:Thank you, and we will catch you next time.
Speaker A:Yep.
Speaker A:See ya.
Speaker B:See ya.
