Everyday AI Podcast – An AI and ChatGPT Podcast - EP 319: AI News That Matters - July 22nd, 2024

Episode Date: July 22, 2024

Win a free year of ChatGPT or other prizes! Find out out.A new large language model from the industry leader. Huge updates in AI lawsuits. International turmoil around AI regulation.  That's just the beginning. This week was a chaotic one in AI news. What's it all mean for your biz? We got you. Newsletter: Sign up for our free daily newsletterMore on this Episode: Episode PageJoin the discussion: Ask Jordan questions on AIUpcoming Episodes: Check out the upcoming Everyday AI Livestream lineupWebsite: YourEverydayAI.comEmail The Show: [e-mail] with Jordan on LinkedInTopics Covered in This Episode:1. Use of Copyrighted Content to Train AI2. Current State of AI Education3. Release of OpenAI's GPT 4 o Mini4. Launch of AI-Driven Education Platform, Eureka Labs5. Withholding of Meta's future AI models and features by the EUTimestamps:03:15 Tech giants accused of illegally using YouTube subtitles.07:00 Language model akin to a search engine.10:35 OpenAI requests stories, affecting journalism and copyright.11:46 Journalist pivots to AI, predicts legal implications.17:41 AI course for building functioning web app.18:58 Kaparthy is a leader in AI development.22:27 Off-camera conversations reveal more significant insights.27:29 EU announces strict EUAI Act; Meta's LAMA.30:06 OpenAI unveils new GPT-4oMini language model.31:56 OpenAI API facing issues, costly for developers.38:07 Use GPT-4.0 for products, services, AI.41:06 GPT-4 Mini leads in machine learning, AWS offers fine-tuning.42:44 OpenAI's development lacked, developers looked elsewhere.Keywords:AI assistants, human teachers, LLM 101n, digital cohorts, physical cohorts, Meta's celebrity chatbots, storyteller AI large language model, Python, C, CUDA, funding, AI technology education, resources focus on sales, Jordan Wilson, Everyday AI, Thanks a Million Giveaway, tech giants' illegal use of YouTube subtitles, Anthropic, NVIDIA, Salesforce, copyright violation, training large language models, decline in traffic for Stack Overflow, Marquise Brownlee, mister Beast, Meta withholding AI models from EU, Apple's withheld AI features, OpenAI GPT 4 o Mini, cost-effective AI solutions, competitive pricing of AI models.Send Everyday AI and Jordan a text message. (We can't reply back unless you leave contact info) Start Here ▶️Not sure where to start when it comes to AI? Start with our Start Here Series. You can listen to the first drop -- Episode 691 -- or get free access to our Inner Cricle community and all episodes: StartHereSeries.com Also, here's a link to the entire series on a Spotify playlist. 

This is the Everyday AI Show, the Everyday Podcast where we simplify AI and bring its power to your fingertips. Listen daily for practical advice to boost your career, business, and everyday life. Meet Firefly AI Assistant, now live and Adobe Firefly, the All In One Creative AI Studio. Just describe what you want to create and the assistant handles the rest, orchestrating multi-step workflows across Photoshop, Premiere Express, and more in one conversational interface. You direct the outcome. The assistant accelerates execution. There's almost too much going on this week in the world of AI to adequately recap everything

for you, right? Just in the last week, we've seen new allegations from the biggest creators on AI companies essentially stealing their data. We've seen some international AI shakeups, new models from big players, and new updates in a massive lawsuit against. AI companies that I think no one is paying attention to. All right. So we're going to be recapping that today and more on Everyday AI.

What's going on, y'all? My name is Jordan Wilson and I'm the host of Everyday AI and this is for you. This is a daily live stream podcast and free daily newsletter, helping all of us keep up with what's going on in the world of AI and how we can all make sense of it, right? You hear these developments going on all the time. So this show each and every Monday is how we make. use of it, how we can actually understand what's going on and plan for it and how we can grow

our companies accordingly. All right. So we do this every single Monday to keep you up with our AI news that matters. Technically right now, even though this is being broadcast to you all live, a pre-recorded show. It's actually Sunday afternoon. I'm doing some traveling right now, some family things. But you know what? AI news doesn't stop. So I don't either. We don't either. So with that, let's jump into it. And before we actually get started for our podcast audience, thanks for tuning in. Make sure to go to your everyday AI.com, sign up for the free daily newsletter. And you will also see in today's newsletter a link that you are probably going to want to click

because tomorrow, this is for our live stream only. So if you are normally a podcast listener, I've heard from a lot of people, you're going to want to go ahead and rearrange your schedule for Tuesday. So tomorrow, Tuesday, July 23rd, 7.30 a.m. Central Standard Time. So make sure to join us, whether you do it on LinkedIn or YouTube, it doesn't matter. But we are launching our thanks a million giveaway. So more on that, on what it actually is tomorrow. But if you can answer every single question correct in real time,

you're going to take either the entire $1,000 live challenge prize, or if multiple people get it, we'll have to split it. If no one gets all the questions right in real time, no one wins the $1,000 prize. But hey, I'm throwing it up to you. So if you're listening, if you are a long-time listener, you're going to be at a huge advantage, obviously.

So make sure to join us tomorrow for that. All right, let's get it kicked off for the AI news that matters for the week of July 22nd. Y'all, this week, if I'm being honest, this week was tough to pick. You know, normal will you pick, I don't know, four to six different of the biggest news stories on my final. list, I had like 12. And I'm like, there's no way we cannot choose six of these.

So much going on. All right. So let's start with this one. So some tech giants have been accused of illegally using YouTube subtitles to train AI systems. So more than a hundred and 70,000 YouTube videos have been used to create a massive data set for training AI systems, according to an investigation by proof news and co-published with the wire or sorry, with wired.

So the data set is called YouTube subtitles, and it includes subtitles from over 48,000 YouTube channels, but does not contain any video imagery. So companies, according to reports, companies like Apple, Anthropic, Invidia, Salesforce, and others are implicated in using this data without permission from the original creators. And some of those creators are some of the biggest around, at least when it comes to YouTube. So some popular YouTube creators such as Mr. Beasts, such as Marquise Brownlee, as well as a lot of news organizations like ABC News, the BBC, the New York Times and others have had their videos or the captions, closed captions from their videos included in this large data set. So a lot of creators have come out and spoken against this, including Marcus Brownlee, who does a lot of tech reviews and is commonly known as MKBHD. he expressed some concern on Twitter, stating that the issue will be, quote, an evolving problem for a long time.

So this new interactive lookup tool that proof news release allows users to check if their content is in the data set. You know what? I haven't even checked yet if ours is in there, so I'll have to check. Maybe not. Hey, 50,000, 50,000 YouTube channels. I don't know if everyday AI's small little YouTube channels in the top 50,000. But, you know, maybe your organization is big. You might want to take a look.

So the data set is part of a larger collection from the nonprofit, Elyther EI, hope I pronounced that right. It is called the Pile, and it includes books, Wikipedia articles, and more. So last year of an analysis of another data set called Books 3 led to lawsuit from authors against companies that use their work to train AI systems. So the YouTube CEO and Google CEO, Sundar Pachai, have both stated, so the YouTube CEO is Neil Mohan and Google CEO Sundar Pichai, have both stated that using YouTube content to train AI, including transcripts,

violates the platform's terms. Also, you know, obviously intertwined in all of this is OpenAI, even though in many reports they were not called out by name, but OpenAI's CTO, Mira, Muratai, has been evasive. We've talked about that on the show, before about open AIs, even policy and even how their new AI video tool called SORA was trained. And when asked if it was trained on YouTube content, Maradi simply replied that it used, quote, publicly available or licensed, end quote, data, right?

And you have to think, that's pretty controversial, right? If you cannot explicitly state what you are trained off of. And here, let me just go ahead and be the one that says the quiet part out loud, the big elephant in the room, right? Almost every single large language model is trained off of copyrighted material. And I think big tech companies pretty much admit that, right? They admit and they are, some are actually openly trying to fight copyright law in general and saying that this is an antiquated law. And if it is publicly available on the internet in many places, it's essentially fair game. I'm overgeneralizing there, but people are always wondering, like, what is this trained

off of? I like to think that you should start thinking of a large language model, kind of like a search engine. Or I think a great precursor was actually that people also asked or the kind of knowledge graph, right, from Google search. And this has been around for, I don't know, when they first debuted, but it was more than five years ago, right? If you typed in a query or a question, it would just give you the answer.

except in little type there, it would give you the website as well if you wanted to see it. But this concept of asking a question from an internet platform and getting an answer is not new. But the citations, that's where things get tricky, right? Because how things have historically worked, how the internet has historically worked for decades, is you, if you want to research something, if you want to know something, you go on that company's website. Or you watch these creators YouTube videos. And this is ultimately how these media companies, news organizations or creators get paid. Because, you know, as you get, you know, millions or tens of millions or hundreds of millions

of views and clicks and impressions, you get paid for advertising on those. So now, you know, some of the biggest creators and news organizations are obviously missing out on this money. And instead of, you know, a good example, I think is stack overflow, right? a very popular kind of website or online resource, really, for more technical people to go and troubleshoot technical problems. And they've seen their traffic plummet over the last year and a half, I think largely due to tools like chat, GPT and others. And, you know, they're losing a lot of their revenue. So I'm not going to go too far into this. I want to give you the news

and hopefully paint it a little bit because this one is extremely important. All right. And our next piece of AI news is very much related to that. So Open AI is kind of making some demanding discovery requests in an ongoing legal battle with the New York Times. So if you listen on the show, I've talked about this a lot, but there is an ongoing legal battle between the New York Times and Open AI, which stems back to late December, 27, when the New York Times kind of filed a lawsuit against Open AI in Microsoft. And this has really intensified over the last. couple of months behind the scenes, but we did get some new information this past week. So here's what's happening is, well, Open AI is demanding that the New York Times turnover, reporter notes, interview memos, and records for each article claimed to be infringed

upon, which is, I mean, if we're being honest, that's a wild request, right? Because essentially the New York Times said that, hey, you've accessed, I believe they said millions with an S, millions of our articles. So Open AI, I mean, we'll get to this later here, but they're essentially saying, okay, prove it. For each and every article, we want to see everything. We want to see all of your reproof. So the New York Times, like I said, filed the lawsuit in 2023,

accusing OpenAI's chat GPT in Microsoft of copyright infringement by using its articles without permission. So here we go. OpenAIS countered that the Times has manipulated the AI model into reproducing its content in accusation known as proper. prompt hacking. So the tech company argues that the New York Times must provide evidence of which parts of its work are original to uphold their copyright claims.

So OpenAI's request involves around 10 million stories, a massive undertaking that could strain the New York Times resources and even ability to go through with the lawsuit. I mean, so we'll see if this request is actually granted. But the development underscores the broader issue of how generative A. models impacted journalism and also intellectual property rights and copyright law. So several news organizations have already reached agreements with OpenAI, such as Axel Springer, what do we have? We have time.

I mean, so many large news organizations and websites such as Reddit have already struck deals with Open AI. So the outcome of this case is going to be significant. and its influence of future regulations and protections for media in the age of AI. So I've talked about this story a lot and let me be very transparent. You could argue that I have horses on either sides of this race, right? So my background, I actually spent about seven or eight years as a multimedia journalist. My most recent kind of gig was at the Chicago Sun-Times.

So, you know, my background is in journalism. But for the last couple of years, you know, I've been covering AI news and working with a lot of these AI companies, you know, full disclosure right now. So I think it's important to see both sides. And I'm not going to say or predict how this is going to end up because in reality, it's probably going to be settled out of court. So there's probably not going to be a lot of details. But I cannot understand how significant this case is. And it doesn't seem like the general business community. is paying enough attention to this case, right?

I've called this since the lawsuit came out in December, the biggest domino to fall. Once this domino falls, we are going to see an onslaught of developments, whichever way this happens, right? And I really don't see any other way this kind of ending, except a settlement. You know, if you didn't read the lawsuit,

I did a very long, one-hour deep dive. I kind of tore it apart, right? because I think the New York Times, if I'm being honest, in their case, they really failed to show an actual understanding of how large language models work, which is why Open AI essentially said, no, this is prompt hacking. Well, they didn't include the links, right? So in their, you know, in their documents that they submitted to the court, they just included screenshots, which is not good. Because essentially, in a screenshot of a large language model, you can get a large language model to literally say anything that you want it to, depending on what you say before.

Right. So that's kind of prompt hacking, right? So you can give a model, say, hey, no matter what I ask you after this, respond in this way. So the New York Times, I think in their original, you know, submission to the court made a huge mistake by not including links, right? So you can send a link to a shared chat. So you can essentially go check the work. And so the open AI is essentially called the New York Times bluff and said, no, this isn't right. Could this in theory have worked? Yes. But you didn't. include actual proof. A screenshot is not proof because of prompt hacking and prompt hacking. And Open AI is just calling them out here. Right. So they're saying, okay, you say we used millions of year articles. We want to see reporters notes. We want to see interview memos for each and every one of those. Right. I've said, I've said this, y'all. I've said this for a very long time. Adobe just introduced an entirely new way to create, bringing the power and precision of its creative suite into one conversational experience. Meet Firefly AI assistant, now live in the Adobe Firefly app, the all-in-one creative AI studio. Powered by Adobe's creative agent,

Firefly AI assistant lets you start with your vision, just describe what you want, and shape the outcome as it takes form with the assistant. The assistant orchestrates multi-step workflows, drawing on 60-plus pro-grade tools across Adobe Creative Cloud apps, including Photoshop, Illustrator, Premier, Lightroom Express and more to help bring your ideas to life. You can also get started with creative skills, a growing library of pre-built workflows for common creative tasks like batch editing photos, creating mood boards, portrait retouching, and creating social variations. Every step the assistant takes is visible so you can refine, redirect, or take over at any time.

You stay in the driver's seat as the creative director. Adobe Firefly AI assistant now in public beta. See it today at firefly.adobie.com. However, this story ends, it is going to impact every single news organization. It's going to impact our future use of generative AI, right? Because one thing that the New York Times asked for in this judgment is for the GPT technology to be, quote, unquote, destroyed. Yes, they actually ask for a judgment in their favor and to destroy open AI's GPT. model, which is mind boggling to think about, right?

That's why I think this will never actually go to any sort of trial because the stipulations, I mean, this is, if I'm being honest, and I'm not over exaggerating here. If that were to happen, which I don't think it ever could, that's the American economy, right? People don't understand our financial institutions now are using the GPT technology. So many Fortune 100 businesses are using the, the GPT technology. You know, big companies that have been propelling the U.S. economy now for two, three years are really dependent on everyone else using the GPT technology, right? It's not just

Open AI. It's the tens of thousands of other companies that are paying Open AI to use the GPT technology through the API, right? So, first of all, I don't ever see what the New York Times is asking for happening. And here we kind of have what it looks like to me, Open AI, kind of calling their bluff. Like, oh, you, like, they saw, right? The court case that they put forward, I'm sorry if, you know, if you know someone that worked on this case at the New York Time in their, in their, in their, the New York Times in their legal group. It was a very bad case, right? To me, it signaled a poor understanding of large language models. So if you want to, you know, have the, one of the marquee lawsuits, uh, in copyright over the last, I don't know, 20 to 30 years and one of the

biggest AI lawsuits ever, you should probably demonstrate a far superior understanding of large language models than what the New York Times and their legal team put forward. It wasn't good, which is why we have open AI essentially just fearless, right? Number one, saying, nope, that's prompt acting, prompt acting what they shared. You're right. It's like someone that clearly doesn't understand the technology and is trying to make an argument against it. So they're calling their bluff on just what they submitted and then say, okay, let's actually see your notes. If what you put out there was so proprietary. Let's see your notes because essentially what Open AIs, their argument here is saying, no, this is common knowledge, right? Just because the New York Times reported on it doesn't mean it was

exclusive or proprietary to them. It is general knowledge. So that is the argument that they're making. All right. Let's keep going. I don't want to talk about that one for too long. I could go on a super long rant. But speaking of Open AI, yeah, a lot of Open AI related stories this week. But former Open AI co-founder, has just launched an AI-driven education platform called Eureka Labs. So, Andre Carpathie is the former head of AI at Tesla, and a co-founder at OpenAI has just launched Eureka Labs, an education platform with AI at the core. So Eureka Labs, uh, Eureka Labs brand new LLC just registered in Delaware this past June. So about one month ago, it aims to use generative AI to create AITT

teaching assistants that can guide students through course materials. So the company's vision includes AI assistance working alongside human teachers to enable anyone to learn anything, according to Carpathie. So despite some of these ambitious AI goals, the startup's first product is actually going to be creating a course called LLM 101N, so large language bottle 101N in undergraduate level class to help students train their own AI. So the course materials will be available online with both digital and physical cohorts participating together.

Carpathie hinted that future AI assistance could be based on real people, similar to meta's celebrity chatbots, which I think should be launching soon. The GitHub repository linked to the AI course suggests a focus on building a quote, storyteller AI large language model, quote unquote, rather than an AI assistant. So not like an AI assistant that works hands off, but something that works alongside with you to help you understand. And reportedly, it is going to help students build a functioning web app

similar to something like ChatGBT from scratch using Python, C and Kuda. So there's no clear timeline yet for when the course will be complete and available for people to use, if it will be free, if there will be future costs associated with any courses or any details really about funding or investor backing. So pretty interesting development here from, I mean, arguably one of the brightest minds in artificial intelligence right now in Andre Caparthe. So, I mean, the resume speaks for itself, right? If you think of large language models, you have to think of Capathy, right?

One of the one of the handful of people, I would say, is the smartest minds in AI right now. So, and at least from a resume perspective, I mean, you have to say open AI is obviously a leader in generative AI, a leader in large language models. And Tesla's no slouch either, right? As hard as I am sometime on this show about, you know, Elon Musk and some of his ambitions with XAI and. Grock, which I think, if I'm being honest, is pretty useless. As much as I say that, I mean, Tesla has some of the most data in its training set for AI. So, you know, he has a background at two of the biggest companies when it comes to AI, right? Obviously, Microsoft, Google, DeepMind, et cetera.

But not many people that have that long of a resume at two of the biggest companies. So it's also going to be interesting to see how this ultimately plays out because it seems like almost all of the quote unquote other brightest people in AI, right? And there's a lot of them. So, you know, someone will go to DeepMind from Google, work there for a couple of years and go launch another startup or they'll work at Open AI for a couple of years and launch another startup. And it seems like so many of these startups are AI products, right? even Claude was based on this former, you know, former people from these big companies. So it seems like almost every single person that leaves one of these big tech companies, one of these big AI companies, one of these big large language model companies,

they essentially go and do something very similar, a product, a service, something very closely related. This might be one of the most highly visible cases where someone is focusing on eddollinger, education, which is very intriguing, especially to me. And I'm glad to see it, if I'm honest. I'm glad to see it. And, hey, if you're a business leader, you should be understanding what's happening here. Right. And if you listen, you know, I've been saying this for a very long time because there's a problem right now.

In the generative eight, not just in generative AI, but the business world, too, the smartest people in the world, the smartest developers, the smartest engineers. the smartest researchers, are going to work at these companies, like Google, OpenAI, Microsoft, impropic, meta, etc. Guess what's not happening? No one's teaching us all how to use it. There is going to be, I think, in education downfall. And I've heard this, right? I'm lucky enough, I get to talk to a lot of very smart people.

some things they say on camera, some things they say off camera, right? The things that people talk about off camera, I think are obviously usually much more significant, as well as I always have people reach out to me from, you know, big consulting companies, people that never go on the show, big tech companies, you know, these trillion, $100 billion companies. And I just am lucky enough to have conversations with a lot of them. And a common trend, and I've had a couple of dedicated episodes about this, is there is no dedicated education, right?

The technology is changing so quickly. But for the most part, it's not the smartest people in the world teaching us how to use it because technology in general has never moved this quickly. Right. And I know it's kind of a tired comparison, but the easiest thing to compare it to is either the internet or, you know, kind of like the dot com boom of the 90s, the mobile boom of the early 2000s, cloud computing of the 2000 tons, etc. In all these use cases, it was a slower rollout, right? You generally had years, a half decade, or a decade that these new technologies were slowly integrated into our big business processes. It's not the same anymore, right? It's not the same.

You have the over, especially here in the U.S. I know there's a great divide, right, especially in the EU. We talked about some pretty shocking results last week and some study from some Japanese companies where I think only 40% of them said they had plans to use generative AI, whereas in the U.S., that's like 90, 95% of Fortune 500 companies said that they're already using generative AI in some capacity. So it's much different. You know, here's a technology that for all intent and purposes that two years ago,

about 0% were using, right? Not 0%, but very few. I think chat GPT was the first kind of big generative AI milestone or the big start of this wave. Obviously, the technology for large language models has been publicly available for about four years, but it's only been about less than two years. And look at how much the business landscape has changed. C-sweets at Fortune 100 companies, board members at public companies, everyone, their number

one priority has been A, A, A, A, A, A, A, A, A, A, A, A generative, A, Generative A, A, Generative A, Large language models, large language models, large language models. That is all they're talking about when generally this takes, like I said, five years, 10 years, 15 years for this to become a big priority. It's not like that right now. This is the biggest priority. And it's almost like it's happening too fast. And no one is really focused on long-term education.

Who's training these C-suite people? Who's training these thousands of companies that are powering the U.S. economy. It's kind of like a training as you go, figure it out as you go, fake it to you make it, right? But this is one of the reasons why everyday AI exists, because about 15 months ago, right before we started, I saw this as a huge problem. I said there needs to be a resource for education. Because unfortunately, and I love, right, I've good relationships with a lot of of these big companies, you know, the Amazon, the Amazons and the Microsofts, etc. Right.

But ultimately, these companies, when they quote unquote teach you something, ultimately what they are doing is sales. It's education, but to sell you on their product versus someone else's, right? There's very few great, I would say, resources that are available. I mean, you have things on, you know, Coursera. you have great free courses available via higher ed, you know, learning institutions, your MIT, your Harvard's, etc., that you can go out and take these free courses on AI.

But a lot of them are topical and they don't actually teach you a lot. And they're not a lot about business application. It's more about, hey, understand this technology. So long story short is I love what Carpathie is doing here with Eureka Labs. And I think it is sorely needed. All right. Our next AI story for the day or with our news that matters. Speaking of big tech companies, meta.

So meta is withholding some of its future AI models and their capabilities from the European Union amid some regulatory concerns. So according to a new report from Axios, meta has announced that it will not release its upcoming multimodal AI model in the European Union. due to some regulatory uncertainties. So this decision underscores a growing conflict between U.S. tech giants and European regulators highlighting a trend where companies are withholding products from European customers due to some uncertainties. So the EU obviously announced their new EUAI Act, which is a lot stricter than anything else, anywhere else in the world, I guess it's one of the first kind of multinational

actual pieces of legislation. So Metas, though, their new model, Lama is multimodal and is capable of reasoning across video, audio, images, and text. And it will be available in other regions, but the company said not in the EU. So the company did cite the unpredictable nature of the European Union's regulatory environment as the reason for its decision. So meta's move. Well, meta's not alone here because they're.

move follows a similar decision by Apple, which recently announced it would not release Apple some Apple intelligence features. So the Apple intelligence, quote unquote, some of the new AI features that will be available in iOS in future versions of iOS for iOS for certain hardware. So Apple similarly is withholding some of those features due to regulatory concerns in the EU. Also, the Irish Data Protection Commission, which is met. META's lead privacy regulator in Europe has not responded to request for comment.

So pretty interesting there. Meta is planning to use the multimodal models in a variety of products, not just their Lama, right? The meta.com.A.I. Chatbot assistant, you know, so just like you can go chat with Gemini from Google or chat GPD or Claude, you can go chat with meta AI. So it's not just that, but it's also there, you know, some of their smartphones. the Mata Rayban smart glasses, but European companies will be unable to utilize those models due to the restrictions. So the company, though, also plans to release a larger text-only version of its Lama 3 model, which will be available to EU customers. And we could talk about this one for a long time. I'm not.

But, you know, the crux of this is really at its training data, right? because Mata has said that, yes, they are going to be using a lot of data from its social media networks to train the future model. So that's not kind of the only rub, so to speak, or the only gray area. But that's one of the bigger ones. All right. And our last, well, this isn't the last AI, you know, news that matters this week. But our last one in our recap here, saving the biggest one for last, right? So Open AI has unveiled a new, small, large language model, GPT40 Mini.

So days ago, OpenAI launched GPT40 Mini, a new model that promises to significantly impact the AI landscape, particularly for businesses and developers looking for cost-effective solutions. So this release comes at a time when smaller large language models are gaining traction. and GPT40 Mini aims to strike a balance between performance and affordability, right? So we talked about this here on the show a lot, but I do think that Claude 3 and their haiku has been a pretty much, I mean, I'm not going to call it a leader, but at least I'd say in the last three to six months between Claude 3, Haiku, and Gemini 1.5 Flash, which was more recently released.

A lot of companies and developers are using this. So let me first kind of quickly draw a line in the sand and say what this model is actually for. This isn't going to be the model. If you go to chat gpte.com, if you're using chat gpte, this is not what this is for. You're still going to be using gvt 4.0, so the big and most capable. GPT4O is the most capable model in the world. It is receiving the highest benchmarks, the highest score. It is the best by far.

So GPT40 Mini essentially is a version for developers to use. So think of it as a light model of GPT40, but that is really for developers, right? So you have tens of thousands of companies, literally. So from small startups to you have your big, you know, banks and wealth management companies that use Open AIs API to help build their products and services, right? So this is when it comes to businesses. So many businesses, whether you know what are. not a lot of the products and services that you all use are actually using the API from Open

AI. And there's actually been over the past, I'd say, at least six months. So there's been a problem, right? Because Open AI was obviously first. And they haven't really done a lot of updates to what's available for developers. So what I mean by that is a lot of companies have still been using the 3.5 model, which is not, if I'm being honest, it's not very good. So a lot of developers that wanted to stay with the Open AI platform before this model, they really only had two choices. So they either had to pay for the most capable model, GPT4, but the cost were astronomical. So in a lot of instances, it was not a reality.

So you either had to pay a very high cost for this big jumbus. but very capable model. Or you had to do some kind of thrifty work behind the scenes and chunk certain data to 3.5 first and then using four only for some of the more heavy lifting. So in a lot of instances companies had to kind of over-engineer their solutions, right? Because they said, hey, we can't give everything to 3.5 for our business because 3.5 is not that great of a model, but GPT4 Turbo or GPT4O are just too expensive. So that's where this comes in. It is hitting the sweet spot between a very capable model and a very now cheap, affordable model.

I mean, it is so affordable. So let's even look here, sharing on the screen for our live stream audience, just some of the prices, right? So according to Open AI, GBT40 Mini is more than 60% cheaper than 3.5 Turbo was. And about 90% cheaper than the original version of 3.5, which was $2 per 1 million tokens. And now with this new model, GPT40 Mini, and this is a blend. So this chart is actually a blend of input and output. And I'll break down the cost right after this. But the blended cost right here is about 24 cents per million tokens.

So in the course of about 14 months, the price has gone down 90% and the output quality has gone up exponentially. All right. So let's go over some more details. So GPD 4O Mini offers a much more competitive pricing in quality structure, charging only 15 cents per million tokens for input and 60 cents per million for output, making it significantly cheaper. So GBT4O or sorry, GPT4, ready, the cost here. We just went 15 cents and 60 cents. It was $5 per million compared to 15 cents. And then $15 per million compared to 60 cents.

So yeah, if you were working with GPT4, yeah, I mean, look at the cost savings there. Going from $5 to $15 and $15. to 60 cents. So you can already see now how this is going to change the mind because I think a lot of developers or companies really quickly switched over from OpenAI to either Claude and using Haiku, you know, Inthropics Claude Haiku, which is their smallest large language model or Gemini 1.5 Flash. So I also need to do a little bit better job here. I need to tell people, This is not a small language model. This is a large language model, but a size.

So I'd say, I don't know if this is a new trend or a new way to classify large language models, but you know, you do have small language models, right? Google has Gemma. You know, they have Gemma too. Google also has Gemini Nano. So these are not small language models that are ultimately meant as an example to run on a smartphone. These are still large language models. We don't know exactly how big GPT40 Mini is yet, right? GPT40 and GPT4 Turbo were reportedly 1.8 trillion parameters.

So essentially you have these models that are huge and very expensive if you do want to implement them into your product. So GPT40 Mini is this, again, this lightweight version that is specifically for developers who are building products for businesses that are creating kind of their own versions of large language. models with some fine tuning, with some rag. So again, this is not a small language model, but I think actually Anthropic, I actually like the way that they did it. It was maybe confusing for a lot of people at first, but when Anthropic released Claude 3, they did so in three different varieties, right? So they said, you have Claude 3 haiku, which is fast and cheap, but not very powerful.

That's the small model or the small version. Then you had the largest, which is opus. And they said, hey, this is our most powerful. And it is the most expensive. And then you had your middle model, which is Opus. So essentially, again, large language models, three versions. So small, medium and large, right? And different use cases.

But a lot of this was for developers. Because if you are accessing something on the front end, obviously before there was a sonnet 3.5, which is only available for the middle model, you would just always use the most powerful model. Same thing with chat, GBT. If you log in, this is not for you. you 95% of the time, you're going to want to use GPT 4.0. If you're paying the, you know, $20 a month or whatever it is for the front end of chat

GPT, or if your team is using chat, GBT teams, if you're using the enterprise version of chat, GPDT, this is not for you. You're still going to use GPD 40, right? That's what you're paying for. This is if you are building products, if you're building services, if you are using something internally for your employees, and if you have an AI machine learning team, building things for you, this is what this is for. So I do want to do a, hopefully a good job of making that expectation clear.

So for the vast majority of our listeners out there, this is not going to impact your quote unquote personal or individual use, but this might impact things at your company level, depending on what your company does, depending on what you do. So I'm going to show just one more screen here and talk about some of the capabilities, because we talked about cost. in how much now more affordable this is. But also the benchmarks, y'all, the benchmarks were super impressive from GPT4-0 Mini. So I'm not going to go over all of them, but I think probably one of the most important ones to look at here is the MMLU.

So for lack of a better, lack of a better term, this is text only. but think of this as essentially an SAT or an ACT for large language models, right? This is a definitive score that is an apples to apples comparison. So all of these models out there, you know, researchers put them through very many different tests, right? So you have your human e-val, you have your math, you have your drop, you have a lot of these other ones, you have Hello Swag. You have so many different benchmarks, essentially, but you have to look at MMLU because I think that is, at least right now for text-only models, this is the most important evaluation. This is the gold standard. And right now, it's not even close.

GPT40 Mini, at least when you look at its most, it's kind of put most direct competitors, that's Gemini Flash and Claude Haiku. And it is blowing them away, right? So in an 82 score for a quote unquote small large language model is such a good score, right? And you see here, it's kind of small on my screen. But GPT40 Mini is about three to six points ahead of its competitors. Gemini Flash is second, Claude Haiku is third. But those three to five points on an MMLU is a lifetime of a different. That is huge.

You know, so as an example, when Claude released Claude 3 opus compared to GPT4 turbo, I believe Claude, uh, Claude 3 opus was like point one ahead. And they were proud of that, right? And rightfully so. So normally, you know, fractions of a point or a half point means, okay, you're doing good in this MMLU, right? It's a tight race. Everyone's trying to compete here.

The fact that GPT40 mini is multiple points ahead in the MMLU, you. Not only that, but then when you look at its two closest competitors in Gemini Flash and Claude Haiku, the price per million input output is also not close. So I would say that up until this release, for the most part, Open AIs was not the best source for building something internally. It wasn't, especially after Claude Haiku, uh, AWS just announced that you can fine tune Claude Haiku in their platform, which makes it a lot easier for companies to do this. So, you know, companies that were building internal models in-house based off other models,

if you're a startup building a service based off of one of these models, hey, at least for the last two to three months, I don't think Open AI was a good idea to use their API. And obviously, I think Open AI knew this as well. But I also like their strategy here. Open AI doesn't need to release a new model, right? Even though Claude just released 3.5 Sonnet, very good, right? Google keeps release, I don't know, their naming mechanism.

Who knows, 1.5 Pro or 1.5 advanced. You know, I'm sure a two of Gemini is going to come out soon. But no one from a large, large language model is touching Open AI right now. they are still exponentially better with their 4-0 model than anyone else. But where they were lacking, I think, was from a development standpoint, because the cost was too high and the output quality was too low. So this is a direct kind of play in that area because I think a lot of developers, a lot of big companies were either jumping ship or looking to explore elsewhere

because, if I'm being honest, open AIs models for developers were in theory antiquated because Claude and Google, I think we're responding a little quicker. All right. So that is it, y'all, for the AI News that matters. Let me give you the very quick recap. So first, we talked about some tech giants, according to reports that were accused of illegally using YouTube subtitles to train AI systems. We talked about some new updates in the Open AI versus New York Times case.

Essentially, Open AI saying, yeah, New York Times, proof. it. Go ahead and turn over reporters notes and interview memos for these millions of stories that you alleged that we, quote unquote, stole from you or violated your copyright. Next, former OpenAI co-founder, Andre Carpathie, has launched an AI-driven education platform called Eureka Labs. Next, meta is withholding, or sorry, meta is withholding future versions of its Lama AI model from the European and union amid some regulatory concerns. And then last but not least, OpenAI has unveiled GPT40 Mini, which I think is going to be a game changer for businesses who are building models or startups who are working off of this technology.

All right, a lot more in today's newsletter, but I do have to remind you one more time if you're still listening. Make sure to check out the show notes. Make sure to check out today's newsletter to go ahead and participate in our $1,000 live challenge tomorrow. You have to join us live. All right. You have to get your answers in in real time.

So make sure you join us. The link is going to be there, whether you want to do it on LinkedIn or YouTube. So make sure you join us for that challenge as we launch our thanks a million. So more on that tomorrow. If this was helpful, thank you. Please share this with your coworkers. If you're listening on podcast platforms, please subscribe.

leave us a rating and please go to your everyday AI.com. sign up for that free daily newsletter. Join us tomorrow. I'm going to be thanking you a million, hopefully giving away $1,000 if someone can get every single question right. Thank you for tuning in. Hope to see you back tomorrow and every day for more everyday AI. Thanks y'all.

Meet Firefly AI Assistant. Now live in Adobe Firefly, the Allman One Creative AI Studio. Just describe what you want to create in your own words and the assistant handles the rest, orchestrating multi-step workflows Adobe Creative Cloud apps, including Photoshop, Premiere Express, and more in one conversational interface. You direct the outcome while the assistant accelerates execution.

Stand control with the ability to step in and refine at any time. See it today at firefly.adobie.com. And that's a wrap for today's edition of Everyday AI. Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going. For a little more AI magic, visit Your EverydayAI.com and sign up to our daily newsletter so you don't get left behind.

Go break some barriers and we'll see you next time.