Resources:
Join the discussion: Ask Jordan questions on AI
Upcoming Episodes: Check out the upcoming Everyday AI Livestream lineup
Connect with Jordan Wilson: LinkedIn Profile
Try Our Free AI Prompting Course: Register for our free Prime, Prompt, and Polish AI Course!
Meta's New Approach, OpenAI's Delays, and Microsoft's Massive Investment
With pace of AI evolution reaching dizzying levels, the ethical spectrum remains a heated debate. There are growing concerns about the potential application of AI-generated content by Meta in developing ad hoc versions of AI models. The fundamental ethical contrasts between the contentious practice of social media scraping and the use of AI-synthesized data were at the epicenter of these discussions and will undoubtedly affect future trends.
Challenges and Hurdles for OpenAI
OpenAI, a formidable name in the AI sphere, has been making waves with its new media manager tool. Originally announced in early 2024 to assist creators in governing how their content is used for AI training, this development has been postponed into 2025. The delay is linked to concerns over intellectual property rights and various legal challenges, casting a shadow on the efficacy of current opt-out methods, which have been criticised as cumbersome and ineffectual. The hurdles faced by OpenAI may underscore a shift in strategic focus.
Generative AI and the Minefield of Copyright Issues
The use of copyrighted content for AI model training by AI providers has caused substantial debate, resulting in legal action. Content scraping is at the heart of these lawsuits, rousing questions about potential infringement. Legal pundits are skeptical whether OpenAI's Media Manager will resolve these IP complications, underlying the broader challenges the industry needs to address.
Microsoft's Mammoth AI Investment and AI Model Research
In contrast to these legal scuffles, Microsoft is concentrating on stepping up its game. The tech giant is planning to pour a staggering $80 billion into AI infrastructure. Aimed at reinforcing AI and cloud computing capabilities, this spend is believed to be a reaction to the increased demand following OpenAI’s ChatGPT release. Findings have unveiled secrecy surrounding the sizes of proprietary AI models, sparking industry intrigue. Trends indicate that these powerful AI models are shrinking, signalling meaningful implications for energy consumption, costs and scalability, and potential reductions in AI's carbon footprint.
The Rise of Autonomous Agents and AGI
While AGI (Artificial General Intelligence) agents began joining the workforce in 2024, a pivotal development in AI innovation continues with AGI agents gradually being integrated into workplaces by 2025. Existing AI solutions such as Microsoft's autonomous agents are already indicating the rise of AI agents in the workforce. OpenAI's new product, known as 'Operator,' is expected to launch in January, with further advancements likely with future model upgrades.
Meta Halting AI-Generated Profiles - A Backlash
Meta's experimental foray into AI-generated profiles faced various technical and ethical hurdles. The key issue was the mounting backlash over concerns of fabricated engagement. Despite its discontinuation, AI chatbots can still be generated by users, although they come with disclaimers over potential inaccuracies.
The AI Landscape - Upcoming Consumer Product Integrations
The forthcoming CES (Consumer Electronic Show) is set to focus heavily on AI advancements. Samsung is planning to unveil AI-enabled fridges, while LG plans to launch "affectionate intelligence," amidst speculation about its actual purpose. Similarly, NVIDIA's new GPU, the RTX 5090, rumoured to be 70% faster than previous models, is hotly anticipated.
In conclusion, from debates over ethical practices in AI content generation to awe-inspiring AI integration in consumer products, the AI landscape is witnessing an exciting, albeit challenging, period. As important stakeholders in this realm, businesses need to keep a close eye on these developments to exploit opportunities and stay ahead of the curve.
Topics Covered in This Episode
1. Meta & AI Content
2. OpenAI Media Manager Delay
3. Microsoft's AI Model Research
4. OpenAI's AGI Progress
5. Grok 3 Model Release Delay
6. AI Implementation in Consumer Products
Podcast Transcript
Jordan Wilson [00:00:16]:
Enough beating around the bush. We're going to have AGI agents in the workforce in 2024. Did I read that right? Yes, I did. That's some of the biggest news this week in AI, as well as, I mean, a little bit of everything else. Meta had a disastrous experimentation with AI generated profiles. OpenAI is looking beyond AGI to ASI, and AI is about to take over CES. Just an alphabet soup here to start your morning. Alright.
Jordan Wilson [00:00:58]:
There's so much going on in the world of AI. You could spend, I don't know, couple hours each and every day trying to keep up with it and say, hey. How is this gonna impact my company, my career, my department? You could do that, or you could let us do that. That's what we do almost every single Monday here on Everyday AI. So welcome. My name is Jordan. I'm the host of Everyday AI. This is a daily livestream podcast and free daily newsletter helping everyday people not just learn AI, but how we can all actually leverage what's happening to grow our companies and our careers.
Jordan Wilson [00:01:32]:
If that's you, maybe you're listening for the first time, welcome. Thank you for joining us. This is live. This is unedited. This is unscripted. It's It's the realest thing in artificial intelligence, and we do this every day. Where you can continue to do this is our website at your everydayai.com. There, you can sign up for our free daily newsletter where each and every day we recap the podcast episode.
Jordan Wilson [00:01:55]:
But you can also go on there, sort anything by category. No matter what you wanna learn, do you wanna learn about AI and HR? Do you wanna learn how artificial intelligence is impacting marketing or communications or the ethics behind it? It's all on our website, sort of by category, where we've talked to 100 of the leading experts in generative AI across the entire world. It's all there for you for free. Go watch, listen, read your everyday AI.com. Alright. Without further ado, let's get into the AI news that matters for the week of January 6th. Livestream audience, thanks for joining as always. We got Wall Street's warlord on YouTube.
Jordan Wilson [00:02:34]:
New face. Thanks for joining. Big big bogey joining. Brian, Joe, Peter joining on the Twitter machine. Marie Rolando. Joe, thank you all. Alright. If you haven't seen this piece of AI news, don't worry.
Jordan Wilson [00:02:49]:
It is fresh and hot off the presses. So OpenAI is now confidently pursuing AGI, artificial general intelligence, with now plans or at least now increased expectations for AGI agents to join the workforce. So in a blog post on his personal website, OpenAI CEO Sam Altman talked about strides the company is making toward achieving artificial general intelligence with a clear road map that introduces the introduction of AGI agents into the workforce by 2025. Yes. Let me repeat that. Open a Open AI CEO, Sam Altman, in a blog post just a couple of hours ago said he feels pretty confident that we will see AGI agents joining the workforce this year. Yeah. Wow.
Jordan Wilson [00:03:47]:
To think about, if you would have said this even, I don't know, 8 months ago, 10 months ago. So, the company has expressed strong confidence in its ability to build AGI, marking a pivotal moment in AI development that could reshape industries and job functions. So couple things, if you didn't read the post, and we'll be linking to it in our daily newsletter. They said that they would be gradually releasing it into the world for safety reasons and that they are confident that they know how to build AGI. So a direct quote from Sam's blog post saying, we believe that in 2025, we may see the first AGI agents. Oh, no. Sorry. AI agents.
Jordan Wilson [00:04:34]:
I won't go back and edit that. Don't worry. Sorry. AI agents, joined the workforce and materially changed the output of companies. So OpenAI's approach emphasizes the importance of gradually releasing technology into society, allowing time for adoption and ensuring that the benefits of AGI are broadly shared. So, yeah, to correct myself, OpenAI said that they are confident and know how to build AGI and that we will see the first AI agents join the workforce in 2025. So in the post, Altman also said OpenAI, yes, slipping this one in here, is actively aiming to superintelligence. So previously, same Altman said that they are thousands of days away to superintelligence, but now, at least on his personal, personal blog post, which, the the company, OpenAI, did, retweet on Twitter, his blog post, but they said that they're actively aiming to superintelligence.
Jordan Wilson [00:05:41]:
So if you are brand new here, maybe you're hearing all these acronyms. I know I've already dropped a couple on you already. I'm sorry. But AGI, artificial general intelligence, that's essentially when, a single AI system can perform, tasks at a higher level than almost any human. Alright? And depending on what definition you're looking at. Right? This is one of those where the goal posts are always moving. I've said this before. I had a dedicated episode.
Jordan Wilson [00:06:08]:
If you look at definitions of AGI from 15 years ago, we're definitely there. But as AI progress progresses, so does to the definition of AGI. So, Sam Altman saying, hey. We've figured out AGI, and we're gonna be slowly releasing it, and we're gonna see AI agents in the workforce. But also saying we're working toward or aiming to superintelligence. So ASI, if you haven't heard of that, the short way of saying it, it's past AGI. It's when you have AI systems that are smarter, not as smart as, but way smarter than the smartest humans at every single task and can also self improve. Right? They can, you know, it's when open or when ChatGPT is making the next version of itself and doesn't need humans to check its work.
Jordan Wilson [00:06:55]:
Right? So that's essentially what ASI is. Also, one other thing, they did share a little bit, which you don't really see, numbers shared a lot. But, OpenAI said that they've seen a weekly surge or sorry, a surge in weekly active users increasing from 100,000,000 to over 300,000,000. Yes. OpenAI and ChatGPT are winning the, kind of LLM or chat AI chatbot race, and it is not even close. Right? 2nd place, 3rd place, 4th place, 5th place combined are not even close to where OpenAI is at. I would say a big, reason behind that leap from 100,000,000 to 300,000,000 active or weekly active users? 2 things. 1, it's it's well, maybe 3.
Jordan Wilson [00:07:47]:
It's having the ChatGPT search. Alright? But then it's also having that available for free. So OpenAI, has really changed its free plan. So giving, at least limited use. Right? So now what they're trying to do is instead of using Google, they're like, hey. Use, ChatGPT. Even if you have a free account, you know, you can use ChatGPT search. Also, you can use ChatGPT as of a couple of months ago without even being logged in.
Jordan Wilson [00:08:13]:
So a lot of those things are really adding, to, kind of their their lead in that department. So we do know so when we talk about AI agents joining the workforce in 2025, is that actually possible? Well, I'd say it's already happening. Right? If you look at, Microsoft with their autonomous agents in Copilot Studio, if you look at, Salesforce, with their agent force, although I don't think that's gonna be very successful early on by charging per conversation. However, it's already it's it's not a far out, you you know, thought to think about. But we do know that OpenAI, according to reports, is set to be releasing operator or at least releasing some details around their operator agent that is supposed to be debuting some point this January. So I don't think that that will be the AGI threshold, right, that we've all been talking about. But, I do think that probably we will see, some sort of release this year from OpenAI that's like, yeah. This is definitely AGI.
Jordan Wilson [00:09:19]:
But I don't think we're gonna see that with this January operator release. I don't think it's until you combine that autonomous, kind of agent nature, if that is called operator, but combining it with a more powerful model. So whether that's a GPT 45, a GPT 5, an o one, an o three. So these o level models right now don't have access to any tools. So, maybe that could be the thing that pushes OpenAI over that line. Alright. Our next piece of AI news, Yeah. You're gonna be hearing about AI a lot this week.
Jordan Wilson [00:09:53]:
Sorry. It's gonna be like AI in fridges, AI in microwaves, AI in your gym sneakers. That's because CES is upon us, well, in a couple of hours. The Consumer Electronic Show, the biggest tech, show in the world is set to kick off tomorrow in Las Vegas, and it promises to be a significant event for tech enthusiasts, particularly those like us who are interested in AI. So couple things that have already been announced or leaked or previewed, Samsung is aiming, to introduce its AI for all everyday, everywhere initiative featuring, yeah, this is real, AI fridges that can anticipate temperature changes when groceries are unloaded. LG is introducing affectionate intelligence. Yeah. Can we stop trying to name rename AI, like Apple Intelligence? And now, LG is is saying they have affectionate intelligence.
Jordan Wilson [00:10:51]:
Can we stop that? That's not smart. Sorry. Anyways, LG is introducing affectionate intelligence, a concept that suggests a more personalized interaction with technology. And although what that actually means is not yet clear. Hey, LG. How about first work on, I don't know, creating dishwashers that work instead of, you know, trying to rebrand affectionate intelligence? Yeah. If you listen to the show, I've I've had a lot of problems with with dishwashers and washers and dryers, over the last couple of months. So, also, the TV industry.
Jordan Wilson [00:11:27]:
Yes. We are going to see generative AI in TVs. That's gonna be fun. NVIDIA is also expected to unveil, which might be one of the biggest announcements coming out of CES. We'll see. NVIDIA is expected to unveil its next generation GPUs with the RTX 5090 rumored to be up to 70% faster than its predecessor. Also, Microsoft is set to announce, and launch some more powerful AI features in its next generations of PCs, urging users to upgrade from Windows 10 to its newest operating system. Alright.
Jordan Wilson [00:12:07]:
So there's gonna be a lot more and we're gonna be, throughout the week, we're gonna be sharing about what's happening at CES. I do believe the majority of the more noteworthy noteworthy, announcements will be coming Tuesday. So make sure you check out our newsletter for that. Alright. So yeah. Big bogey face says currently at 0% of artificial superintelligence. Marie, I am not going to CES this year. Maybe next year.
Jordan Wilson [00:12:38]:
We'll see. Here's one that's fun. And by fun, I mean, I'm glad this happened. Meta has shut down AI profiles amid user backlash and technical issues. So Meta has decided to delete Facebook and Instagram profiles of fake humans. Right? So they essentially it was announced last week even though these, AI characters had actually been out incognito ish in the wild since September, but Meta did finally say, alright. We're gonna not do this whole AI profiles thing anymore. So, essentially, the company initially launched these AI powered profiles in September 2023 as part of an experiment.
Jordan Wilson [00:13:27]:
But user interactions have raised significant concerns, leading the company to essentially squash this project. So if you missed it, it did come out last week, and we covered it in our newsletter, news of this, even though Meta had been testing these AI generated profiles out, for more than a year. But news came out last week that Meta was going to start to roll these AI profiles out more broadly. And you might be thinking, why? Isn't that like spam? Isn't that like, you know, just potential for misinformation, disinformation? Well, Meta wanted to do it to increase stickiness and increase engagement on the platform. So, when you, take a photo of your, Thanksgiving dinner and put it on Instagram and you only got 3 likes and 2 of those were from your mom's normal account and your mom's burner Instagram account. Right? Meta's like, hey. If we just get a bunch of, you know, millions of AI generated profiles and they can start commenting and liking people's content, people will spend more time on our platform, which is an absolutely terrible idea. I am sometimes baffled, at the, you know, AI ideas or AI initiatives at some of the biggest companies in the world.
Jordan Wilson [00:14:44]:
Like, I know that people in Silicon Valley live in a bubble, but come on. Everyone everyone saw this and said, this is dumb. I don't know anyone that thought you you know what? You know what? We could use on social media some some some AI profiles, millions of them, please. Yeah. I would love to have fake engagement. So, anyways, the company, like I said, had previously introduced 28 AI personas in 2023, but decided to shut down most of them by summer of 2024 with the last profiles now disappearing recently. So here is an AI, feature that I'm glad gets squashed. We don't need AI everywhere.
Jordan Wilson [00:15:30]:
Like, we don't need, AI fake generated profiles on social media, please. So despite the removal of these profiles, users can still create their own AI chatbots, which can be used on the platform. So that includes various roles such as therapists and relationship coaches. Right? So this is more of like creating your GPTs, and then you can use them across Meta's different platforms. Also, Meta in those chatbots has included disclaimers warning that some messages may be, quote, inaccurate or inappropriate, end quote. So legal ambiguities regarding the responsibility of chatbot creators for the actions and statements of their AI companions as highlighted by a lawsuit against character AI, which claims the platform encourage harmful behavior in a user. Yeah. I don't think and I don't personally want to see these, you know, AI profiles on any platform.
Jordan Wilson [00:16:27]:
Right? There's already way too many bot farms, right, with fake, profiles on social media that are powered by artificial intelligence, but these are not affiliated with the actual companies that own said social media platforms. I think fake profiles are a huge problem on social media. And I don't think this was a good idea by Facebook to even or sorry, by Meta to even run these experiments, but I am glad, that they, squash them. So, yeah, livestream audience. What do you think? Juliet said this is the opposite of intelligence. Yeah. And, Michael said because social media isn't addictive enough. Yeah.
Jordan Wilson [00:17:07]:
They were just trying to make it stickier. Right? You go on and you post, I don't know, some quote you think is super inspiring, but doesn't inspire literally anyone. And then you get, you know, 30 comments from AI bots that are like, oh, this is the best thing ever. Right? Keep posting. No. Stop. Stop with this madness. Alright.
Jordan Wilson [00:17:26]:
Speaking of madness, yeah, we didn't see GROC 3. So Elon Musk's GROC 3 model is facing delays. So the anticipated launch of Elon Musk's GROC 3, which is axe AI's next major AI model, has been pushed back highlighting ongoing issues in the AI sector regarding product timelines and scaling limitations. So, Musk previously stated that GROC 3 would be released by the end of 2024. But, yeah, here we are in 2025, and there's no signs of Grock 3, which is to the surprise of probably no one unless you are absolutely, addicted to everything that Elon Musk says, and you eat out of his Twitter account like it's serial on a Saturday morning. So, yeah, I've said this all along. I'm not shy in saying that I don't think Grok is a serious model mainly because it is largely trained on Twitter data, which is a cesspool. So reports suggest that an intermediate model, GROC 2.5, may be released before GROC 3, indicating potential setbacks in development.
Jordan Wilson [00:18:35]:
So Musk admitted in an August interview that Groc 3's availability in 2024 would depend on luck. So first, he said, yeah. We're gonna release it in 2024. And then a couple of weeks ago, he said that, Groc 3's pretraining was complete. So everyone's like, oh, the pretraining's complete. That means we're gonna be getting the model. No. I do like XAI or Grok's move here, because they did kind of preview a web only interface.
Jordan Wilson [00:19:04]:
So at least moving the Grok AI platform off of Twitter, which I think is a smart move. However, because so much in the model uses real time data from Twitter, I don't think any companies will ever actually use it in production. Yeah. It might be fun to go on there and just have a second model or a 5th model to back up what you're already researching or if you wanna, you know, go use their roast me feature or to roast a certain, you know, x profile. But aside from that, I don't see any business anytime soon actually using Grok until they separate it from Twitter. So at least I like this direction, that they're going with having a dedicated web only interface, which I think will give business leaders the, proper way to test, Grok. And yes, they had they built their super gigafactory of compute and they're built like, buying like a trillion NVIDIA GPUs. But that does not matter if you still are using the Twitter slash x platform as one of your biggest pieces of pulling in new information.
Jordan Wilson [00:20:11]:
That is dumb. Right? Like, that's like meta, hopefully, is not using all of the content that was created by those AI generated profiles to create its next version of llama. Right? I think there's a difference between thoughtfully using synthetic data from a model, to either distill other models or to train models versus using what's on social media. That is a big difference. Yeah. Doctor Scott saying Musk question mark delays? Really? Yeah. Shocker. Right? Michael said I use Grok a couple of times because it's the quote, unquote free speech quote, unquote less restrictions model, and I came across the same restrictions and biases.
Jordan Wilson [00:20:54]:
Yeah. It's a bunch of marketing, not very good. Alright. Speaking of failed promises, OpenAI's media manager tool also delayed amid IP concerns and legal challenges. So OpenAI's anticipated media manager tool, which was designed to allow creators to manage how their works are used in AI training, has yet to launch 7 months after its announcement. This is another one of those things OpenAI said that it would be released in 2024, yet we are in 2025. And we haven't heard a peep on the media manager tool. So like I said, it was intended to identify copyrighted content across very various media types and reflect creators' preferences aiming to address criticism and legal challenges faced by OpenAI.
Jordan Wilson [00:21:44]:
Insiders suggest that the tool was not seen as a priority within OpenAI with reports indicating little internal progress or focus on its development. OpenAI has missed its self imposed deadline to have the media manager operational in the 2024 calendar year. So right now, leaving content creators, media publishers, and everyone else without a comprehensive opt out solution for their works. So right now, current methods for creators to opt out of AI training, such as submission forms for removal, has been criticized as cumbersome and ineffective. So, yeah, there actually is a way on OpenAI's website that you can, essentially say, hey, OpenAI. Stop scraping my website. Stop using my content for your models. It is not super intuitive.
Jordan Wilson [00:22:33]:
So that's why OpenAI did announce the media manager, which when we first covered it in our newsletter, in early 2024, I thought it was a great idea if it came to fruition. However, it seems like, I don't know, maybe AGI and ASI have suddenly, become more important than, you know, the data. And here's the other thing. These large language models, these big tech companies, they need the Internet. There's this you know, the big elephant in the room is, well, these models are essentially scraping copyrighted content and using it to train their models. And we're gonna be talking about a lot of lawsuits in 2025 that may or may not prove that exact same thing. Right? That's the the the big secret in the room that is not really a secret. All of these big tech companies are scraping the open Internet, which includes obviously copyrighted works.
Jordan Wilson [00:23:23]:
Right? And it's it's essentially that's how generative AI works. Right? It's instead of reproducing a one to one copy of something that is copyrighted, instead, it uses, you know, maybe dozens or 100 or thousands of variations of different copyrighted content and essentially blends them all together and spits something out that is technically unique, but technically, just derived from dozens or 100 or thousands of pieces of copyrighted content. Right? So there's a a very, informal lesson in how generative AI and large language models work. So OpenAI faces, like every other big tech company, faces class action lawsuits from multiple creators and large content organizations, including authors in media conglomerates claiming that their works were used for training without permission. Yeah. But it is hard, to prove unless a model spits something out verbatim, which obviously, you know, you can, kind of scheme to do that with some, you know, fancy prompt engineering. So, yeah, we're gonna be seeing this a lot in the courts in, 2025. So legal experts have expressed skepticism about the media manager's ability to resolve ongoing intellectual property issues, highlighting challenges and ensuring compliance and the burden it places on creators to manage their own rights.
Jordan Wilson [00:24:42]:
Right? Yeah. It's kinda like and this isn't just OpenAI. This is everyone. Right? But kinda like the unofficial, unwritten rules right now are like, hey. Yeah. We're gonna essentially take every single piece of content that's on the Internet, whether you give us access to it or not. Right? Unless you jump through some hoops and make it hard for big tech companies to not scrape it and to not use it all. But that's how the Internet works now.
Jordan Wilson [00:25:07]:
Right? FYI. Speaking of big tech companies, Microsoft is planning for an $80,000,000,000 investment in AI infrastructure for the calendar year. So Microsoft has announced a significant investment aimed at enhancing its capabilities in artificial intelligence and cloud computing. So the tech giant plans to invest approximately 80,000,000,000, that's with a b y'all, $80,000,000,000 this year to develop data centers specifically designed for training AI models and deploying AI driven applications. So the investment comes in the wake of a surge in generative AI interest, obviously, following OpenAI's ChatGPT launch in late 2022 as businesses across the globe scramble to try to use as much AI as possible. Well, what does that mean on the back end? Number 1, that's why NVIDIA is has been the most valuable company in the world by market cap, which I told y'all like 2 years ago and no one believed me. But also it means that data centers are growing. It it it's meaning that we need more, more power, more energy to power all of these, you know, giant clusters of GPU chips that are running, you know, inference in the cloud.
Jordan Wilson [00:26:24]:
Right? So, you know, if your company is using AI in pretty big way or maybe, you know, you or your team are heavy users of, you know, ChatGPT or Claude or Gemini or whatever. Right? Somewhere in the world, that inference or that compute is happening. It's happening in the cloud. Right? So, these data centers are gonna get bigger and bigger. You know, that's why we're talking about, nuclear power. That's why we're talking about, you know, where these, data centers are even located and building, you know, these huge new data centers that are essentially, you know, tens of thousands or hundreds of thousands of GPUs. So that's what happens, right, unless you're using edge AI, which is a large language model that can run locally. Right? Because then all of that is happening on your device.
Jordan Wilson [00:27:12]:
But the majority of the power applications of these, you know, AI chatbots or whatever you wanna say, it's cloud compute. Right? So Microsoft announced they are doing $80,000,000,000 this year in new AI infrastructure developments. So as the primary backer of OpenAI, Microsoft is obviously positioned as a key player in the competitive AI landscape benefiting from its exclusive partnership with the a partnership with the AI chatbot developer. Also, more than half of that planned $80,000,000,000 investment will be directed toward projects in the US. Alright. Speaking of Microsoft and large language models, some dorky news here, which I loved getting my hands on this one. Instant screenshot, instant read from me. But Microsoft's new research has revealed some unforeseen insights on AI model performance and parameters.
Jordan Wilson [00:28:10]:
So in a new study from Microsoft, they've estimated the sizes of powerful AI models. Why does that matter? Well, the size of all of these large language models is typically shrouded in secrecy. No one knows. No one knows. Right? No one knows the size of these large language models. For the most part, companies aren't, at least when we're talking about proprietary models, for the most part, companies aren't announcing it. It's different with open source models that you can download and fork. Well, because you're downloading them, so you can kind of deduce, the size or the parameters in these models.
Jordan Wilson [00:28:49]:
Right? So even if you're a nontechnical person, here's why this is important. That scenario that I just described for you. Right? Let's say you're, your company, there's 10,000 employees and all of you are using generative AI. Well, there's a good chance that that company, like OpenAI, is maybe losing money on that. Right? Especially if you are power users. Right? If you're paying $20 a month for ChatGPT Plus, you're paying, I think it's like 50 or $100 now for an enterprise seat. Right? But all of that power that you are needing from the model is being drawn from a data center. It is very costly.
Jordan Wilson [00:29:24]:
It is sucking up a lot of energy. So models, presumably, even these large language models have been getting smaller. But until this study from Microsoft, we really didn't know it because the smaller the models get, the essentially less power, they suck. However, most proprietary models don't want to say, hey, here's what's, you know, under the hood of our model, because that's proprietary information. That's the secret sauce. Right? Like the weights of these models. So the the research is showing that the models, even the large ones, are getting significantly smaller. So here's some of the deets.
Jordan Wilson [00:30:02]:
Ready? So, Microsoft suggests that as an example, Claude 3.5 SONNET contains a 175,000,000,000 parameters, whereas o one preview from OpenAI boasts 300,000,000,000 parameters. A good benchmark, which is one of the kind of first, models, large language models that it was widely reported to know the size was GPT 4. The original GPT 4 was reportedly about 1.8 trillion parameters. So they're they're as an example, Claude 3.5 SONNET. Let's see. I'm not good at math on the fly, but I believe that's about 10% of the size. So as models get smarter, as the process of, you you know, reinforcement learning with human feedback gets a little better, as data source quality improves. Post training improves.
Jordan Wilson [00:30:56]:
The models are essentially able to get smaller, which is huge. So OpenAI's smaller models, the o one minutei and GPT 4 o minutei are estimated to have, 100,000,000,000 and 8,000,000,000 parameters, respectively. So let me repeat that. So GPT 4 o mini, which is more capable now than the original GPT 4 turbo or sorry, the original GPT 4. GPT 4, 1.8 trillion parameters. GPT 4 0 Mini, 8,000,000,000. Alright? It's a fraction of the size. So why is this important? Well, there's been all this conversation lately about AI hitting a wall, scaling, hitting a new wall.
Jordan Wilson [00:31:47]:
Right? That would be true. I think people who are saying that don't truly understand the impact of these models being smaller and smaller because what that means, essentially, I think, you know, let's say we have a GPT 5. Right? And there's a GPT 5 minutei. Alright? There's a likelihood that that GPT 5 minutei could live on a device. Right? As computer hardware, so as these GPU chips in our computers, as these NPU chips, these neural processing units that are kind of like referred to as AI chips, right, as these TPU, these data center chips, Right? As the chips get more powerful and at the same time as models get smaller, guess what that means? That re that reduces kind of this carbon footprint, but here's the big piece. Right? One of the biggest reasons that AI adaption hasn't been like a 1000% is companies rightfully so are worried about their data. They're worried about their privacy. Right? Because they're essentially like, oh, we don't wanna send all of our data and all of our information to the cloud, which is not smart because guess what? Big companies, you're already storing all that information anyways on cloud servers, but I'll I'll leave you to just go ahead and, you know, think about that yourself in the quarter.
Jordan Wilson [00:33:09]:
Like, oh, okay. Why am I okay with us, you know, using cloud hosting for all of our important data, but I'm not gonna use that same provider to use AI models? Whatever. That's your own, you know, really smart decision to ponder over. However, you know, as these models become smaller in parameter size, that means in theory, we will be able to run proprietary state of the art large language models on device edge AI, running these locally in the very near future. Therefore, you won't need to worry as much about privacy and security because you're not sending all of that data, which a lot of times it's proprietary sensitive, whatever. You're not sending it, to a third party. It is on prem. It is on your device.
Jordan Wilson [00:33:54]:
It is edge AI. So this, new study, from Microsoft might seem dorky. It's a bunch of numbers, but it is actually huge, because, like I said, we haven't really known, aside from GPT 4, which we kind of knew from a lot of reporting, was about 1.8 trillion parameters. But we didn't really know how much are these how big are these newer models. Well, they are a fraction of the size, and that is why, you know, when everyone's talking about, oh, AI is gonna hit a wall. No. It's not. Because these models now are less than 1% of the size and exponentially more powerful.
Jordan Wilson [00:34:36]:
Those two things cannot coexist and not lead to exponential scalability. Period. I don't know. I think sometimes I like to just make people like to make these, you know, weird claims, like, oh, hey, I see the wall. Right? It's like, oh, right? Like, as a reason for not using it or as a reason to, like, I don't know, be like a hipster skeptic of AI. Right? But that's I don't know. 1 plus 1 in this instance is going to equal 2. No matter how much you wanna throw in this common core math that I don't understand.
Jordan Wilson [00:35:06]:
Does anyone know common core math? Does 1 plus 1 still equal 2? I don't know. Alright. That's it y'all. Let me go over a quick recap of all the AI news that matters. Sorry. God, I don't know. A little weird tangent there. So, here we go.
Jordan Wilson [00:35:24]:
Story number 1, OpenAI through, CEO Sam Altman's blog post saying that they, are confidently pursuing AGI. They know how to build it, and they're planning for AI agents to join the workforce in 2025. Next, we're gonna be seeing a ton of AI, probably too much, at the CES Consumer Electronics, show in, Las Vegas this week. Next, Meta thankfully shut down their AI profile experiment amid user backlash and technical issues. Next, Elon Musk GROC 3 model is facing delays and was not released in 2024 as promised. Speaking of delays, OpenAI's media manager tool was delayed amid IP concerns and legal challenges. Microsoft is reportedly planning $80,000,000,000 investment in AI infrastructure, new AI infrastructure for this year. And new Microsoft research has is revealing some new and, I think, exciting insights on the size of large language models.
Jordan Wilson [00:36:31]:
Alright. Was this helpful, or do you prefer spending hours every single day sifting through reading the news and worrying about how it might impact your career? That's what I do. Stop wasting your time. We do this. So almost every single Monday, we do this AI news that matters. Right? It's makes sense. Right? I do this every day for a living. I talk to the smartest people in AI.
Jordan Wilson [00:36:55]:
We put together, an AI newsletter. I read over the AI news every day, so just join us Mondays. Please Please join us every day. But if you can only join us once, join us Mondays to go over the AI news that matters. If this was helpful, if you're listening on the podcast, please subscribe to the show. Leave us a rating. I'd super appreciate it. Also share this with your network.
Jordan Wilson [00:37:12]:
If you're, on LinkedIn, if this is helpful yeah. I know people tell us everyday AI is your cheat code. It's your secret helper. Okay. That doesn't help us stick around. Right? If you, keep us as a secret, we can't grow. So please share this with your friends, your neighbors, your coworkers, your friends' neighbors' coworkers. So thank you for tuning in.
Jordan Wilson [00:37:32]:
Please join us tomorrow and every day for more everyday AI. Thanks y'all.