Ep 542: Apple’s controversial AI study, Google’s new model and more AI News That Matters

Topics Covered in This Episode:

1. OpenAI's Advanced Voice Mode Update
2. Reddit's Lawsuit Against Anthropic
3. OpenAI's New Cloud Connectors
4. Google's Gemini 2.5 Pro Release
5. DeepSeek Accused of Data Sourcing
6. Anthropic Cuts Windsurf Claude Access
7. Apple's AI Reasoning Models Study
8. Meta's Investment in Scale AI

Keywords:

Apple AI study, AI reasoning models, Google Gemini, OpenAI, ChatGPT, Anthropic, Reddit lawsuit, Large Language Model, AI voice mode, Advanced voice mode, Real-time language translation, Cloud connectors, Dynamic data integration, Meeting recorder, Coding benchmarks, DeepSeek, R1 model, Distillation method, AI ethics, Windsurf, Claude 3.x, Model access, Privacy and data rights, AI research, Meta investment, Scale AI, WWDC, Apple's AI announcements, Gap year, On-device AI models, Siri 2.0, AI market strategy, ChatGPT teams, SharePoint, OneDrive, HubSpot, Scheduled actions, Sparkify, Vo3, Google AI Pro plan, Creative AI, Innovation in AI, Data infrastructure.

Podcast Transcript

Jordan Wilson [00:00:17]:
Why is anthropic in hot water with Reddit? Will OpenAI's ChatGPT become the de facto business AI tool? Did Apple make a huge mistake in its buzz worthy AI study that had just released on large reasoning models? And why the heck did Google release a brand new version of Google Gemini when it was already on top? Yeah. A lot happening as always this week in the world of AI news. And if you missed anything or if you're just spending so much time wondering what all of these updates mean for you and your company or your department. Don't spend hours a day doing that. Instead, just spend your Mondays with us here on everyday AI as we break down the AI news that matters. Alright. What's going on y'all? My name is Jordan Wilson and welcome to everyday AI. This is your daily livestream podcast and free daily newsletter helping us all not just keep up with AI, but how we can use it to leverage, all this new information, all this new technology to grow your company and career.

Jordan Wilson [00:01:28]:
So if that's what you're trying to do, you are definitely in the right place. It starts here on the unedited, unscripted daily livestream, but you finish. You actually grow your company by reading our newsletter. It's for free on our website at youreverydayai.com. So make sure you go there and sign up. Make sure you can go check out also more than 400 and, 40 now back episodes on our website, all for free categorized. So whatever you're trying to learn to get ahead, to get that edge, it's already on our website. We've already probably interviewed a world leader in that field.

Jordan Wilson [00:01:59]:
So make sure you go check that out. Alright. Enough chit chat. Like I said, everything from Infrapping being in trouble with Reddit to open AI's, pretty big new connectors, that I think are gonna change how companies use AI. I think this new study from Apple got a lot of things wrong in Google. Yeah. We have a new version of Gemini 2.5 Pro. Alright.

Jordan Wilson [00:02:24]:
Let's get into the AI news that matters for this week. What's up, livestream audience? Good to see you. Everyone from, Al joining us on YouTube from Scotland, Georgie joining us, from Jamaica, Kimberly on the LinkedIn machine, from New York City. Jay woke up somehow in West Virginia. Doctor Harvey Castro, thanks everyone for joining us. Alright. Let's start first. Probably the most recent update in terms of it's only been a couple of hours, but OpenAI has rolled out a major new advanced voice mode upgrade.

Jordan Wilson [00:02:58]:
So OpenAI has launched a significant upgrade to its advanced voice mode in ChatGPT, which is now available to all paid users across platforms according to the company. So the revamped voice mode delivers much more natural speech with improved annotation, realistic cadence, and expressiveness that can capture emotions like empathy and sarcasm. So users can now access real time language translation as well. That one's pretty cool. By simply requesting it, allowing continuous two way translation throughout conversations. So that's an enhancement aimed at, you know, travelers, global professionals, but that one's pretty big. Right? By just saying, like, hey. Act as a translator.

Jordan Wilson [00:03:43]:
You're gonna hear, you know, two different people speaking two different languages and, the new advanced voice mode. Well, if it works correctly, we'll take care of the rest. So this new update builds on earlier improvements to accents and interruption reduction, making voice mode more reliable for diverse users. So these new features are designed to make interactions feel more human and seamless and a little less robotic, which could benefit anyone relying on voice AI for communication. So, I gave this a try. One thing I think this does a little bit better on, is is starting out in advanced voice mode. One thing that I've especially when I'm on the phone. One thing I I never really know if advanced voice mode is, like, ready to go.

Jordan Wilson [00:04:27]:
So this new, update, I think, helps with that just a little bit kind of that first response. That's just my, take right now. OpenAI did also note that there's some limitations, including occasional drops in audio quality and some rare instances of unintended sound. So not quite, you know, as, eerie as the, the one last year, right, where, the advanced voice mode was, like, saying help. Right? Like, help. Get me out of there. So nothing will write that, but nobody I did say that this isn't perfect, and there's probably gonna be some, mistakes. So, livestream on it, have you guys, you know, tested the new advanced voice mode out? And I think probably I haven't actually done a lot of, kind of coverage on the show of these different voice modes.

Jordan Wilson [00:05:18]:
Right? I think I did the original one, on, a show here. I think one of the reasons is it's sometimes hard to capture, to capture that, on the phone. And so maybe I'll just have to, like, grab my wife's phone because I use, you know, I use my phone to, you know, shoot the, shoot the video here for the live stream. So, yeah, let me know if if if I should do the, you know, the the new advanced voice mode. You know, Google Live is obviously really good. Perplexity has theirs. I just don't know if our audience, podcast people as well. I always put my information in the show notes.

Jordan Wilson [00:05:51]:
Let me know if that's something you'd wanna see, kind of a a voice off battle or at least going over, what these voice modes can and can't do. Let me know if that's something you wanna see. Doctor Harvey Castro says Eleven Labs and Hume AI are better at emotional intelligence. Yeah. We did talk about this on the show, you know, last week on the AI news that matters. Eleven Labs has their v three. I don't know. I'd say this new version, maybe not.

Jordan Wilson [00:06:19]:
I'd say the, obviously, the Eleven Labs and Hume, maybe four days ago were better, in in this regard. I don't know. I'd say right now, this new advanced voice mode update that just rolled out hours ago, would at least put, OpenAI in the, you know, either one a or one b conversation where I think maybe before, there was a little bit of a drop off in terms of quality. Cecilia says, yeah. Do a voice off show. Brian says it as well. Al here on the YouTube machine says, would like a feature where it could just listen and transcribe. Yes.

Jordan Wilson [00:06:52]:
My gosh. I would love that. Right? Which is, you you you know, it's not that big of a deal. You know, generally, I would just use a third party tool, but I would love that as well to just be like, yo. Just don't say anything. I'm just gonna, you know, word vomit here for fifteen minutes. I've tried it before, and a lot of times it just cuts out. So yeah.

Jordan Wilson [00:07:10]:
Al, I like that idea, you know, just to be, like, have a translation mode. But just to be able to talk to advanced voice mode for that, I think that would be cool. Alright. Our next piece of AI news, anthropics in trouble. Well, because, some of the other big companies like OpenAI, and Google have paid tens of millions of dollars to partner, with Reddit for their data. And apparently, according to a new lawsuit, Anthropic is just trying to grab it all for free. So Reddit has filed a lawsuit against Anthropic in California superior courts, accusing the AI startup of illegally scraping Reddit user comments to train its clawed chatbots. So the lawsuit claims that Anthropic used automated bots to collect content from Reddit despite explicit requests not to and without user consent, raising fresh concerns over privacy and data rights as AI companies raise to improve their models.

Jordan Wilson [00:08:08]:
So unlike other lawsuits targeting copyright infringement, Reddit's case focuses on a breach of its terms of use and claims anthropics actions accounts to an unfair competition. So like I said, Reddit has pretty lucrative licensing agreements right now with OpenAI and Google and other companies that pay for access to its data, enabling the platform to enforce privacy protections and user rights. So yeah. Like, when you, are using Reddit, you know, you're essentially being like, yo. Yeah. I'm giving, you know, Reddit, permission to give this data to OpenAI and Google to train their models. These are agreements that helped Reddit prepare for its public stock market debut last year. So Infropix CEO and researchers have previously documented the value of Reddit's subject matter, forums for AI trading, and the company maintains its use of web data is lawful and essential for developing language models.

Jordan Wilson [00:09:07]:
So this lawsuit highlights growing tensions between the content platforms and AI developers, underscoring the stakes for companies relying on user generated content and the potential impact of how publicly available data may be used or restricted in future AI advancements. So it's no secret how valuable Reddit data actually is. Right? So many times if you're looking at the sources of information when you're asking in in AI chatbot a question, so many times it comes from Reddit. You could actually make the argument that Reddit aside from, you know, maybe a source like Wikipedia, you can make us you you can make a claim that Reddit is potentially one of the most or if not the most important, single website for training data. One of the reasons is a lot of the information that, you know, these these AI, labs kind of just scrape on the open Internet are available in kind of, like, there's redundancy. Right? Like, even with Wikipedia. Right? A lot of that information, especially when it's factual, exist in a handful or dozens or hundreds of different sources. Reddit, not so much because these are individual, a lot of times, subject matter experts sharing their expertise.

Jordan Wilson [00:10:23]:
Right? If you don't know a ton about Reddit, Reddit has actually been a huge part of many growing organizations' content strategy. So, you know, whether small business, entrepreneurs, big business, will put out original helpful content on Reddit either asking or answering people's questions or just, you know, starting topics on, you know, news developments, etcetera. There's actually been a lot of companies that literally exclusively post on Reddit and don't even post on their company blog just because Reddit is so highly trafficked. So this one might be at least for me, the second or third most interesting, kind of AI lawsuits to keep an eye on outside of the New York Times versus OpenAI and Microsoft because this one is huge just because of the value of the data. Again, I would even venture to say Reddit data is more valuable ultimately than data from the New York Times. That pains me as a former journalist to say, but so much of what's in the New York Times is available on dozens of other news publications, whereas a lot of time data on Reddit is exclusive to Reddit. And, you know, sometimes it's more of the the nuance and human expertise that really helps make these large language models better and, more capable. So, yeah, Marie saying anthropic scraping the Internet.

Jordan Wilson [00:11:42]:
What? So what else is new? Yeah. Reddit data is also all over perplexity. It's everywhere. Right? And obviously, people write, about Reddit data and source insight Reddit data on their own websites and other websites as well. You know? So there may be some some workarounds or technical loopholes, on that. And this is obviously extremely hard to police, but it is one that we should all be keeping an eye on. All right. Next piece of AI news seems small, actually big Chad GPT and open AI making a huge play in the business, sphere here with its new, connectors for paid users as well as a new chat g b t meeting recorder.

Jordan Wilson [00:12:27]:
So OpenAI has introduced a new record feature in chat g b t teams for Mac OS, letting users record, transcribe, and summarize meetings on voice notes directly within the app. They've also, released their new cloud connectors, which I'll talk about here in a minute. But the new recorded tool can transcribe up to a hundred and twenty minutes per session, and it can turn audio into editable summaries and generate emails, project plans, or even code from conversations positioning ChatGPT now as a new competitor, to companies such as Otter dot ai or Zoom's transcription features. The OpenAI update also brings cloud connectors, and this one is big for Google Drive, OneDrive, Dropbox, Box, and SharePoint, allowing ChatGPT business users to search and analyze documents from these platforms while respecting existing user positions, permissions. Sorry. So I did kind of a show on this, but it was technically on, Claude's version of this. So, if you scroll back a couple of episodes to episode five thirty nine, that show was the one Claude the one new Claude feature that changes knowledge work and how, to use it. And, actually, in that show, I called.

Jordan Wilson [00:13:51]:
I'm like, Chad GPT should be releasing this feature any day now. And, funny enough, they announced it later that day. But this is pretty big. And and and I'm not saying that rag is is dead. Right? So traditional retrieval augmented generation. And you might be wondering, like, okay. What the heck is this and why is this actually important? Well, let's look at some of these connectors that I have on screen for our live stream audience. So I mentioned things like, you know, Google Drive, Gmail, I mean, Google Calendar, SharePoint, Outlook, Teams.

Jordan Wilson [00:14:23]:
One I'm excited to try is HubSpot. Right? Being able to chat with your dynamic data is huge. The downside at least on the OpenAI side and maybe one reason that, Anthropic's Claude still has an advantage at least in, connecting to this enterprise, user data is because right now, at least in chat, it's only available in deep research mode. So you might have to, you know, wait anywhere from three to fifteen minutes, and you can select which of your connectors that you want OpenAI's deep research to go through. So I don't think that's necessarily a bad thing because in almost everyone's experience, you have a, lower likelihood of hallucinations when you do the deep research mode. It uses a more powerful tool. The last we heard from OpenAI, they actually use dual o3 modes. It's actually using two versions, two separate versions of their very powerful o3 model, to go through and, do different things.

Jordan Wilson [00:15:22]:
So, this one, the connectors space is booming, and I called this last week. I said, you know, even though technically, anthropic beat everyone else, I said this is gonna be rolling out to all the major labs. Obviously, you know, OpenAI responded like that that day. This one's this one's big. Right? And like I said, I don't think this kills traditional rag. Right? So what that means is if you're asking, let's say you're in a, in a marketing role at a huge logistics company. Right? If you're talking to ChatGPT or Gemini or Claude or anything else, right, about, let's say, industry specific news, you don't always have a ton of control. It might bring it might bring in things from its own internal training, data which could be extremely outdated.

Jordan Wilson [00:16:10]:
It could be wrong. It could not have the level of expertise that you'd expect by being a subject matter expert and being a marketer in the logistics industry. Right? Or it might browse the web, and that's again, roll the dice. So by having it first look at your data, that is huge. That is essentially, you know, the the the promise of retrieval augmented generation, which is, you know, companies, you know, two to3 years ago spent millions of dollars essentially fine tuning models with rag. Right? Creating these, embeddings in their vector databases, you know, essentially creating, their own version of a large language model, but that first and primarily used its own internal data, which was a very expensive and laborious process. And now we're getting that with a couple of clicks. Right? And this is one of the reasons, if I'm being honest, I was never pushing RAG, super hard because I knew that this day would come.

Jordan Wilson [00:17:03]:
Right? Competing on the application layer is how companies like Google, OpenAI, and even Claude are going to, compete in the long run. You know, as we talk about, like, you know, knowledge being commoditized, but also that large language models could largely just be swappable. Like, right, replaceable. So I think how companies are ultimately gonna connect is these deep, and dynamic integrations with our dynamic data. So pretty, pretty big pretty big news here. And and, yeah, live stream audience. Let me know if you've tried these connectors at all. I've tried them, very impressive.

Jordan Wilson [00:17:38]:
I'll probably be doing follow-up, shows on this either later in June and July. There's so many new, kind of areas of, these large language models. It's it's difficult to even just understand how people are using them. And, yes, I talk to people aside from just being on this livestream every day. I talk to businesses, you know, from small businesses, startups, Fortune 500 companies, huge companies doing tens of billions of dollars of revenue. I'm constantly talking with them or consulting. You know, they hire us to help them, you know, learn a a a large language model. So it is actually difficult right now over the last couple of weeks to even understand where the interest is.

Jordan Wilson [00:18:16]:
Right. So please reach out and let me even know what are you interested in learning about. All right. Our next piece of AI news, for some reason, Google has just woken up and chosen to dominate even more, because they released a new version of its Gemini 2.5 pro model, even though their previous version of Gemini 2.5 pro was already by far the most powerful and capable large language model. So Google has begun rolling out a preview of its upgraded Gemini 2.5 pro model, which will be generally available in the coming weeks. So this new version is labeled six zero five or June 5. Right? So it's been out for just a couple of days, not to be confused with their previous version, which was actually five zero six. I kinda wait like, I kinda wish that Google would have, like, maybe waited a day or roll this out a day sooner.

Jordan Wilson [00:19:16]:
Because if you're dyslexic, you're probably gonna see these as the same model and you're gonna be like, wait, which is this the o six zero five or the o five zero six? Regardless, it's showing significant improvements in coding benchmarks, challenging tasks such as GPQA and even humanities last exam, which is one of the most challenging, kind of, AI benchmarks for large language models that ex that assesses math, science knowledge, and reasoning skills. So this new update also addresses user feedback about performance drops from their previous model outside of coding, and Google promise improved style and structure for more creative and better formatted responses. And, yeah, in terms of Elo, right, we talk about the LM arena, pretty often here on the show. This is where you can go in. You put in input. You put a prompt in. You get outputs from two different models. You don't know which one's which.

Jordan Wilson [00:20:13]:
You choose which one's better, and then you get the ELO score. And Gemini 2.5 pro, the new version, the June 5 version, actually got a 24 jump over its previous version. So now Google, their 2.5 pro is number one and number two on the LM arena board. So the model upgrade is available through the Gemini API or via Google AI Studio and Vertex AI. So, pretty exciting. And if you remember when I covered Claude four, what what episode was that? If you wanna go back and, and listen to it, where where was that? Here we go. Episode five thirty four. This was on May 28 when we, talked about Claude four because one of the things that Anthropic really hung its hat on is, you know, all of their scores in software engineering, web development, etcetera.

Jordan Wilson [00:21:07]:
And I said, yeah. Good luck with that anthropic because that's gonna last a whole maybe week or two until Google, updates their Gemini 2.5 pro. Sure enough, within about ten days, you know, they wipe out at least external benchmarks, that Anthropic's Claude had yet. They're gone. They're gone. So Gemini, 2.5 pro on the L M Arena now holds the top mark in every single category. It's silly. It's absolutely silly.

Jordan Wilson [00:21:41]:
So, alright. Let's take a quick break for words from our sponsors at Google.

Google Gemini [00:21:50]:
This podcast is supported by Google. Hey, everyone. David here, one of the product leads for Google Gemini. Check out v o3, our state of the art AI video generation model in the Gemini app, which lets you create high quality eight second videos with native audio generation. Try it with the Google AI Pro plan or get the highest access with the Ultra plan. Sign up at gemini.google to get started and show us what you create. This podcast is supported by Google. Hey, everyone.

Google Gemini [00:22:27]:
David here, one of the product leads for Google Gemini. Check out v o3, our state of the art AI video generation model in the Gemini app, which lets you create high quality eight second videos with native audio generation. Try it with the Google AI Pro plan or get the highest access with the Ultra plan. Sign up at Gemini.Google to get started and show us what you create.

Jordan Wilson [00:22:59]:
Alright. And thank you to our partners at Google for sponsoring the show. Alright. Speaking of Google, our next piece of AI news, well, Chinese AI Lab DeepSeek is accused of using Google Gemini's data to train its powerful new updated r one model. So Chinese Lab DeepSeek's newly released r one, and this is the May 28 version, is making headlines for its strong performance on math and coding benchmarks, but some researchers suspect it was partly trained on data from Google Gemini's family of models. So, developer Sam Paik and others have published evidence that DeepSeek's r one zero five twenty eight uses language and thought traces strikingly similar to Gemini 2.5 pro fueling speculation about data sourcing in model training ethics. And if you pay attention to the show, you know this is not DeepSeek's first time under scrutiny for doing this exact same thing. As last December, its b three model was observed identifying itself as ChatGPT, further raising concerns about unauthorized use of rival AI outputs.

Jordan Wilson [00:24:18]:
So OpenAI previously told the Financial Times that it found evidence that DeepSeek had used distillation on OpenAI's models, which is a method that extracts training data from larger models, and Microsoft also flagged suspicious data, exfiltration from OpenAI developer accounts affiliated with DeepSeek in late twenty twenty four. So, essentially, now you have a pretty serious, at least, whether it's accusations or, you you could say solid proof that DeepSeek has distilled, or, you know, kinda just borrowed from, OpenAI, Microsoft, and Google to help train its data. So, you know, when when when you read all these stories about, oh, you you know, DeepSeek is the future of AI because they can train their their models at a fraction of the cost. Well, according to reports and executives at these exact same companies, they're doing this because they're literally just distilling, so from the companies that are actually paying the money to create, train, fine tune the models. So yeah. Hey. Where where are you at all you, deep seek people? You know, I'm pretty sure you were in the comments back in, December saying how, hey. In a couple of months, you know, OpenAI and and Google are gonna be irrelevant, you you know, because deep sea can train models for, you know, one one hundredth of the cost.

Jordan Wilson [00:25:43]:
Yeah. No. They can't. Right? Go look at the semi analysis, reports on that. I'm gonna go pull up, what what show was this. I did a a a deep dive on DeepSeek, a couple of weeks after all the all the hoopla. So go listen to episode four sixty on that, and I break down the artificial analysis report that essentially looked at DeepSeek and they're like, wait. They did not train this model for, you know, $5,600,000, which is like 5% or 2% of what it would actually cost to train.

Jordan Wilson [00:26:16]:
They actually broke down their actual cost. So, if if you wanna know the truth on deep seek aside from this recent news that's, you know, there, some researchers are saying, yeah, they're just distilling from other models. Go listen to episode four sixty if you want the deep dive. All right. Joe here saying won't touch deep seek, Chinese government surveillance, worm. Yeah. So, if you're using deep seek FYI via the API or on their website, you are sending all of the in any information that you upload, straight to the Chinese government. So this isn't me being political or me doubting open source.

Jordan Wilson [00:26:56]:
I'm all for open source. Right? I'm all for open source, open weight models, but I don't know. Do you wanna send all of your data if you're using the API, if you're using the online version, straight to the Chinese government? That's up to you. Maybe you don't care. But, yeah, I would definitely be, cautious, especially for, enterprises that aren't reading the fine print. Yeah. You should probably read the fine print on that one. All right.

Jordan Wilson [00:27:22]:
Our next piece of AI news, anthropic in even more drama. It's been the week of drama for Anthropic. After last week, we saw, you know, over the last two weeks, we saw a week of releases, right? We saw Claude for their Opus and their sonnets getting, Claude for pretty big bumps. And then we saw their, version of, connectors called integrations. But now it's just all the drama because, like, we just talked about, facing a pretty big lawsuit and consequential lawsuit from Reddit and now, some drama or a potential breakup with Wind Surf. So open, sorry. Anthropic has withdrawn nearly all access to its Claude three dot x model. So, you know, that's, three five and, three seven, models from Wind Surf after reports surfaced that OpenAI is acquiring Wind Surf for $3,000,000,000.

Jordan Wilson [00:28:20]:
So, Jared Kaplan, Infropix cofounder and chief science officer, confirmed the move was driven by a desire to avoid enabling a direct competitor, to focus and to focus on lasting partnerships, citing limited computing resources as a secondary factor. So Windsurf users. So if you don't know what Windsurf is, it's one of these kind of AI, vibe coding platforms. It's much more than that, but if you had to put into a category, it's a vibe coding platform, like Cursor. Right? So Windsurf users, including both their free and pro, customers lost direct cloud access with less than a week's notice. And access now requires users to bring their own API key while Gemini 2.5 pro is being offered at a discounted price as an alternative. So Windsurf, formerly Codium, has criticized the decision from Anthropic as anti industry and warned it could negatively impact many other companies reliant on AI model access. The timing suggests Anthropic wants to prevent its clawed models from supporting soon to be open AI owned rival and possibly to protect its proprietary, data from leaking into competitors' ecosystem.

Jordan Wilson [00:29:36]:
So a lot of people online are losing their noodles on this, and they're like, oh, this is such a terrible move from Anthropic. I don't know. My two senses, I don't blame Anthropic, for this. I think it was, in bad taste that they cut off, access to these models with, I think they said, five days notice. Right? We've seen reports on this OpenAI, acquisition on Wind Surf now for three weeks. So, yes, maybe, Anthropic needed to do its due diligence to independently, kind of confirm this rumored or reported on acquisition and to clear things internally. But still, to do this with only five days, even though I understand and agree ultimately with their decision, to do it with only five days notice is kind of bad form. Right? Especially given the fact that anthropic, I mean, they're really only future customer base, if I'm being honest, for the most part, is software developers, web developers, people doing agentic coding, and they're using tools like Windsurf and Cursor for that.

Jordan Wilson [00:30:45]:
So I I I know this angered, a big portion of their current customer base. So probably not a good move, at least on the optics from Anthropic, even though I agree with it. Y'all, Anthropics has, like a PR problem. Right? We talked about it last week, last week on the show, kind of with their, how they talked about, oh, we we we found all these problems, with Claude four. They're really bad, and then they deleted the tweet. Yeah. It it it's bad. Internal, I don't understand how a company as big as Infropic is essentially just abysmally bad.

Jordan Wilson [00:31:26]:
I don't even know if that's a word, but they're just very, very bad, at comms. Right? It's it's I don't I don't understand it because this is a simple business one zero one. Right? Don't don't piss off your biggest user base. And they did by doing this, even though I think ultimately they made the right decision, you probably should have given at least a couple of weeks of notice. Alright. Our next piece of AI news, speaking of kind of hot takes, I might have one on this one tomorrow. So a new Apple research paper called the illusion of thinking argues that AI reasoning models offer only marginal improvements only over standard language models and often fail as tasks grow complex, challenging a central, narrative in recent AI development. So according to this Apple study, standard large language models outperform reasoning models on simple tasks, while both types collapse on highly complex problems, with reasoning models regressing as complexity increases, contradicting claims that chains of reasoning yield smarter AI.

Jordan Wilson [00:32:37]:
So the study highlights a critical, critical vulnerability introducing small irrelevant changes to prompts can degrade model performance by up to 65%, revealing the model's reliance on pattern recognition rather than genuine logic or deductive reasoning. So Apple, in their study, found no evidence that current reasoning models perform true logical problem solving. Instead, they predict responses based on statistical patterns from their training data, casting doubt on the practical value of chain of thought outputs or these reasoning models. Critics have accused Apple of being shortsighted or self interested, especially as the company's own AI products like Apple Intelligence and Siri two point o face AI challenges, but Apple maintains its focus on privacy preserving efficient on device AI aligns better with real world use cases. And the paper's conclusion that large scale reasoning offers limited benefits also happens to coincide with Apple's public strategy of focusing on smaller efficient on device models, leading to accusations that the research is just marketing designed to justify their current position. The research exclusively uses abstract logic puzzles with a single correct answer as its benchmark. This is a terrible study, ignoring the primary real world applications of reasoning models, which often involve creative collaboration, coding, and drafting, where the step by step thinking process itself, is a valuable output. So let me just say it now.

Jordan Wilson [00:34:17]:
I'm gonna not sleep a lot today and tonight because I read this study over the weekend. And like, at first I'm scratching my head and I'm like, how is this come? Like like, who approved this? Right? This did not seem like a research led initiative. This seemed like someone high up in the business development side of Apple said, hey. FYI, June 9 today, right, today is our big WWDC announcement. And we are essentially right according to reports. Bloomberg said that Apple is taking a quote, unquote gap year, on AI. Whereas last year at their WWDC conference, they just said, hey. Everything.

Jordan Wilson [00:34:59]:
AI. AI. AI. Apple Intelligence. Apple Intelligence. Right? So reports are today at Apple's WWDC. They're they're gonna have some quote, unquote AI announcements, but they're essentially being like, oh, whoops. We couldn't deliver.

Jordan Wilson [00:35:10]:
We're facing multiple class action lawsuits that we promised all of this Apple intelligence and didn't deliver it. They had to pull their, you know, some of their simplest, AI features, because they were getting it wrong. Right? Like, email summaries were, hilariously wrong. So Apple clearly has an AI problem. So you have to question the validity of this research even though the research itself is valid, how they framed it, I think actually is extremely disingenuous to the research field. Right? I'm not a researcher by any means. I've obviously read a hundreds of research papers over the last, you know, three to five years as I become more interested and more involved in artificial intelligence. And out of all the ones I read, this is probably the most questionable research paper I've ever read.

Jordan Wilson [00:36:06]:
Like I said, the observations are sound, but this seemed like marketing. This seemed like cherry picking. This seemed like just something that, again, just seemed very disingenuous to just the AI research community. Right? This, this seemed like almost apple, you know, bending over backwards to try to cherry pick and frame, certain research data to justify how they're absolutely so bad at AI. It is unfathomable that a company like Apple many years later, after we saw reports they were spending millions of dollars a day internally on their own AI systems couldn't roll out anything that didn't work or didn't get them sued. So instead of facing the facts and doubling down and getting it right, now instead, they're taking a gap year, and they're putting out a suspiciously timed research paper that cast doubt on the future of large language models literally hours before their announcement where their future of large language models is being swept under the rug. Interesting. Right? Yeah.

Jordan Wilson [00:37:19]:
Joe here saying Apple marketing commissioned a study to prove that AI reasoning models are overhyped. Yeah. Little tongue in cheek, but I'm with you there, Joe. This doesn't seem like a research led, like, survey. Right? Like, even looking at the actual facts the researchers make. So the facts are sound. Right? But they're also illogical. Right? And the way that they're framed, I'm like, this seems like a company with a huge agenda that is trying to, shift markets because they know by putting this out at this time, it's getting a ton of news coverage.

Jordan Wilson [00:37:56]:
And news coverage, whether you believe it or not, it shifts markets. So, yeah. Actually, tune in tomorrow. I'm gonna commit to it right now. We're gonna do a hot take Tuesday on this paper. I know that sounds dorky, but I got takes. Alright. And our last piece of AI news, Meta is reportedly in talks for a $10,000,000,000 plus investment in Scale AI.

Jordan Wilson [00:38:19]:
So according to Bloomberg reports, in Reuters, Meta Platforms is exploring an investment in Scale AI that could exceed $10,000,000,000. So the deal, if finalized, would be one of the largest investments in an artificial intelligence startup ever and underscores the intensifying race among tech giants to secure AI capabilities. So Scale AI was founded in 2016 and was last valued near $14,000,000,000, specializes in data labeling, and already counts NVIDIA, Amazon, and Meta as backers. The company also operates a platform for AI researchers to share information with participation from con, from contributors across more than 9,000 different cities. So according to this new report on Meta's, purported 10,000,000,000 investments, terms of the potential deal are not yet finalized and could still change with Meta and Scale AI both declining to comment on the talks. If completed, the investment could further accelerate the advancements in AI technology and provide new opportunities for professionals and companies seeking to leverage large scale data infrastructure. It's a big move here. This is a big move.

Jordan Wilson [00:39:36]:
I think this is a potential deal that could catapult Meta into this one, kind of tier one discussion. Right? And and I don't know. Maybe it's because of, Meta takes is the only big tech trillionaire company to take this, primarily open source, open weights approach with their llama models. Although Google does probably have the most capable, open model, at least small open small language model in Gemma three n. But this is pretty big. I think this is something if it works out well and if Meta takes good advantage of this data, of this potential partnership, with Scale AI, it could make the race at the top much more interesting because at least since, quarter four of last year, so around November, December, it's really just been Google and OpenAI at the top. Right? You could say they're flip flopping in the one a spot, and then you have kind of, anthropic in Microsoft, in that one b spot. But you kinda have a big drop off, I would say, until you get to Meta and their llama models even after their most recent releases a couple of months ago.

Jordan Wilson [00:40:46]:
Alright. And, you all said you wanted this, this new little segment in here at the end of the AI news that matters are rumors and what's next. So like I said, what's very next? Well, in about, less than four hours, we're going to see Apple's WWDC, their worldwide developer conference, today at noon Central Standard Time, so noon Chicago time. But it's been reported they're essentially taking a gap year on AI. They're gonna be, opening up their three different, internal large language models to developers and some other small AI announcements. But they're essentially, you know, tucking their tail between their legs, and not saying AI every fifth word like they did last year because that blew up in their face. Still no grok 3.5, although the Twitter verse is saying it's gonna be dropping at any day soon, but, Elon Musk might be a little busy with his current breakup with president, president Donald Trump. G p t four o is showing some tracks of thinking for some users, which is interesting.

Jordan Wilson [00:41:50]:
A non reasoning model is showing traces of reasoning for many users sharing online. Google's new, I think I might've got that that, that one wrong there. I think it's called spark, sparkify. Yeah. Google's new sparkify, is an invite only creative machine that uses, Google's different, AI platforms to create videos up to two minutes long. You know, what's weird guys. I told you all that by the end of twenty twenty five, we were going to be able to create our own very high quality bespoke versions of like Pixar movies. And y'all laughed at me and I'm like, no, you guys, I talked to the people developing this technology.

Jordan Wilson [00:42:38]:
It's coming. Trust me. And a lot of you people probably thought I was crazy, but here we are. It's already starting to roll out where you can literally create up to a two minute video with a simple prompt, in this, Google Sparkify, which right now is invite only. Google v o3 has started rolling out, a fast version in the flow video editor. Google yeah. A lot of Google, what's next in rumors. Google is going to be rolling out scheduled actions in Gemini soon, which I'm extremely excited about that one.

Jordan Wilson [00:43:13]:
That's one of the things I love most, and I think one of Chesapeake Key's most underrated features is scheduled, tasks. So, now, looks like Google is testing that out in certain Gemini accounts. And NotebookLM in kind of an ask me anything session, they kind of laid out some of their road map, some things that are probably coming soon. Video over, video overviews, new source types, and even an API. A NotebookLM API is mind boggling what that could do for the industry. Alright. I hope this was helpful. So let me quickly recap the AI news that matters.

Jordan Wilson [00:43:52]:
So OpenAI has released an updated and more human like version of its advanced voice mode. Reddit is suing Anthropic for allegedly scraping user data. It didn't have a, a license to do. OpenAI has launched its new cloud connectors and ChatGPT meeting recorder. Although, I haven't seen the meeting recorder, pop up in my Mac app yet, the cloud connectors are there. Google has unveiled its upgraded Gemini 2.5 pro model even though it was already winning the race. New reports are saying that Chinese AI Lab DeepSeek is, using Gemini data to train its new powerful r one model. Anthropic has cut cut clawed access for Wind Surf amid OpenAI acquisition rumors.

Jordan Wilson [00:44:41]:
A new Apple study is challenging the hype around AI reasoning models, but I'm gonna break that down in tomorrow's hot take Tuesday. And last but not least, Meta is reportedly in talks for a $10,000,000,000 plus investment into Scale AI, which we want one of the largest investments in an AI company ever. Alright. I hope this was helpful. If so, please share this with your friends. Click repost if you're listening, here on Twitter or on LinkedIn. I'd appreciate it. If you're listening on the podcast, as always, please click that subscribe button, on Spotify or Apple Podcasts wherever you're listening.

Jordan Wilson [00:45:15]:
Yeah. Maybe you just listen on the live stream. This thing goes out of the podcast, FYI. So we'd love for you to follow us there. Leave us a rating if this is helpful. Please go to youreverydayAI.com. Sign up for the free daily newsletter. FYI, I'm I'm laying it out tomorrow.

Jordan Wilson [00:45:28]:
We're gonna do a hot take Tuesday on this new Apple paper and break it down a little bit for you. And then on Wednesday, the new AI at work Wednesday, you said you guys liked it. So we're gonna be doing, a segment called AI magic converts outdated content into engagement gold. You're not gonna wanna miss that one. I've been planning it for, like, three or four days. It's gonna be really good. You're gonna wanna join. Thank you for tuning in.

Jordan Wilson [00:45:53]:
Go to youreverydayai.com. Sign up for the free daily newsletter. I'll see you back tomorrow in everyday for more everyday AI. Thanks, y'all.