How Apple Could Get Their AI Revenge

0:00Today on the AI Daily Brief, how Apple could get its AI revenge before that in the headlines, friends, good AI gets much cheaper. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

0:18All right, friends, quick notes before we dive in. First of all, thank you to today's sponsors, Notion, Superintelligent, Robots and Pencils, and Blitzy. To get an ad-free version of the show, go to patreon.com slash aidailybrief. And for sponsorship opportunities, shoot us a note at sponsors at AIDailyBrief.ai. Welcome back to the AIDailyBrief Headlines Edition, all the daily AI news you need in around five minutes. We kick off with a new model out of XAI, which is called Grok4 Fast. According to XAI, the model is capable of delivering similar results to Grok4 while using 40 % fewer reasoning tokens.

0:52Combined with a big drop in cost per token, an XAI says this results in a 98 % reduction in price to achieve the same performance on Frontier benchmarks as GroK4. For example, on the AIME 2024 and 2025 math benchmarks, GroK4 Fast achieves very similar results to GroK4 while using around 60 % as many tokens. Meanwhile, the model's performance on the GPQA Diamond Scientific Knowledge benchmark was slightly worse than GroK4 but matched GPT-5 high. Artificial analysis ran their independent benchmarking and came up with equally impressive results. The model scored a 60 on their intelligence index, placing it in line with Gemini 2.5 Pro and Claude 4.1 Opus, but slightly below O3 or GPT-5 high.

1:30Alongside the efficiency characteristics, Grok4Fast also features a 2 million token context window, which is clearly aimed at things like large coding tasks. The model also has a unified architecture, allowing both reasoning and non-reasoning variants to operate based on the same model weights. The operation of reasoning is controlled via system prompt rather than being baked into the model. Overall, what's very clear is that this release is about pushing the cost-performance efficiency barrier even farther. I think it's fair to say that this is the narrowest gap we've seen between the frontier of performance and the fast and cheap model variant.

2:01Artificial analysis remarked, stepping back, this release follows the trend of the cost of accessing AI intelligence falling quickly. In the past one and a half years, the cost of accessing GPT-4 level intelligence has fallen around 500 times and falls have continued as intelligence frontiers have been reached. Professor Ethan Malek writes, with the new Grok4 fast, the price performance curve for AI has shifted again. I updated my chart to reflect this change. I have to update it every couple weeks now. I also think the GPQA diamond benchmark is likely maxed out. The tests themselves have errors, making it impossible to get to 100%.

2:34I'm going to need to start doing this with a harder benchmark. The chart, for those of you watching, have the benchmark score over on the y-axis and the cost per million tokens descending on the x-axis, meaning that up and to the right means higher performance for cheaper. And you can see Grok4 fast out way up here on the top right of that curve. We've talked a lot recently about why for the moment, most organizations are still choosing to use the state-of-the-art model, no matter what the cost implications, but that I think that part of that reflects the fact that each new model advance is still really unlocking new use cases and that as more models are over the threshold into high performance for a lot of use cases, those choices might start to shift.

3:13Basically, if you have the option to use Grok4 fast, which involves so little compromise on performance for so much gain in terms of cost, You're likely going to do that even if you were not willing to make the trade-off between Gemini 2.5 Pro and Gemini 2.5 Flash just six months ago. One other addendum on XAI news. According to sources familiar with the deal, XAI has raised$10 billion in debt and equity from investors including Valor Capital, the Qatar Investment Authority, and the Saudi investment firm Kingdom Holding Company at a$200 billion valuation. CNBC reports that this is additional to the$10 billion that was raised a few weeks ago.

3:47Seems like we're unlikely to get any official clarity on the deal, as on Friday when the news broke, Elon Musk posted fake news. XAI is not raising any capital right now. The secrecy around the fundraising is noteworthy considering a string of reporting of unrest at XAI. Last week, the Wall Street Journal claimed that a string of executive departures, including the CFO, were related to concerns about the company's management and financial health. That article claimed that Valor Capital Antonio Garcias had played a hands-on role in mediating disputes with executives, a claim which XAI has denied.

4:16On the brighter side, the New York Times reports that Rock has reached 64 million monthly users, according to Musk. The claim was reportedly made in an all-hands on Thursday. Now, moving back to benchmarks and their saturation, this is obviously a big concern, and not just for Ethan Mollick to be able to show off the AI price performance curve. The more saturated the benchmarks are, the harder it is to really understand new model performance gains. To help with that, Scale AI has introduced a new coding benchmark. At this stage, Frontier AI models are all clustered around 70 to 80 % on SweetBench verified, and users don't have a lot of confidence that a one or two percentage point difference is all that meaningful.

4:50To solve this problem, Scale has introduced a new benchmark called Sweebench Pro. The new test will source problems from commercial, proprietary, and copy-left-style open-source codebases to reduce the chances that problems are contained in training data. It's also designed to more closely reflect production environments and real-world problems. Bing Liu, the director of research at Scale, wrote, Current benchmarks like Sweebench have driven real progress, but they no longer reflect the Frontier challenges faced in production systems. Alongside contamination resistance measures, Bing noted that the benchmark now includes more difficult tasks like changing 100-plus lines at a time and working across large codebases.

5:23Testing the current Frontier models, Scale found that GPT-5 had the best performance, with 23.26%, while Claude 4 Opus was at 22.71%. The next pair were Claude 4 Sonnet at around 17 % and Gemini 2.5 Pro at around 13%. If nothing else, the spread of scores demonstrates that new benchmarks have the potential to more easily differentiate between leading models. On top of that, the generally low scores show that there's a lot of room to measure improvement before the benchmark becomes saturated. Lou commented that the scores were even lower on the problems based on commercial codebases, adding that this highlights, quote, how much harder working with real enterprise codebases remains.

5:58I think that this shift not only to find new benchmarks, but also to have them more directly related to real-world performance and the environments that AI is actually going to operate in is a great shift. Moving over to the rumor mill section of our episode, OpenAI appears to have some interesting new releases coming soon. On Sunday, Sam Altman posted on X, over the next few weeks, we are launching some new compute intensive offerings. Because of the associated costs, some features will initially only be available to pro subscribers and some new products will have additional fees. Our intention remains to drive the cost of intelligence down as aggressively as we can and make our services widely available, and we are confident we will get there over time.

6:34But we also want to learn what's possible when we throw a lot of compute at interesting new ideas. Resource constraints were one of the big themes around the release of GPT-5. It was the first time that OpenAI's flagship model didn't increase in size from the previous iteration. According to Semi-Analysis, OpenAI added more compute this quarter than in any previous quarter, so they might have a little space capacity to play around with, although given that Sam is already preparing us to pay more for whatever it is they're releasing, it seems like it's going to be pretty compute-hungry. The speculation, of course, ran rampant on X.

7:04Vraser wrote, Sam just teased new compute-heavy drops. This could be Sora 2 or the reasoning beast that won golden math and code. Maybe even native video generation edit every frame, every angle, every detail like Photoshop for reality. If that's true, we're weeks away from collapsing the gap between imagination and creation. I would say that from my exploration, Sora 2 certainly seemed to be the odds-on favorite guess. Others, though, think it will be that experimental reasoning model that solved the final problem in the ICPC coding competition last week. Some also recalled Noam Brown's recent comments about a civilization of geniuses in the data center.

7:37Prins commented, Reminder that Noam Brown's multi-agent team has been working on significantly increasing the time horizon over which a model can reason. Meanwhile, the team at OpenAI has been pretty tight-lipped, with a researcher named Tristan kind of tamping down speculation posting new features, improvements. The next few weeks are going to be fun. Meanwhile, OpenAI's inference bill just keeps going up, with a company planning to spend an extra$100 billion on backup servers over the next five years. The information reports that investors were told that backups would be monetizable by either powering research breakthroughs or servicing surges in product usage.

8:08That amount is in addition to the$350 billion that OpenAI had already projected to spend on server rentals through 2030. OpenAI is now expecting to spend an average of$85 billion a year on server rentals over the next five years. Meaning that even if this year's projected$20 billion in revenue is achieved and rapid growth continues, they will still have a large shortfall that will need to be made up with regular fundraising. CFO Sarah Fryer told investors that the company is quote massively compute constrained and said the plan was necessary to ensure OpenAI doesn't need to hold up features or new models.

8:37Information executive editor Amir Afradi commented, the implication of OpenAI's plan to rent$450 billion worth of servers before the end of this decade are mind-blowing. Certainly the new projection backs up the idea that OpenAI might actually need the$300 billion contract they recently signed with Oracle that sent that company's stock soaring. Speaking of Oracle, last one today. Oracle is in talks with Meta for a cloud computing deal worth$20 billion. Tiny, tiny numbers after what we were just talking about. Bloomberg reports that Meta is looking to sign a multi-year deal to purchase compute for training and inference.

9:07The deal would diversify Oracle's revenue streams, which are currently heavily concentrated with OpenAI. It does, however, raise questions on whether Oracle and their infrastructure partners will be able to build the multiple gigawatt-scale data centers required to fulfill their contracts. On Meta's side of the equation, it's very clear that CEO Mark Zuckerberg is willing to spend whatever it takes on AI infrastructure. In an interview last week, he said, we're going to spend aggressively. Even if we lose a couple hundred billion, it would suck, but it's better than being behind in the race for superintelligence.

9:34Amit has investing commented, to even hear Zuck say this, it shows how important they believe the opportunity is. CapEx is not slowing down. That's going to do it for today's headlines. Next up, the main episode. Chatbots are great, but they can only take you so far. I've recently been testing Notion's new AI agents, and they are a very different type of experience. These are agents that actually complete entire workflows for you in your style. And best of all, they work in a channel that you already know and love because they are purpose-built Notion super users. Notion's new AI agents completely expands the range of what Notion can do.

10:10It can now build documents from your entire company's knowledge base, organize scattered information into organized reports, basically do tasks that used to take days, and get them complete in minutes. These agents don't just help with work, they finish it. Getting started with building on Notion is easier than ever. Notion agents are now your very own super user to help you onboard in minutes. Your AI teammates are ready to work. Try Notion AI for free at the link in our show notes. Today's episode is brought to you by Superintelligent. Now, one thing that we are having a lot of conversations with folks about is the fact that for some of you, your fiscal year is coming to an end.

10:43And that means two things. One, it means planning and thinking about what you're going to do in the next year. And two, it means using up those last of budgets so you don't lose them. If you are an enterprise that happens to find yourself in that situation, Superintelligent would love to help on both fronts. We are moving increasingly towards an annual AI planning model where we map out how you can create an action map of your organization's agent opportunities that represents an executable backlog of AI and agent use cases that you can deliver on over the course of the next year. Additionally, for those end-of-year budgets, we have worked out deals with a number of partners where we can pre-lock in general implementation packages even before you figured out exactly what use cases are going to require them.

11:24If you'd like to learn more about superintelligence agent readiness audits and this new end-of-fiscal-year plan, visit us at bsuper.ai, click Get Started, and make sure to use the word fiscal somewhere in the description. Today's episode is brought to you by Robots & Pencils. When competitive advantage lasts mere moments, speed to value wins the AI race. While big consultancies bury progress under layers of process, Robots & Pencils builds impact at AI speed. They partner with clients to enhance human potential through AI, modernizing apps, strengthening data pipelines, and accelerating cloud transformation.

11:57With AWS-certified teams across U.S., Canada, Europe, and Latin America, clients get local expertise and global scale. And with a laser focus on real outcomes, their solutions help organizers work smarter and serve customers better. They're your nimble, high-service alternative to big integrators. Turn your AI vision into value fast. Stay ahead with a partner built for progress. Partner with Robots & Pencils at robotsandpencils.com slash AI Daily Brief. This episode is brought to you by Blitzy, the enterprise autonomous software development platform with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise-scale code bases with millions of lines of code.

12:37Enterprise engineering leaders start every development sprint with the Blitzy platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and precompiles code for each task. Blitzy delivers 80 % plus of the development work autonomously while providing a guide for the final 20 % of human development work required to complete the sprint. Public companies are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding co-pilot of choice to bring an AI-native STLC into their org.

13:05Blitzy is providing a limited-time, 30-day free proof-of-concept for qualifying enterprises. The team will provide a 5x velocity increase on a real development project in your org. Visit blitzy.com and press Book Demo to learn how Blitzy transforms your STLC from AI-assisted to AI-native. That's B-L-I-T-Z-Y dot com. Welcome back to the AI Daily Brief. Every few months, we get renewed chatter about AI-native devices, AI wearables. Sometimes it's because of new product launches. Other times it's because of rumors. But it's this fascinating discourse where people are debating what a category will look like that, frankly, isn't even really a category yet.

13:44It is just something that some companies are betting on. Today, we're going to look at the latest news, a bunch of which has come in the last week or so, get a sense of the state of play, and talk about why after all of their flops, failures, and frankly, just unimpressiveness in AI, Apple could secretly be sitting on the AI device Trojan horse. Let's start with the latest news from over the weekend, though. The short of it is the latest burst of reporting around the much-ballyhooed and anticipated OpenAI device that represents a collaboration between Sam Altman and Johnny Ive. Now, it's been pretty clear that OpenAI and Altman are trying to capture some of the magic of the iMac, iPod, and iPhone-era Apple team.

14:23Obviously, the hiring of the designer of many of those products, Johnny Ive, is integral to that. But the information also reports that the company has been hiring Apple engineers and contracting with iPhone manufacturing partners over recent months. The information writes that Luxshare, which is a major assembler of both iPhones and AirPods in China, has a contract to assemble at least one of OpenAI's devices. And sources say that OpenAI has also approached Gore-Tec, which assembles AirPods, HomePods, and Apple Watches. Sources say that the first release target is late next year or early 2027, and the reporting also included some confirmations and updates on the form factor the device would take.

14:57The information writes, one of the products OpenAI has talked to suppliers about making resembles a smart speaker without a display. Another source said that the company has also considered building glasses, a digital voice recorder, and a wearable pin. Now the first part, the idea of it resembling a smart speaker without a display, seems in line with rumors from earlier this year, that the device would be something like a pocket-sized puck. Back in May, the Wall Street Journal reported that the idea was to have a third core device a person would put on a desk after a laptop and an iPhone. They wrote that the product would be capable of being fully aware of a user's surroundings in life and would be unobtrusive.

15:33On the flip side, the reporting that OpenAI has considered a wearable pin is very much not in line with previous comments, where Johnny Ive has been quite negative about previous devices like the Rabbit and the Humane pin, saying, quote, there has been an absence of new ways of thinking expressed in products. Sam Altman also dashed the rumors that the first device would be smart glasses, perhaps recognizing that Meta has a big lead in that area. But all of this does line up with the reporting that they were considering not just a single device, but a family of devices. It may simply be that they're at the stage where they're just considering their options and understanding if then sort of implications of if they were to develop a particular type of device, what the constraints would be.

16:10Overall, in 2025, OpenAI has now recorded more than two dozen employees from Apple who previously worked on consumer hardware. That's up from about 10 last year and none in 2023. It's not all that surprising that they're able to be successful with this poaching. The information writes, Some longtime Apple employees working on the company's hardware products have become bored with the incremental changes in the type of products they're working on and frustrated with the bureaucracy at Apple. It hasn't helped that employees have seen their income suffer due to Apple's lackluster stock over the past year.

16:40Unsurprisingly, they also write, IVE's involvement with OpenAI, too, has enhanced the company's credibility in the eyes of its recruits from Apple. So the takeaways for those keeping track home, one, OpenAI recruitment of Apple hardware folks is increasing, planning for the supply chain for this device or devices is happening. From a timeline perspective, we're looking at somewhere between 12 and 18 months, call it, and the form factor remains up in the air. Now, presumably, one of the reasons that we're hearing more about OpenAI's device right now is the buzz around Meta's next generation of smart glasses that were announced last week.

17:10So far, the tech press has seemed to have fairly universally positive experiences. The glasses now feature a tiny built-in screen that's invisible to the outside world, and Meta also seems to have figured out how to do gesture controls with a haptic wristband. Lance Ulanoff of TechRadar commented that the new version of the Ray-Bans feel like they succeeded in every way that Google Glasses failed. They look less conspicuous, they're more comfortable to wear, and they have a much more significant battery life. The device is also not months and months out. They will, according to Mark Zuckerberg, be available within a few weeks, and at a price point of$799, they're expensive but nowhere near the, for example,$3 ,500 for the Apple Vision Pro.

17:50The fact that these Ray-Bans are succeeding in the ways that Google Glass failed is obviously very intentional. It's quite clear that Zuckerberg is making his big bet over the form factor for AI devices. At last week's announcement, he said, glasses are the ideal form factor for personal superintelligence because they let you stay present in the moment while getting access to all of these AI capabilities to make you smarter, help you communicate better, improve your memory, improve your senses. And this is the sort of grandiose language that the people behind these devices are using. Sam Altman has made similar proclamations about his device.

18:24During the swirling news cycle following their announcement back in May, he said, just the way we think of our current computers were designed for a world without AI. And now we're in a different world, and what you want out of hardware and software is changing quite rapidly. And if Altman's device really is some version of this little puck that you either keep in your pocket or wear as a necklace or something else, there is an interesting difference in the bet about what the future of AI devices is built for. Altman, from the little we know, appears to be betting on a vision of ambient AI. In other words, a series of devices all talking to each other and maintaining context that can create the feeling of an omnipresent AI assistant.

19:01Zuckerberg's bet, meanwhile, is still very much rooted in the idea of a device acting as a portal to use AI. The glasses can serve as a comfortable way to access your AI, but they're still fundamentally a device that you turn on and use and control, rather than an AI enhancement to your surroundings. So far, the ambient AI vision has been a little stuck on go. One of the more prominent devices to get previewed in recent months is a product called Friend. The device is a pendant offering that had some similar features to earlier AI wearables. It records conversations, can help users keep track of their day and keep a diary of events, and provide observations throughout or once the day is done.

19:37A series of fairly dismal reviews were published earlier this month, epitomized by the Wired headline, I hate my AI friend. Now, what was interesting about this is that whereas with previous generations of AI wearables, the big complaint was that the devices fundamentally didn't work and really didn't do anything, this time around, the complaints were around, one, how people reacted to the pin and the general hostility that they found, and two, the fact that they simply didn't like its personality. In fact, engineer Ali Bell commented at the time, extraordinary leap forward that we can critique a wearable's personality and not just its usable hardware.

20:12Still, there are some early positive reviews of the friend as well. Robert Scoble writes, now that I have used a few of the always listening devices like Avi Schiffman's friend, It makes me want to display in a camera so I can show the AI things, a display so it can show me things, but they are cool. They just leave me wanting a lot more, like a virtual being on my screen. Now, addressing this multimodal capability is obviously a clear goal for Meta. For example, during last week's event, Mark Zuckerberg said, glasses are the only form factor where you can let AI see what you see, hear what you hear, talk to you throughout the day, and very soon generate whatever AI you need right in your vision in real time.

20:47The question is, if the use cases that people are looking for involve explicitly deciding to call up and use the AI rather than it being on all the time, is the perfect AI device already sitting in every pocket in the world? Right now, a huge percentage of people access AI using their smartphone, and it's not necessarily all that inconvenient. Now, some of the knocks on a smartphone are that they don't offer enough bandwidth. In other words, typing and reading responses is too slow, although presumably that could be solved with better voice mode. For Zuckerberg and Altman, another issue with the phone, though, is a matter of screen use and the way that it potentially disconnects people.

21:20A big philosophy that both of those two seem to share these days is about getting people to look up from their phone and connect with others around them. Certainly so far, it looks like Google's big AI device bet is the smartphone. The recent release of the Pixel 10 was all about AI, even though some of the AI features weren't labeled as such. The chipset for the phone was specifically chosen to give better AI performance at the cost of a better CPU. And while this might not be the generation to do it, you get the sense that Google wants the smartphone to become an AI-enhanced experience in every way.

21:48What's more, now that people have gotten their hands on the iPhone 17, it seems like Apple is making a similar bet. Apple lacks the advanced software models that Google has, but the hardware is no joke. Match.com developer Adrian Gronlund posted a video of what he's been tinkering with, writing, here's Apple's foundation model running on iPhone 17 Pro. It's just so fast. Apple was not joking. The A19 Pro chip is really good for running LLMs on device. And yet, if one was looking at the recent Apple event, for their real Trojan horse for AI devices, I think that it was not in fact the iPhone, but instead another device that got a new version announced.