The AI Daily Brief: Artificial Intelligence News and Analysis - Everything We Know About GPT-5 So Far

Starting point is 00:00:00 Today on the AI Daily Brief, everything we know so far about GBT5. Before then in the headlines, vibe coding is finally coming to the enterprise. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. All right, friends, quick announcements before we dive in. One, thank you to today's sponsors, Blitzy, Plum, and Vanta. To get an ad-free version of the show, go to patreon.com slash AI Daily Brief. Second, if you were interested in sponsoring the show, we are starting to fill up for our fall and winter slots, Shoot me a note at NLW at Breakdown.net.

Starting point is 00:00:35 And I can send you more information. Lastly, I mentioned this yesterday. But if you are a dev shop or a build agency that actually builds agents, I want to hear from you. Super Intelligent is, of course, routing people to the use cases that are most relevant for their companies. But then they've got to build things. And if you have the capabilities to build things, especially at meaningful enterprise

Starting point is 00:00:56 scale, shoot me a note at NLW at BSUPER.AI with agent builder in the subject line, because I want to hear from you. But with that, let's get into today's episode. Welcome back to the AI Daily Brief Headlines edition, all the daily AI news you need in around five minutes. We kick off today with something that was on the one hand, completely inevitable, but on the other hand is still extremely exciting, which is the fact that vibe coding is finally coming to the enterprise in a major way.

Starting point is 00:01:22 Replit and Microsoft have announced a strategic partnership that will see Replit added to Microsoft's Enterprise Cloud Store. The platform will also be integrated into Microsoft's cloud services like containers, virtual machines, and their version of Postgres. This integration will allow app builders to have easy access to enterprise-grade backend infrastructure while allowing Microsoft to earn revenue from the developer ecosystem. Now, while Microsoft already offers their own AI coding platform in the form of GitHub copilot, they say that Replit will be a complementary addition to their product lineup. Instead of a replacement for co-pilot, they're pitching Replit as a substitute for no-code

Starting point is 00:01:55 prototyping and design tools like Figma. The pitch is that non-technical business managers can use Replit to build their own apps. Deb Kupp, the president of Microsoft America, said, at Microsoft, we believe every person in every organization should be empowered to achieve more through technology. Our collaboration with Replit democratizes application development, enabling business teams across enterprises to innovate and solve problems without traditional technical barriers. Repplet CEO, I'm John Mossad added, we aspire for Replit to be the most trusted name for enterprise in this new era of agenetic coding. Now, TechCrunch noted that there is one big loser from the new partnership writing, if there is any competitor taking an L from this

Starting point is 00:02:30 partnership, it's Google Cloud. The apps built and run through Replett are typically hosted on Google Cloud. In fact, Replit has been such a feather in Google's cap that the cloud giant has profiled the partnership. However, the deal is non-exclusive, meaning that the startup is not leaving Google Cloud, but is growing to support Microsoft shops. Still to me, from where I'm sitting, this is a completely obvious type of tie-up, and I agree wholeheartedly with Carthick Harry Horan, who writes, I predict Microsoft will acquire a company in the AI-Vy coding space within the next six to 12 months, just like their acquisition of GitHub in 2018. Bursale, lovable, replet, et cetera, will all be acquisition targets.

Starting point is 00:03:04 To me, this is just incredibly obvious. We are already seeing, certainly in the startup domain, but also creeping into the enterprise how these vibe coding tools are changing how people do their work, especially in non-technical roles. But enterprise vibe coding really is a different, more complicated beast than consumer-level vibe coding. It's going to take someone like a Microsoft who has deep pockets and patience to create a context where it's worth pursuing.

Starting point is 00:03:25 But boy, is the pot of gold at the end of that rainbow, a big one. Next up today, some funding news. Mistral is in talks to raise a billion dollar funding round as they grow into a regional champion in Europe. The French AI company is reportedly in talks to raise a billion dollars in equity from several investors, including Abu Dhabi Fund, MGX. Alongside, the firm is seeking hundreds of millions of euros in debt capital from French lenders, including BPI France. To date, Mistral has raised around a billion euros since its 2020-Founding, reaching a valuation of just under $6 billion after a round last year. This round then would be a significant jump in available capital, enabling the firm to pursue European Data Center projects and compete in the next generation of foundation models.

Starting point is 00:04:03 The fundraising also has a distinctly geopolitical angle. French President Emmanuel Macron has said that Mistral is central to the idea of European AI sovereignty. This deal would deepen ties between France and the UAE, ensuring some level of independence from China and the US in AI competition. MGX already partnered with Mistral and Nvidia to construct Europe's largest data center campus in May. The 8.9 billion euro facility will be built in France and is projected to be operational by 2028. Beyond that, the UAE has also committed to spend 50 billion euros on AI projects in France

Starting point is 00:04:33 in support of Macron's push for AI sovereignty. In other funding news, Agent Tooling Platform Langchain is about to become a unicorn. According to TechCrunch sources, the company is set to raise a new round of funding at a billion dollar valuation led by IVP. Langchain started life in late 2022 as an open source project self-funded by founder Harrison Chase. In early 2020, it's developed Interest grew, Chase transformed the project into a startup. He closed a $10 million seed round from benchmark that April, and a week later, he raised another $25 million in a Series A led by Sequoia, which valued the startup at $200 million. Langchain was a breakout hit early during this AI era. It provided agentic tooling for LLMs before those terms were well-defined. The platform was

Starting point is 00:05:11 one of the first to allow LLMs to do things like search the web, call APIs, or interact with databases, enabling developers to build AI apps even in the early days. Since then, they've added observability, evals, and monitoring through their closed-source-Lang Smith platform. Now, TechCrunch noted just how many unicorns are emerging from the ecosystem of AI startups. According to them, there were 36 AI unicorns minted in the first half of this year, and Y Combinator has said that they believe that 300 unicorns were created across the entire SaaS boom, a figure that looks like it will be exceeded by AI by the end of the year. One more investment to profile today, although of a very different sort, meta is making a

Starting point is 00:05:45 multi-billion dollar bet on AI wearables. Zuck Shop has bought a minority stake in Rayband Maker SLOLuxatica. Sources say they purchased around 3% of the company for $3.5 billion. Those sources also say that meta is considering adding to that investment, bringing their stake to 5% over time. Now, the investment certainly cements the idea that smart glasses are a key meta device play for the AI era. They've been offering the meta-ray bands for almost four years and recently launched a pair

Starting point is 00:06:09 of Oakley-branded smart glasses, which is another make from the same parent company. SLOLLexatica is the largest eyewear company in the world, so the partnership gives meta access to industry-leading manufacturing and distribution networks. Now, leaning this hard into AI devices is a reversal from how META addressed the smartphone era. Jamath Palahapatia, who served as a senior executive back then, has said he pushed the company to develop their own phone in 2007 following the release of the iPhone. That project was never completed and Facebook was forced to develop on rival hardware for the entire tech cycle. Zuckerberg later said that this was one of his biggest regrets, meaning that

Starting point is 00:06:40 his company didn't get to shape the way mobile platforms developed. He is clearly not looking to make the same mistake twice, with meta entering the AI era with the most established platform launched well ahead of its relevance. Then again, the tech is quickly catching up to make smart glasses a functional and ubiquitous AI platform. The International Data Corporation expects sales of smart glasses to grow by 47% each year through 20209. So clearly with this deal, meta is looking to lock in their early dominance of the category. Now, I'm not sure that I think that this will be meta's only play in the AI hardware space, but it's certainly an unexpected beachhead that has already paid some amount of dividends for them. With that,

Starting point is 00:07:14 though, we will wrap the headlines. Next up, the main episode. This episode is brought to you by Blitzy. If you're a technology leader, here's something that probably sounds familiar. Your organization's competitive edge is buried in legacy code that desperately needs modernization, but the resources required feel out of reach. That was the case for a global investment analysis firm. They needed to migrate 70,000 lines of complex MATLAB financial algorithms to Python, algorithms that drive investment decisions for trillions in assets.

Starting point is 00:07:42 Their estimate? Months of high-cost specialized engineering work. Instead, they partnered with Blitzy. Blitzy's autonomous AI preserved mathematical precision and generated over 80% of the new codebase, completing the migration with just five days of engineering time. They cut the timeline by 95% and saved 880 engineering hours. If your organization is facing similar modernization challenges, visit blitzy.com to schedule a consultation and discover how AI power development can transform your technical capabilities.

Starting point is 00:08:11 Today's episode is brought to you by Plum. You put in the hours, testing the prompts, refining JSON, and wrangling nodes on the canvas. Now it's time to get paid for it. Plum is the only platform designed for technical creators who want to productize their AI workflows. With Plum, you can build, share, and monetize your flows without giving away your prompts or configuration. When you're ready to make improvements, you can push updates to your subscribers with a single click. Launch your first paid workflow at useplum.com. That's Plum with a B and start scaling your impact.

Starting point is 00:08:44 Today's episode is brought to you by Vanta. In today's business landscape, businesses can't just claim security, they have to prove it. Achieving compliance with a framework like SOC2, ISO-27-01, HIPAA, GDPR, and more, is how businesses can demonstrate strong security practices. The problem is that navigating security and compliance is time-consuming and complicated. It can take months of work and use up valuable time and resources. Vanta makes it easy and faster by automating compliance across 35-plus frameworks. It gets you audit-ready in weeks instead of months and saves you up to 85% of associated.

Starting point is 00:09:16 costs. In fact, a recent IDC White Paper found that Vanta customers achieved $535,000 per year in benefits, and the platform pays for itself in just three months. The proof is in the numbers. More than 10,000 global companies trust Vanta. For a limited time, listeners get $1,000 off at Vanta.com slash NLW. That's VANTA.com for $1,000 off. We'll go back to the AI Daily Brief. Today we are talking about one of potentially the next big things to hit AI, which is, of course, GPT-5. There is a growing sense that this model is right around the corner and that it may be very significant. So today we're going to talk about everything that we know about it, how it might impact the AI space, and yes, a little T-leave

Starting point is 00:10:04 speculation around when we might actually get the thing. Now, GPT-5 has been coming soon for almost a year at this point, but the model that finally gets released will be very different from where it started. In the middle of last year, rumors started circulating about the next big models from OpenAI, codenamed Orion. Rumors in the fall were that the model was going to come as early as December, but it wasn't too long after that in around November that we started hearing about industry-wide issues around training with a building narrative that pre-training it hit a wall. What we ultimately got in the place of a new flagship model in the series of GPT3, 3.5, and 4,

Starting point is 00:10:40 was instead OpenAI's first reasoning model, first 01, and then 03. Now, as we've previously discussed, the launch of reasoning models, it turns out, were a huge inflection point in the adoption of AI. A whole array of new use cases came online, enterprise adoption started going up significantly, and the new era of agentics became a viable possibility rather than just something for the future. Now, in and around the reasoning models, we did get a GPT 4.5. It clearly wasn't a big enough leap to warrant the title. And while the model was a niche hit among people who wanted a better LLM for writing,

Starting point is 00:11:13 it failed to capture much attention or usage in the broader public. In fact, the model is actually being sunset in the API next week. Swicks from latent space had this interesting observation, saying, I think the quote-unquote failure of GPT 4.5 relative to 03 and 04 mini is actually fantastic validation of the bet on the reasoning paradigm. 10 to 100x larger model consistently lost out to smaller reasoning model for an important open domain task. And still, hold aside what's great about the reasoning models.

Starting point is 00:11:41 The point is that there hasn't been a new, Open AI flagship in the series of GPTs for about a year now, and it's been over two years since they felt confident enough to herald a new generation of models with a full numerical upgrade from GPT4. At the same time, the industry has changed quite a bit since the last time we got excited about GPT5. Last fall, as we just discussed, the narrative was that scaling hit a wall, and people were really wondering if there was any more juice left to squeeze when it came to pre-training or if we were just in the era of new methodologies like test time compute, the reasoning era and agentics with tool use being the vectors for improvement going forward. At this point, now,

Starting point is 00:12:17 six to nine months later, there is a pervasive sense of optimism that AI is nowhere near at Zenith and that there are still plenty of avenues to pursue in order to improve the technology. Logan Kilpatrick, the head of product at Google AI Studio recently posted, the next six months of AI are likely to be the most wild we have seen so far. Everything keeps scaling up. More hardware, more model progress, more product knowledge, more AI momentum, more product market fit. Now, Logan went to pains to say that this wasn't inspired by any specific thing. He said it was just inspired by me feeling the progress and then looking out over the next six months and being reminded it's going to continue. So bringing it back to GPT5, what part of that

Starting point is 00:12:53 excitement can we attribute to this new model? The main expectation around GPT5 is that it will represent a unification of OpenAI's technology. In a recent podcast, OpenAI's head of developer experience, Romaine Hewitt said, we're truly excited not to just making that new great frontier model, we're also going to unify our two series. The breakthrough of reasoning in the O series and the breakthroughs in multimodality in the GPT series will be unified, and that will be GPT5. One of the core promises is that GPT5 will do away with model switching. In a Reddit AMA from May, OpenAI VP, Jerry Twork wrote, GPT5 is our next foundation model that is meant to just make everything our models can currently do better and with less model switching. Interestingly, he also referred to their operator

Starting point is 00:13:34 agent as a, quote, product surface, suggesting perhaps that GPT5 will feature tighter integration with agendic tools. Back in February, Sam Altman had said something similar about the model switching idea. In a post on X, he wrote, we want AI to just work for you. We realize how complicated our model and product offerings have gotten. We hate the model picker as much as you do, and want to return to Magic Unified Intelligence. Now, this post was a couple weeks before GPT5, and he said, we will next ship GPT4.5, the model we called Orion internally as our last non-chain-of-thought model. After that, a top goal for us is to unify O-Series models and GPT series models by creating systems that can use all of our tools know when to think for a long time or not and generally be useful

Starting point is 00:14:15 for a wide range of tasks. Now, getting a little bit more up to date, developer Bueh-Denock compiled some other nuggets of information gleaned from recent interviews and leaks. He expects a 256K context window, which would put GPT-5 roughly in line with most competitors, but not as large as Google's million token context window. Multimodality likely means will see native video, image and audio inputs, and perhaps even outputs. Another current rumor is that OpenAI will adopt the mixture of experts architecture brought to prominence recently by several Chinese labs. This architecture means that only part of the model is engaged

Starting point is 00:14:46 at a time, allowing for a higher number of total parameters while keeping inference costs down. There are estimates that inference costs could be 60% lower per token than GPT-40. Memory is another factor that could see improvement which would lead to more perform an agent operation as well. Not commented, for builders, this likely means you have to rethink prompt design for giant context, expect richer tool calling that mixes text with time-based media, and budget for lower latency cheaper API calls despite a larger model. GBT5 looks less like a parameter bump and more like a systems integration milestone, folding multiple specialized capabilities into one cohesive model. And frankly, even without the rumors, that seems like a reasonable

Starting point is 00:15:24 assumption. Since its release, OpenAI has been tacking features on to GPT4, including things like memory and updated image generation. GPD5 represents their first opportunity to bake these features in natively, allowing the model to be trained from the ground up on how to make best use of the tools it has access to. Rassarax commented, GPD5 might be the first model that feels like true AGI. If OpenAI integrates full O4 or O4 Pro reasoning plus agentic tool use within the chain of thought, operator, codex, and deep research, we're talking about a model that can think, plan, act, and adapt like never before. AGI vibes incoming. Now, one of the reasons we think we think that GPT5 is nearing release is that users are starting to report seeing a ton of AB testing

Starting point is 00:16:04 of something new on the platform. Specifically, it appears that OpenAI is testing how reasoning traces are presented. Onex user, for example, suggested that a recent interaction they had looked like hybrid reasoning. In response to a prompt using the 4-0 model, which again is not a natively reasoning model, chat GPT said, just a moment, I want to give this one the extra thought it deserves. That certainly would imply that OpenAI has figured out a way to differentiate between the prompts that should use reasoning and those that shouldn't. Another interesting feature on that post is a button labeled Answer Now that presumably cuts the reasoning short and forces the model to just spit out an answer based on how much thinking it's done at that point. One of the problems with hybrid reasoning

Starting point is 00:16:41 is that if the model starts thinking and goes down some rabbit hole, there hasn't historically been a good way to make it stop. So again, while it's just a guess, it seems like Answer Now might be a Ux change to address that issue. So if the idea is that GPT5 will natively bring together all of these features that have been bolted on, perhaps we also need a slightly different way of thinking about what an LLM actually is. Karina Nguyen, a researcher and product staffer at OpenAI, recently posted super intelligent operating system. Believe it or not, she is not talking about my startup.

Starting point is 00:17:12 Instead, cryptically, it sounds like it's sort of what's being described in the rumors of GPD5. Rather than treating it like an individual model you interact with, GPD5 kind of sounds like an AI operating system. T.J. Ridgeway writes, so GPD5 is the AGI framework. This is why they're all pivoting to superintelligence discussion, the one ring to rule them all. In response to the idea that superintelligence won't just be a scaled LLM, he added, yes, I pretty much agree.

Starting point is 00:17:38 What I'm saying here is this roadmap they put out is the framework through which AGI will be achieved, not AGI itself. I believe innovations in long-term memory are also a crucial aspect that is just now being explored. It is worth noting that back in January, Sam Almond did mention, quote, we're now confident we know how to build AGI as we have traditionally understood it. Still a practical question after all of this is whether any of it will be enough to amaze heavy AI users. Specifically, as much as we're summing up all of these different rumors here, none of them really suggest at least not yet a massive change in capability or anything particularly new. Instead, these rumors all focus on bringing everything together and making the

Starting point is 00:18:15 user experience more seamless. Chubby, for example, wrote, For hardcore users, GPD 5 will be a bit of a disappointment if the rumors are to be believed. Rumor has it that Sam Altman is not particularly impressed with the performance and improvements compared to older models such as GPD40 and O3. GPT5 is more of an iterative improvement that certainly shows significant leaps in benchmarks, but compared to benchmarks such as reasoning at the end of last year, 01, or deep research at the beginning of this year, GPT5 is more of the same, just slightly better. In this respect, GPT5 is probably not the qualitative leap that hardcore users had hoped for.

Starting point is 00:18:47 And yet Chubby wrote, maybe the hardcore users are not the point. They continue, for the vast majority, however, GPT5 will be a quantum leap. When I talk to friends, I almost always hear the same thing. ChatGPT is great. They say it will help them get more out of their university studies, answer all their questions, and even provide excellent advice on medical matters. When asked which model they would use for this, the answer is always the same. GPT40, of course.

Starting point is 00:19:11 They either don't know anything about 03 or due to the complicated nomenclature. They consider 03 to be the inferior model because it's older, because 3 becomes before 4. Some people think that simply. However, GPT5 will be an all-in-one model. Depending on the request, the appropriate amount of inference will be applied and reasoning will be carried out. This means that all those who previously used only GPT-40 will suddenly receive much better answers with GPT 5 than before, because they did not use or were not aware of the full

Starting point is 00:19:36 strength and range of Chad-G-GPT's model capacity. Now, this I think is a super-important point. The vast majority of ChatGPT users, even at this stage, are not subscribers. They haven't used deep research and they don't understand what a reasoning model is, or at least how it's different. In other words, removing ChatGPT's model selector isn't just a minor UX improvement. Instead, it fundamentally will broaden the average user's experience by making all of the myriad features available to them without them having to know about what those features

Starting point is 00:20:02 actually do. Think back to the deep seek moment at the beginning of the year. The viral breakthrough was not that there was better reasoning. The full version of 01 had already been out for two months and was clearly the better model. The innovation was simply putting reasoning right in front of users who had never experienced it before with the full reasoning traces on display. The average comment about Deepseek wasn't just that it was powerful, it's that it was really cute or cool because it talked to itself before responding.

Starting point is 00:20:27 OpenAI then has the opportunity to give the average and experienced user that type of moment of delight with GPT5. Even if every single feature already exists and the performance bump is minor, improved accessibility is a huge deal for the average GPT-5 user. And yet, I do think that unfortunately for OpenAI, that still might not be enough. Or at least not for long, it's very clear that in the wake of Mark Zuckerberg's poaching spree, the stakes for Open AI are building. Zuck's superintelligence team is now largely in place, and then it's worth then trying to speculate on what their big play is. In the lead-up to the release of Lama 4 during the spring, Zuckerberg set off on a media tour

Starting point is 00:21:04 discussing his AI plans at length. They were really two big themes. The first was making an automated advertising platform, and the second was introducing AI friends to meta-social media platforms. Hold aside what you think of those ideas. It is pretty undeniable that they're both fairly modest ambitions. In other words, neither is something you would necessarily want to spend hundreds of millions of dollars in payroll to achieve. And so whether Zuckerberg is aiming for a straight shot superintelligence play or iterative

Starting point is 00:21:30 model releases that are actually keeping up with and pushing the state of the art, he clearly has something much larger than just better advertising automation in mind. One observation that many have made is that Zuckerberg has put together a team of experts with a wide range of skills. He poached the reasoning team lead from OpenAI, a multimodal expert from Google, an edge model developer from Apple, and the list goes on. In other words, a full-stack team capable of recreating everyone Open AI or anyone else has on offer from scratch. Point being, whatever they're building, it's not AI friends. Altman says he's not concerned. In an interview at the Sun Valley Conference, he was asked how he's feeling about the talent war

Starting point is 00:22:05 and responded, fine, good. We have obviously an incredibly talented team, and I think they really love what they're doing. Obviously, some people will go to different places. There's a lot of excitement, I guess you could say, in the industry, but no, I think we feel fine. At the same time, it is pretty clearly undeniable that whether Sam is a part of this or not, OpenAI leadership in general is starting to feel the heat. We've seen changes in compensation packages, memos that suggest the feeling was of having their house broken into. And so for this reason, I think that the GPT5 moment, whether OpenAI wants it to be or not, is going to be seen as hugely significant and reflective of the state of play when it comes to the broader industry. Back in April, even before this aggressive talent war, Altman tweeted,

Starting point is 00:22:46 Change of plans. We are going to release 03 and 04 mini after all, probably in a couple of weeks, and then do GBT5 in a few months. There are a bunch of reasons for this, but the most exciting one is that we are going to be able to make GPD5 much better than we originally thought. We also found it harder than we thought it was going to be to smoothly integrate everything, and we want to make sure we have enough capacity to support what we expect to be unprecedented demand. So just how soon is this thing coming? Ultimately that among all of this is the biggest rumor. We have gotten a bunch of hints recently from OpenAI insiders that something big is coming in the next week or two, but I also think that OpenAI knows the stakes of GPT5. I don't think that they're

Starting point is 00:23:21 feeling so much pressure that they're going to release something that is anything less than extremely impressive. Still, maybe we have an exciting midsummer tree coming up. Certainly the chorus of rumors is getting louder, and as they get more credible, I will be sure to let you know it here. For now that that is going to do it for today's AI Daily Brief. Until next time, peace.

The AI Daily Brief: Artificial Intelligence News and Analysis - Everything We Know About GPT-5 So Far

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.