← The AI Daily Brief: Artificial Intelligence News and Analysis
The AI Daily Brief: Artificial Intelligence News and Analysis
The 10 Biggest AI Stories of 2025
00:28:50
Loading summary
Transcript1 lines
- [00:01]
A
Today on the AI Daily Brief, the 10 biggest AI stories of 2025. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. Alright friends, quick announcements before we dive in. First of all, thank you to today's sponsors, kpmg, Super Intelligent Robots and Pencils and Blitzi. To get an ad free version of the show go to patreon.com aidaily Brief or you can subscribe on Apple Podcasts. And for all of the information that you could possibly be looking about for the show, sponsorship, speaking, etc. Go to aidaily Brief AI now, we are in the early stages of our end of year coverage. From here on out, most of our episodes will be either looking back or looking forward. And today we're starting with the 10 biggest AI stories of 2025. Now these are not in ranked order. Instead I put them in a combination of a linear and narrative sequence. But I will call out when I hit my vote for the biggest story of the year. And we're going to kick off with the very first big story of the year, which was the absolute hullabaloo around the release of deep seek R1. Now, deep seek started to have models that people were paying attention to at the end of 2024. But in January, when they released their first reasoning model, R1, every everyone stood up and took notice. There were a couple of reasons for that. First of all, while all the American labs were spending hundreds of millions, if not billions of dollars to train their models, Deepseek was saying that R1 was trained for just a few million dollars. On top of that, however, Alongside the model, DeepSeek also released their very own Chatbot app and it rocketed to the top of the App Store charts, even displacing ChatGPT. For a while, as markets tried to digest the news, there was a deep sell off of AI stocks. Nvidia lost $593 billion in market cap in a single day, the single biggest one day loss in stock history. Now, of course, markets recovered, but this Deep Seq story set up so many of the themes that would shape the rest of the year. One that we'll discuss in a few minutes is the rise of reasoning. Part of what made the Deep Seq application so popular was that while OpenAI had released their O1 reasoning model at that point, and while O1 remained ahead of what you could get with deep seq R1, O1 was at the time entirely behind a paywall. So the vast majority of people had never seen a reasoning model. They were delighted both with the reasoning traces that Deepseek exposed in their app, as well as just the differentiated quality of the results. Of course, that market squirm would portend everything that we've been dealing with for the past five months around the AI bubble debate and from a lasting legacy perspective, one thing that was absolutely true about Deepseek was that Chinese models were much closer and nipping on the heels of Western closed source models than the vast majority of people had thought coming into the year. That has played out throughout the year with models like Canon QEMI as well as later Deepseek models being right up in the thick of things as some of the best models available. You can see Kimik 2 and Deepseek 3.2 behind Gemini 3 GPT 5.2 and Opus 4.5, but ahead of pretty much everything else, it would also kick off a back and forth debate around the appropriate US policy vis a vis China that has continued to be dynamic throughout the year, with the latest big change of course being the Trump White House deciding to allow Nvidia to sell H200 chips into China, the most advanced chip we've allowed to be sold to China in a number of years. All in all, the Deep seq story started 2025 off with a bang and it has not let down ever since. Our second big AI story for the year also kicked off in January, which was the massive AI infrastructure buildout. It started oh so innocently, just OpenAI and a couple of friends like SoftBank, MGX and Oracle announcing their intention to invest a half trillion over the next four years to build AI infrastructure in the United States. The initiative was called Project Stargate and it was announced at the White house on Tuesday, January 21st with President Trump in attendance with Oracle founder Larry Ellison, OpenAI CEO Sam Altman and SoftBank CEO Masayoshi Son. Now of course since then the AI infrastructure deals have done nothing but increase. Throughout the year we have seen a massive amount of hyperscaler capex and expansion with basically every major company Microsoft, Google, Amazon, Meta all increasing their guidance around their capex. For 25 and 26 we saw initiatives like the Global AI Infrastructure investment partnership between BlackRock, Microsoft, MGX and others, which was $100 billion investment vehicle focused on data centers and the electricity to power them. We had Elon Musk's XAI Colossus expansion which sees that company attempting to scale from their current a hundred thousand GPUs to a million GPUs or more. And of course with all this data center build out. There is also going to be energy requirements leading to announcements like the Google and Nextera Energy Partnership, which is an agreement to develop gigawatt scale data center campuses that have power generation on site thanks to an investment in nuclear. Now as we discussed, this was a theme throughout the year and right up until the end of the summer it was a major theme driving up stock prices. But then came the Oracle and OpenAI deal. At the end of August, Oracle revealed that it had added 317 billion in future contract revenue during its quarter that ended Aug. 31. That led the company's stock price to surge by as much as 43%, temporarily pushing his net worth up over even Elon Musk when a couple days later it was revealed that OpenAI was the customer driving about 300 billion of that markets started to get a little bit more nervous and this of course brings us to our next big story of the year, which is the AI Bubble debate. Now if we were just looking for what theme or topic was most discussed, particularly in mainstream media, for sure this is the biggest AI story of the year. Like I said, at least in terms of the amount of sheer ink spilt on it. Every week even to now sees an endless stream of AI bubble debate related articles. And interestingly, a lot of the focus is on Oracle, that big deal with OpenAI and the debt that they're taking on to finance the build out. One of the key themes of the bubble conversation is the circularity of revenue. I'm sure you've seen some version of this chart which shows the dense web of investment and customer relationships between major companies including Microsoft, OpenAI, Intel, Oracle, Nvidia, XAI and AMD. Now to some this screams House of Cards. To others it shows the dense web of relationships that is driving the mass AI ification of the economy writ large. AI bubble talk is so ubiquitous that it now has its very own Wikipedia entry, complete with a section on that circular financing. Now part of what makes this such a juicy and resonant theme is that it's one that's impossible to prove or disprove in the short term. In other words, even if we are in the midst of an AI bubble, the way that that would be manifest and problematic in terms of, for example, OpenAI missing financial obligations with these big deals is not coming to bear in the short term. That that means that it's ripe territory for narrative debates as market actors try to drag participants to their view of the world. Now one good resource that I've pointed to before, if you are interested in. This story comes from Exponential View, who put together a boom and bubble monitor. This came out of a blog post where they looked at five historic indicators for financial bubbles, economic strain, industry strain, revenue momentum, valuation, heat and funding quality and now turn them into a live tracker. Now at this stage, they argue, we are still firmly in boom territory with only one in the five gauges in the red, which is the industry strain. That said, there is a lot to watch here and it's a great resource. You can find it at boomerbubble AI now moving on to our next story, one that I have to begrudgingly include. If the AI bubble debate was the most debated topic of the year, the most referenced media of the year, to my great chagrin, was the MIT report that argued that 95% of generative AI pilots are failing. Now, in my notes about the 10 biggest stories, I called this Enterprise Adoption and the MIT Lie. And while I've talked about the MIT report a lot, I do want to, one more time and for posterity as part of this recap episode, rip it to shreds for the utter garbage that it is. Two big reasons for that. First of all, the methodology, and second of all, the incredible and incorrect leaps of logic that are embedded in the analysis. So first of all, from a methodology perspective, this study, which I say in the biggest, most aggressive air quotes I can manage, looked at a couple of things. First, it looked at recent earnings report of public companies who mentioned AI to see if any of them talked about revenue acceleration. It then paired that with around 50 convenience interviews from random executives they apparently had access to. This is the entire methodology for this thing. Not only is that a radically underwhelming data source, but the idea that an organization not mentioning revenue gains from AI in a report means that their pilots are failing is absolutely ludicrous. Again, one would think that with a headline backed by someone as prestigious as MIT that says that 95% of pilots are failing, you would assume that they asked a bunch of enterprise AI leaders if their pilots were succeeding or failing and 95% of them said that they were failing, right? But no. This is an inference from a missing articulation of revenue gains in earnings reports and nothing more. Now, if we can be charitable to the study authors for a moment, they obviously didn't know that it was going to have the impact that it had, and it became caught up in something that was much bigger than just the one report. However, still, frankly, it didn't befit the MIT name. And I do think they should be embarrassed at the quality of their thinking. Now, packing my soapbox away for the rest of the episode. We do have to acknowledge that there was a reason that this report was so resonant. It came into a ready and waiting environment where a combination of factors made this report have an element of confirmation bias. The first was that markets were starting to turn, and this seemed like perfect evidence of why. A huge part of the amplification that happened in the first couple weeks after this was announced came from Wall street analysts and investors who, who were talking about it as part of their assessment of AI markets. But the second thing was hold aside the AI bubble debate, a lot of the learnings of 2025 from an enterprise perspective were around the theme that to be good at AI and to really get the value out of this technology, it was going to take more than just dropping a chatbot on top of your people. Obviously, sophisticated organizations never thought it was going to be that simple. But at the time this study came out, there was the beginning of a broad recognition that, okay, to really get the full value out of AI, we're going to have to think in bigger, more comprehensive and systemic terms. We're going to have to redesign systems, we're going to have to address data readiness, we're going to have to think about the context that we give agents, and that is the real substantive piece that it interacted with. Still, if you want to know the actual story of enterprise adoption over the course of 2025, it was that even as all that learning that I was just mentioning and realization was happening, that to really get the full value out of AI and agents, it was going to take more. Adoption was still just a steady lineup. And not only was it a steady lineup, the AI implementations that were happening were leading to value. In our AI ROI benchmarking study, we found that around 44% of use cases were reporting modest ROI and about 38% were reporting high ROI of either significant or transformational impact. Only 5%. Basically the exact inverse of the MIT study reporting negative roi. And keep in mind, negative ROI does not mean failure. Negative ROI means failure to reach roi. Yet where the outlay of resources is still higher than the gain from that outlay of resources in the short term. But if you look at leaders who are interacting with AI, 2025 saw their optimism about the value of this technology go nothing but up. Comparing KPMG's global CEO study from 2024 to 2025, in 2024, the majority of CEOs, 63% said that they expected to see ROI from AI in three to five years. 20% of optimists thought it would be one to three years and 16% of pessimists thought it was going to be more than five years. By 2025 that had pulled forward massively. Two thirds of CEOs surveyed in 2025 thought that they would see ROI within one to three years. Instead, 19% said that it was just six months to a year away and now less than 2% thought it was going to take more than five years. Look, I do think it is worth understanding why this MIT report, as bad as it was, struck such a nerve. But when you peel the layers away, the story of Enterprise adoption in 2025 is more adoption, starting ROI and a real recognition that to get the next set of value, it's going to take more work. Sure, there's hype about AI, but KPMG is turning AI potential into business value. They've embedded AI and agents across their entire enterprise to boost efficiency, improve quality, and create better experiences for clients and employees. KPMG has done it themselves. Now they can help you do the same. Discover how their journey can accelerate yours at www.kpmg.usagents. that's www.kpmg.us agents. Today's episode is brought to you by my company, Superintelligent. Superintelligent is an AI planning platform, and right now, as we head into 2026, the big theme that we're seeing among the enterprises that we work with is a real determination to make 2026 a year of scaled AI deployments, not just more pilots and experiments. However, many of our partners are stuck on some AI plateau. It might be issues of governance, it might be issues of data readiness, it might be issues of process mapping. Whatever the case, we're launching a new type of assessment called Plateau Breaker that, as you probably guessed from that name, is about breaking through AI plateaus. We'll deploy voice agents to collect information and diagnose what the real bottlenecks are that are keeping you on that plateau. From there, we put together a blueprint and an action plan that helps you move right through that plateau into full scale deployment and real roi. If you're interested in learning more about Plateau Breaker, shoot us a note. ContactSuper AI with Plateau in the subject line, AI isn't a one off project. It's a partnership that has to evolve as the technology does. Robots and pencils work side by side with clients to bring practical AI into every phase. Automation, personalization, decision support and optimization. They prove what works through applied experimentation and build systems that amplify human potential. As an AWS Certified Partner with Global Delivery Centers, Robots and Pencils combines reach with high touch service where others hand off. They stay engaged because partnership isn't a project plan, it's a commitment. As AI advances, so will their solutions. That's long term value. Progress starts with the right partner. Start with Robots and pencils@ropotsandpencils.com aidaily 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. Enterprise engineering leaders start every development sprint with the Blitzi platform, bringing in their development requirements. The Blitzi platform provides a plan, then generates and pre compiles code for each task. Blitzi 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 Blitzi as their pre IDE development tool, pairing it with their coding pilot of choice. To bring an AI native SDLC into their org, visit blitzi.com and press get a demo to learn how Blitzy transforms your SDLC from AI Assisted to AI Native. Our next major story of the AI year has to be the AI talent wars. Now, talent was always valued in AI. That was never a question. Top researchers inside the big labs have for a number of years been making very, very hardy salaries that would make most people extremely happy. However, around the middle of this year that started to get to new extreme levels as competition between the labs for talent started to ratchet up. Now a little bit of that was spin outs from the labs who were bringing people along with them. OpenAI's former CTO, Mira Muradi, started her own Thinking Machines lab, bringing a bunch of talent with her. Another former OpenAI leader, Ilya Sutskever, started his SAFE superintelligence, once again recruiting a bunch of talent away from the other labs. But where things really heated up was the middle of the summer when Mark Zuckerberg started recruiting for his superintelligence. Lab reports started coming in of just absolutely crazy offers. In June, Sam Altman said that Meta had offered some OpenAI staff up to $100 million, bragging at the time that no one had taken him up on that offer, although that wouldn't last for long. And the numbers just got crazier. From there we started to see more and more of those nine figure offers, and people started making the comparison to professional athletes. Sequoia even wrote a piece called why AI Labs Are Starting to Look Like Sports Teams Now. In many ways, this culminated with the sort of, but not exactly acquisition of scale AI by Meta, which cost Meta $15 billion and seemed like mostly a way for them to get their hands on Scale CEO Alexander Wang to lead that superintelligence lab. And while the insane headlines about nine figure deals may have died down over the course of the fall, the AI talent wars continue apace. More recently, what we've been seeing is the gutting of some incumbents, particularly Apple, who are having an extremely hard time keeping talent right as their AI strategy flounders. Now we'll see how this all shakes out heading into 2026, but my guess is that talent is going to continue to be a key battleground for all these labs as we head into next year. From here we move into some stories that are a little bit more about the substance of AI rather than the market and the ecosystem around it. The next story is one that is so ubiquitous and surrounding us that it might not even seem like a story, as it was just our reality throughout the year. And that's what I'm calling the rise of Reasoning. I mentioned back in the Deep SEQ story that a big part of why their app rocketed to the top of the app charts was that it was the first time that most free AI users, which obviously represents the vast majority of them, had used a reasoning model. And of course once you use a reasoning model, it is very hard to go back towards the end of the year. We got some numbers around this from Open Router Open Router is a platform that allows developers to connect their applications to a variety of LLMs, meaning that they don't necessarily have to be locked into one ecosystem, but there can be model switching based on different needs or based on the models going down or whatever the reason is. And over the course of the year and a hundred trillion tokens from a starting point of basically zero at the beginning of the year, reasoning tokens now represent over 50% of the total consumed. If you used O3 or Gemini 2.5 Pro or Claude after 3.7 or Gemini 3 or GPT5 or basically any model in the second half of this year, chances are that by default you are using reasoning models. Now. That said, and the reason that I wanted to call this out as an explicit story is that while this may be obvious to us in the space, the difference between reasoning and non reasoning models is not necessarily widely known outside of AI users. Professor Ethan Malik referenced a recent study that found that clinical LLMs could ace medical exams but at the same time perform weakly on realistic clinical tasks. The problem is that the study was using GPT 4 and Claude 3 opus, Ethan wrote, I hate to keep bringing this up, but studies cannot lump reasoners with earlier models when considering AI abilities. And while studies don't always need to use the latest models, they should test to see if there are trends in ability as model size scale to anticipate the future. Now, of course, part of what the reasoning models opened up is our next Big Story of the year, and the one that if I had to commit to a single biggest story of the year, would absolutely be my number one is the emergence and growing ubiquity of five Coding man, what to say about Vibe Coding that hasn't already been said? It started with such humble origins, this tweet from back in February from Andrej Karpathy. He said, there's a new kind of coding I call Vibe coding, where you fully give in to the vibes, embrace exponentials and forget that the code even exists. It's possible because the LLMs, for example cursor, composer with Sonnet, are getting too good. When I get error messages, I just copy paste them in with no comment. Usually that fixes it. The code grows beyond my usual comprehension. I'd have to really read through it for a while. Sometimes the LLMs can't fix a bug, so I just work around it or ask for random changes until it goes away. I'm building a project or a web app, but it's not really coding. I just see stuff, say stuff, run stuff and copy paste stuff. And it mostly works. Now of course, Vibe coding was shorthand for a much broader array of AI and agentic enabled coding. We saw massive growth in consumer apps like Lovable and Replit. But then we also saw the rise of Cursor and Cognition and these tools that were for AI enabled and agentic coding. But for professional developers, pretty much universally, it's acknowledged that coding became the first most important use case of Gen AI, which was expressed in the numbers. Menlo Ventures, in their annual study of enterprise AI, found that 55% of departmental AI spend about $4 billion could be attributed to coding. The next highest category was it at 700 million. Replit and Lovable both surged ahead of $100 million in ARR and have continued to grow. Meanwhile, Cursor is closing in on 800 million, making these companies some of the fastest growing revenue companies in history. Indeed, Vive coding became so ubiquitous that by the end of the year, the conversation had shifted a little bit inside and around professional developers and software engineers. That group is now in many cases wrestling with the downsides of Vibe coding, whether it's the amount of review that's required, or technical debt that gets created, or the atrophy of key coding skills. On top of those issues, there's also just questions of how to design the modern AI coding stack. How much and in what context do people want super fast AI assistance versus full automation that does just big chunks of the coding work for them? Whatever the Case Model releases throughout the year have shown that for the big model labs, there is nothing more important than the coding use case, with basically all of them seeing it as key not only to unlocking the coding market, but as key to making AI capable for other general use cases. I think as we head into next year, we're going to start to see a fork in the Vibe coding conversation. Right now we're still talking about AI enabled and agentic coding that professional software engineers and software engineering organizations are doing with the same set of terminology and in the same breath as what non technical people are doing with code for the first time. I don't think those are really the same thing and I think those conversations are going to break apart a bit. I also think frankly that as ubiquitous as Vibe coding was this year, the impact that it is poised to have in 2026 is even greater. In other words, I don't see this as a trend that will dissipate into all the other things that you can do with AI. Instead, I think this is a fundamental capability shift that will change how a huge portion of knowledge workers do their work forever going forward. I think we've barely scratched the surface on that, which is why of course, I'm exploring it through a couple of different interviews throughout the course of these end of year episodes. Now, staying on the coding and agent theme a little bit, a lot of people had 2025 pegged as the Year of Agents. I actually tend to think that was true, although it meant something different than we thought going into the year. Part of that was that it was the year of coding agents, but another part of that was that a lot of the key events of this year were about agent infrastructure, the rise of context, and the decisions that all of the competing Model labs made to go in on the same set of standards. And in order to all move further faster, Anthropic introduced the model context protocol at the end of 2024, and it got some initial attention. However, towards the end of February and into March is when it really started to capture people's attention and became a major theme for AI builders everywhere. Mcp, of course, was a way for agents to connect to external services and data sources, greatly expanding what those agents can do. And one of the things that was really interesting is that if you look back at the history of computing, there have often been standards wars that lasted years at a time where groups who wanted one set of standards fought against groups who wanted another set of standards, all of which ultimately served to slow down overall development in whatever field they were in. That did not happen this year. You could tell as soon as MCP hit that inflection point that the other labs considered competing and then ultimately decided to just get on board. On March 26, Sam Altman tweeted, people love MCP and we are excited to add support across our products. On March 30, Alphabet CEO Sundar Pichai wrote to MCP or not to MCP? That's the question. Let me know in the comments. Followed up on April 9 with Love. The feedback to MCP it is. And it wasn't just mcp. Other parts of agent infrastructure also saw similar uptake across the labs. Also on April 9, Google announced the Agent to Agent protocol. Agent to Agent, like it sounds, is an agent communication protocol. It was explicitly framed when it was announced as a complement to mcp, and within a month you even had Google competitor Microsoft embracing A2AMore recently, we've seen a similar phenomenon with anthropic skills. Skills are a way to give generalized agents access to specialized context knowledge or instructions using a file and folder system. And in December, OpenAI started supporting the framework as well. Now, on top of all this agent infrastructure, we also had the emergent discipline of context engineering. Whereas prompt engineering was all about figuring out the right way to prompt an LLM to get the results that you wanted, Context engineering is all about making sure that the LLM has access to the right information or context to do the work that you're hoping to have it do. Taken together, all of this kind of makes 2025 the year of agent infrastructure, which of course sets up 2026 to be the year of agent impact in practice. Now, of course, I'm sure there's a lot more infrastructure still to be built, and these lines are ultimately pretty blurry, but I think that this focus on context and the emergence and rallying around of key agent infrastructure is a key AI story of the year lastly, today in our 10 biggest AI stories of the Year are what I'm calling collectively the Next Leap models. With this I'm referring to Gemini 3, Opus 4.5 and GPT 5.2. Now, I had initially planned to include in this episode my countdown of the most impactful model releases of the year, but obviously at this point you can tell the show is getting pretty long, which I probably could have predicted, and so I'm going to move that into its own episode. But for our purposes here for this last story, when GPT5 came out, it was for many a big disappointment. In fact, it helped really fuel the AI bubble debate. The chorus of people saying that AI had hit a wall got louder and louder, pointing to GPT5 as evidence. All of this meant that there was huge pressure on Google leading into the release of Gemini 3, and it was a challenge that they undeniably met. Google in fact released two incredibly important models in November, Gemini 3, as well as their image model Nanobanana 2. One impact was for Google itself. For the first time really since ChatGPT launched, Google appeared to be in the driver's seat across the industry as a whole. But even beyond that, there were impacts on the market as well. Gemini 3 served to counteract the narrative that AI had hit a wall, giving more optimism that we'd see continued growth and adoption, which of course could help justify these big deals over the next five years that markets were trying to figure out how they should price in. Just as the AI community was digesting, Gemini 3 anthropic dropped its most advanced model, Opus 4.5. Now it's been a few weeks since this came out and I don't know that I've ever seen a model that started on such a high note in terms of people's perceptions and just continued to grow in esteem. Many people have argued that Opus 4.5 represents a fundamental level up when it comes to the coding capabilities of AI. I've seen people reset their timelines and how they think about the future of software engineering jobs because of Opus 4.5. Even for non software engineers, Opus 4.5's capabilities have found their way into the vibe. Coding apps like Repl.it and Lovable Transforming with those platforms can do competently as well. Now of course, all of this prompted some concern from OpenAI in the weeks leading up to Gemini 3. An internal memo that was later leaked saw Altman forecast some rough vibes due to a resurgent Google. The rough vibes memo was upgraded to a full on Code Red and a shift in priorities away from a bunch of long term and short term efforts to just focus on ChatGPT and new model releases. It was that Code Red effort that got us an advanced release of GPT 5.2, which, while not necessarily seeing the same universal praise as Gemini 3 and Opus 4.5, certainly has a lot of proponents, including myself, who think that GPT 5.2 Pro is for my use cases around business strategy. The best model out there. But you take it all together and this set of NextLeap models have not only demonstrated that AI development hasn't really hit a plateau, but also leaves us heading into 2026 with veritable superpowers compared to where we were heading into 2025. And of course it doesn't appear like that's going to slow down anytime soon. We are anticipating at least another OpenAI model in January and you gotta think that the other labs won't be far behind. And so friends, that is my list of the 10 biggest AI stories of 2025. Like I said in another episode I will countdown and it will actually be a countdown of the most impactful model releases of the year. But for now, that is gonna do it for today's AI Daily Brief. Appreciate you listening or watching as always. And until next time, peace.