My Autumn AI Predictions

0:00Today on the AI Daily Brief, as the brr months begin, you are getting my autumn AI predictions. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

0:17Hello, friends. Welcome back to the AI Daily Brief. Quick announcements before we dive in. First of all, thank you to today's sponsors, KPMG, Pencils and Robots, Superintelligent and Blitzy. To get an ad-free version of the show, go to patreon.com slash aidailybrief. And if you are interested in sponsoring the show, send us a note at sponsors at aidailybrief.ai. Now, I have a fun one for you today. The news cycle is slowly starting back up. It is back to school time here in the US, which also means back to work time for all intents and purposes. This is really the time of year that everyone shakes off the summer sunshine and gets back down to business.

0:51And I love it. It has since I was very young and this was the school year beginning felt more like New Year's to me than New Year's. And so rather than just do a retrospective on what happened this summer, which if you were listening to the show, you'll already have heard, I want to translate that into some predictions. So what I'm going to do is I'm going to look at where we are and how I think it plays out over the course of the next couple of months. And I think where we have to start is with continued simmering skepticism around AI. We right now have an unholy alliance of market shorters, AI skeptics, and media headline editors who are all keeping this MIT study headline alive.

1:28This, of course, is the study where MIT found that 95 % of all AI pilots were going nowhere and creating no value, which was translated by headline writers into 95 % of AI efforts are completely worthless. Never mind the fact that this was based on around 50 interviews and a reading of public financial statements to see whether companies were declaring that AI had increased their revenue or not, a very dubious methodology at best. Now, interestingly, however, you're also starting to see a lot of companies who want to sell you something, lean into the MIT story as well. Basically, the extension of the narrative is that yes, of course, many pilots are meaningless, but it's not because AI doesn't work, it's because it wasn't implemented well.

2:07The implication, of course, being that if you work with them, they can help you implement it well, which honestly, I'm totally fine with this narrative. Secure the bag, my AI entrepreneur friends. But the point is that surrounding all of this, we are heading into this fall with a little bit of AI skepticism. Usually at this time, the summer AI antagonistic narrative has sort of worn off. But of course, it wasn't just the MIT study. It was also Sam Altman sort of a little bit saying we're in a bubble, although even that was widely overreported, I think. But as I said before, I think that part of the reason that this hit the way that it did when it did is that it came in the context of the broader market cycle.

2:42Summer is a historically low liquidity kind of time when there's not all that much market activity, and narratives that otherwise wouldn't move the needle have a more outsized impact. Take on top of that, anxiety around the rate cutting cycle, as well as tariff policy, although really for most of August it was about the rate cutting cycle, and Wall Street's general concern that AI is the only thing propping up this market, the TLDR for me is that I think that the response had a lot more to do with where specifically the market is than anything that you have to be concerned about if you're thinking about how to use AI for your own benefits or for your company's benefits.

3:18So, predictions. First of all, I think that this story has some more legs from where it is right now, but ultimately will feel much less significant in the future than it was presented when it came. Same as previous AI skeptic narratives in years past. I think this is especially true if markets get the rate cut that they want in September. At the same time, I also predict that there is going to be a positive outcome of this, which is a return to focus on some things that have gotten lost, which is the real challenge of actually implementing AI in the enterprise. Remember, another part of the reason that the MIT study hit when it did was that it came right after the announcement of GPT-5, which was, for a time, viewed with a lot of skepticism.

4:00Now, at this point, the narrative, I think, has shifted fairly dramatically, and people are enjoying GPT-5, figuring out how to get more out of it, et cetera, et cetera. But all of this is combined to create a conversation where companies are thinking about not just how well-performant a model is, and assuming that a more powerful model means that their experiments are going to work. But instead, this is creating context to go even deeper on the real challenges of implementation, the real challenges and opportunity of the last mile. You're starting to see these other types of pieces pop up, like this one in the Harvard Business Review that's all about how companies can escape the AI experimentation trap.

4:37That's their term, and I think it's a good one. And hold aside their advice, which is basically to focus on the use cases that have real business opportunity and double down on them rather than spreading yourself too thin. The point isn't so much whether that advice is right or not, although I do think it's fairly sound. And instead, the fact that this is the conversation that's coming out of this. Organizations by and large are not looking at that MIT study and justifying a lack of behavior. Instead, they're saying, how can we do this better? And I think that's a really positive outcome. I think, in fact, that it is going to create the context for a much broader conversation about infrastructure and why that should be where companies place a lot of their emphasis.

5:17Which brings us to our next prediction. I recently said that I thought that 2026 was going to be the year where the big theme in enterprise AI was context orchestration or context engineering. And the idea of context orchestration or engineering is basically that AI and agents can only get you so far in an enterprise if they don't have good access to the context in which they are operating. In other words, if your agents don't have the relevant background, the relevant data, the relevant information that can help them contextualize their work to your organization, there are just going to be really severe walls that you run up against very quickly.

5:49It's very natural at the beginning of a technology cycle to focus on big flashy pilots and exciting proof of concepts to basically move the ball down the field by doing things that are fun to show off rather than a lot of the laborious hidden behind the scenes work that ultimately is going to produce good results. Now, frankly, AI has always been a little bit of an exception to this rule in the sense that pretty quickly organizations started to try to get their data houses in order and understood that there was likely to be a direct relationship between how good at giving AI access to information and context they were and how well it was going to perform for them.

6:22But I think that this moment with GPT-5 and the MIT study are going to give enterprises even more narrative cloud cover to really have the new hotness be the messy, difficult infrastructure work that is so necessary for getting to the next level of performance when it comes to AI and agents. Now, a lot of this has already started this year. One of the major themes, I think, when we look back on 2025 will have been MCP, Model Context Protocol, and these new standards around agent access to information that so many enterprises are adopting quite quickly. You're also seeing conversations spring up around other standards like A to A.

6:56Ravine, the CEO of Credle, recently wrote on his LinkedIn, I'm not a fortune teller, but my educated guess on what will dominate AI headlines in 2026 is agent to agent or A to A workflows. Over the past few months, my conversations with AI experts, customers, and other builders all point in the same direction. Single agents aren't enough. We've seen AI evolve in waves. 2023 was the year of AI chat. 2024 was all about LLM wrappers. 2025 has been dominated by AI agents. And I believe 2026 will be the year A-to-A workflows take center stage. He argues that the two reasons why are first, massive tool and data sprawl that happens inside companies that make a single agent managing all tasks across those systems really unrealistic, and the fact that communications infrastructure like A to A helps us build more specialized agents that can then work together.

7:44You already started to get a taste this year of MCP showing how infrastructure build-out can be sexy. I think we're going to see that do nothing but increase, and I think that some of the skepticism that we got in the summer is actually going to create some momentum to move even faster in that area. Then again, I also think that we're starting to see a bunch of counter signals to that skepticism from the summer. And one of my predictions for the fall is that you're going to see a bunch of people race in the other direction over the next few weeks. An example of this came on Monday when Evercore bucked the trend of calling things an AI bubble and instead released a note predicting a 20 % rally for US stocks by 2026, driven in large part by excitement around AI.

8:24Evercore's report basically said that if you zoom past all of the hand-wringing and nervousness, earnings continue to defy expectations despite tariff and policy uncertainties. Now, they're not saying it's a sure thing that we're going to see another big rally. But remember, in the last quarter, we saw CoreWeave triple their revenue. Microsoft's Azure jumped 39 percent and NVIDIA was up 56 percent, which, yes, was less than the 200 percent the year before, but is still a remarkable result for the biggest company in the world. I think that this sort of market positivity is going to surge, especially after we get whatever we're going to get when it comes to rates at the end of this month.

8:58And when it does, people are going to look back on some of the stories that they decided to conveniently not focus on in the summer, like Sam Altman saying that they were expecting to spend trillions on infrastructure quite quickly despite the fact that they knew that economists were going to say they were nuts, and that is going to help fuel this narrative even further. Another counter-signal comes around fundraising. This isn't so much a prediction, just an observation, although to the extent that I think there's a prediction here, I think it's going to do nothing but increase. But the TLDR is that at least when it comes to private markets, appetite for AI investment remains unabated.

9:27Biggest example of this is that Anthropic has now closed this round that has been increasing in both total size and total valuation over the last couple of months, coming in at a hot$183 billion valuation. The company ended up raising$13 billion, which was like three or four times as much as it seemed like they were going to at the beginning, and at a significantly juiced valuation from where they started. Now, of course, the reason that Anthropic was able to command such a big increase in their valuation is that the company has just been absolutely crushing it. They 5X their revenue from$1 billion annualized in January to$5 billion annualized in August, thanks in large part to the rise of the agentic coding use case.

10:03Speaking of which, more evidence of the countersignal in fundraising. Vibe Coding Platform Lovable, fresh off of a raise at a$1.8 billion valuation, the year-old startup, is now apparently receiving offers at double that at a$4 billion valuation. Point is, I don't think these are outliers. I think these are reflective of the fact that we are just going to see nothing but more private investor enthusiasm for hot AI deals. And boy, at this point, we are definitely in the okay, but that's so obvious portion of these predictions that Vibe, or perhaps as it will now be known, agentic coding is going to continue to dominate as maybe the key theme of 2025.

10:38Vibe coding wasn't even a term when we started this year. Tools for AI coding had only just started to really get good enough, but obviously over the last six months, it has become a dominant force. We saw this expressed in Andreessen Horowitz's recent Top 100 Gen AI Consumer Apps, where Lovable jumped from just off the list six months ago to number 23, and was one of the big themes that A16Z called out, in part because we were seeing an increase not only in traffic to the core websites of these companies, but also to the domains where they hosted users' creations, suggesting that people are actually building real stuff that they're then publishing for other people.

11:10I do predict that the language of vibe coding is going to go a little bit by the wayside. I'm already noticing people starting to call it simply agentic coding more often, And I think that's a more accurate reflection of the full range of how it's being used. Now, we might still use Vibe Coding to refer to people who are not traditional software engineers using these tools, or maybe we'll come up with some other term for that. But I think that agentic coding will continue to be a force. It will continue to be used in higher and higher value and more complete production uses. I think that inside companies, it'll start to jump out of the prototype phase and will have specific buckets of work that become normalized to be completed with agentic coding.

11:45I think that we're going to see more startups who start to fill in the gaps of agentic coding. such as the just released code review agent from Lindy that can look at your Vibe Coding app and see where it needs to be improved. And again, prediction for the fall is that by the time we're doing end of year lists, agentic coding will be seen as perhaps the most important force in AI of 2025. What if AI wasn't just a buzzword, but a business imperative? On You Can With AI, we take you inside the boardrooms and strategy sessions of the world's most forward-thinking enterprises. Hosted by me, Nathaniel Whittemore, and powered by KPMG, this seven-part series delivers real-world insights from leaders who are scaling AI with purpose, from aligning culture and leadership to building trust, data readiness, and deploying AI agents.

12:29Whether you're a C-suite executive, strategist, or innovator, this podcast is your front-row seat to the future of enterprise AI. So go check it out at www.kpmg.us slash AI podcasts, or search You Can With AI on Spotify, Apple Podcasts, or wherever you get your podcasts. Today's episode is brought to you by Robots and Pencils. When competitive advantage lasts mere moments, speed to value wins the AI race. While big consultancies bury progress under layers of process, Robots and 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.

13:09With AWS-certified teams across US, 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. If you are a regular listener, you will have heard about superintelligence agent readiness audits at this point. But I wanted to tell you today about the full suite of agent readiness products that go beyond just the initial readiness report.

13:46Over the last six months, Superintelligent has built out an entire agent planning suite. We help you move from discovery to planning to implementation. After you've completed your agent readiness audits, we help you double-click on your most important use cases with what we call our use case planning reports. These reports are going to help you understand what sort of technical preparation you need to do to be ready for a use case, what challenges you might face in implementation, and whether you should be thinking about building, buying, partnering, or some combination. After that, you can even get a spec document in what we call our technical blueprint that gives either your developers or the developers of the partner you work with what they need to build exactly the agent that you're looking for.

14:24If you want to learn more about Superintelligence Agent Planning Suite, we've built a custom GPT to answer your questions. Just go to bit.ly slash super super agent. That's bit.ly slash super super agent, all one word. And if you have any questions, the agent can even help you book an appointment with our team. 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 Blitzy platform, bringing in their development requirements.

15:03The Blitzy platform provides a plan, then generates and pre-compiles 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 SDLC into their org. Blitzy 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.

15:36Visit blitzy.com and press Book Demo to learn how Blitzy transforms your STLC from AI-assisted to AI-native. That's blitzy.com. Let's talk about what we can expect from models. One prediction that I have is that as we get out of the summer doldrums, and we get farther away from the release of GPT-5 and further past the MIT study, etc. I think that we will realize that while yes, we may be seeing the saturation of chatbot-style use cases, and progress is moving more incrementally when it comes to the difference between GPT-4.5 and 5, and eventually Gemini 2.5 to 3 and GroK 4 to 5, etc., that while we were all focused on that and looking over in that direction, multimodal capabilities were just absolutely exploding all around us.

16:22There are a bunch of big examples of this from the summer, a huge number of which, by the way, come from Google. Most recently, of course, we got NanoBanana. I've talked extensively about NanoBanana, which is, by the way, technically Gemini 2.5 Flash Image is its real name, but where NanoBanana excels is not in how much better it is at native generation, it's how much better it is at making changes to existing photos. I'm working on an episode for sometime this week or next week where I talk about this new idea for a benchmark I have called the Unlock Score that was inspired directly by seeing how many new use cases this simple change opens up.

16:57Basically, while some people have complained that Nano Banana doesn't represent some major increase in raw generative capacity, the fact that it can allow for more precision photo changes creates a huge number of commercially viable use cases that used to either be A, only possible through an extended set of workflows that were all cobbled together, or B, not possible at all. It is, of course, not just images that have gotten a major upgrade. We haven't really gotten an update in a while, but what we are seeing now is VO3 actually making its way into production for a lot of different use cases, particularly around social media and advertising.

17:34VO3 is the first model that has made it fully into production for ads that are airing on national TV in major places. And that is a trend that I think is going to do nothing but increase, especially as it gets plugged in with models like Nano Banana that make the process of going from image to video more viable and just better. Indeed, a lot of what you're seeing right now on the internet is people combining the capabilities of these models, as well as apps like Higgsfield, to make AI videos that can more natively translate crazy imaginings into something really powerful and clear. There's this other X factor as well of world models.

18:07At the beginning of August, we got a preview of Genie 3, which made some major advancements when it comes to the ability to create 3D interactive worlds with just a text prompt. Specifically, we got much more extended memory that again opens up a whole variety of new use cases. Most people are barely touching images and video to say nothing of these world models, but this could be a whole new frontier for creation and is happening at a more rapid clip than I think most people would have predicted. So my prediction for this area of multimodality is that again, as we shake off the slough of debates from the summer about MIT and GPT-5, I think people will realize that they have this totally expanded creative canvas that they have barely scratched the surface on.

18:47And I think we're going to see a lot of that come to actual production as people dig in and try these rapidly improved, step function improved type of multimodal models out. Now, let's talk about cost. The state of the cost of AI is really fascinating to me. On the one hand, we're seeing the cost of inference go down and down and down in a way that basically no one would have thought possible. I mean, it just massively exceeds anything that we would have expected or any comparison to Moore's Law. At the same time, so far, most people for most use cases have opted to just focus on the highest-end models, and so total cost continues to rise.

19:24The Wall Street Journal wrote about this over the weekend with a piece called Cutting-Edge AI Was Supposed to Get Cheaper. It's More Expensive Than Ever. And obviously what's going on here is not about AI not getting cheaper. It's actually about the fact that as it gets more powerful and cheaper on a unit basis, we use more of those units. Aaron Levy from Box wrote about this, specifically in the context of the journal piece, and said, This is precisely Javon's paradox in action in the purest form. Because the cost of AI tokens have gone down, we can now afford to use far more of them for increasingly complex tasks.

19:56The key point, thus, is not that AI is getting more expensive. Instead, it's that because it's getting cheaper and more capable, we're using more of it to solve problems better. For almost every like-for-like task, we're just using way more tokens to complete the task to deliver far better output. Whether it's writing code, answering a healthcare question, or analyzing a contract, we're using far more AI today to perform that work because we need the additional points of performance. Getting a 99 % correct answer when working with a legal contract is very different from a 90 % correct answer, and it's easily worth the 10x to 100x increase in tokens.

20:28Now at some point, we will start to reach plateaus for certain types of tasks, and then the cost per task will go down. For instance, we probably don't need 100x more tokens than we use today for answering a simple medical question or summarizing a document. So then eventually, on a like-for-like basis, these workloads will become cheaper as we're able to capture the efficiency gains from the models. But the general cycle will go on essentially forever, because we will just keep raising the bar of what we do with AI. As tokens continue to get cheaper due to algorithmic breakthroughs, competition in GPU prices, general compute efficiencies, and open-weight alternatives, we will find the next set of ways to consume the tokens.

21:02We'll deploy far more agents in parallel to speed up tasks, we'll use multi-agent systems to compare answers and get to consensus, we'll solve more complex knowledge work problems, and we'll have far longer running agents in the background. AI will both simultaneously always be getting cheaper and more expensive. It is very rare that someone says exactly what I would have said, but this is exactly one of those times. Going back to this idea that agent decoding is the key use and the breakout use of AI in 2025. We've moved from sitting there telling a coding tool what to do to now running multiple coding agents in the background in a way that just simply consumes more tokens.

21:38We also continue to ride the frontier where each increase in AI capability opens up new uses that weren't possible before. And to the extent that anyone's most important work is in that set of new use cases, they're kind of priced into using the most expensive versions of the models. However, at some point, there will be entire categories of tasks that can comfortably use models that aren't the state of the art and can take advantage of ridiculously reduced prices and inference. So my prediction, exactly as Aaron said, for a little while at least, even though AI costs are coming down dramatically, the things that get enabled both by new capabilities and by the cost coming down are going to ironically increase the total amount that we spend on AI because we're simply going to be using more of it.

22:25Last day to the models conversation, new data sources. We continue to have lurking around us a conversation that's been going on since about this time last year, which is whether we've reached a plateau in pre-training as a scaling methodology. Certainly for some people, the more incremental performance between something like GPT 4.5 and GPT 5 suggests that yes, we have. And while I think I've clearly shown that there are other areas where models are improving at every bit as exponential a clip as anything we've seen before, There is still this conversation around plateaus and specifically LLMs and what it means in terms of what we can predict for the future.

22:58Part of the challenge is the incremental cost of compute, but part of the challenge is also access to data. At this point, every model has been trained on the entire corpus of publicly available human information. That's just basically what you can assume. And so to train on more novel data, companies are going to have to go find more novel data. I predict that rather than just resting on their laurels, AI companies are increasingly going to be looking for those novel sources of data. One example of this is it's going to be very default, even more default than it is now, for AI to be training on your interaction with it.

23:31Anthropic was a long-term holdout around this, but even they have finally changed their terms of service to force an opt-out when it comes to general consumer usage being included in the training data. You're also starting to see companies go out and look for novel sources of data. The information today posted this piece, OpenAI and XAI show interest in cursors coding data. Startups that sell AI-powered coding assistants have created some of the fastest growing businesses in Silicon Valley, making them ripe acquisition targets for OpenAI and other large AI developers. So far, Cursor's owner AnySphere isn't selling.