WEBVTT 00:00:00.000 --> 00:00:02.600 In just 30 days, the whole game for large language 00:00:02.600 --> 00:00:05.280 models got, well, it got flipped on its head. 00:00:05.500 --> 00:00:07.900 A major player, the one everyone thinks of as 00:00:07.900 --> 00:00:10.359 the leader, saw its market share on the web drop 00:00:10.359 --> 00:00:13.380 by a massive 19 percentage points. Yeah, it was 00:00:13.380 --> 00:00:16.179 a huge alarm bell, a real signal that the rules 00:00:16.179 --> 00:00:20.280 have changed and speed is just everything now. 00:00:20.460 --> 00:00:22.859 So today we're doing a deep dive into why that 00:00:22.859 --> 00:00:24.699 happened, both strategically and technically. 00:00:24.940 --> 00:00:26.980 We're going to look at what AI agents really 00:00:26.980 --> 00:00:29.879 need to survive this new phase. Welcome to the 00:00:29.879 --> 00:00:32.579 Deep Dive. We've gone through some key findings 00:00:32.579 --> 00:00:35.939 from the 2026 AI landscape and a few other reports 00:00:35.939 --> 00:00:37.820 to give you the knowledge you need. Our goal 00:00:37.820 --> 00:00:40.359 is pretty simple, help you navigate these incredibly 00:00:40.359 --> 00:00:43.079 fast shifts in the market and the tech. So our 00:00:43.079 --> 00:00:45.859 roadmap today starts with that market data, the 00:00:45.859 --> 00:00:48.420 big reversal, and why distribution is king again. 00:00:48.740 --> 00:00:50.979 Then we'll pivot to what's happening on the ground, 00:00:51.079 --> 00:00:52.859 you know, why businesses are moving from simple 00:00:52.859 --> 00:00:55.679 automation to building real AI agents. And finally, 00:00:55.679 --> 00:00:57.159 we're going to break down a brand new way of 00:00:57.159 --> 00:00:59.780 thinking about agent memory, sort of the brain 00:00:59.780 --> 00:01:02.299 architecture for the AI of the future. Okay, 00:01:02.340 --> 00:01:04.200 let's start with that market reversal. The numbers 00:01:04.200 --> 00:01:07.060 are pretty stark. Stark is a good word for it. 00:01:07.120 --> 00:01:09.340 I mean, the sources show the dominant player 00:01:10.039 --> 00:01:13.980 ChatGPT fell from 87 .2 % of public web traffic 00:01:13.980 --> 00:01:19.079 all the way down to 68 .0%. That 19 -point drop 00:01:19.079 --> 00:01:21.620 in one month. That's why you hear people calling 00:01:21.620 --> 00:01:24.450 it a code red. And at the exact same time, Gemini 00:01:24.450 --> 00:01:27.290 was just surging. They jumped from 13 .7 % up 00:01:27.290 --> 00:01:30.670 to 18 .2%. That's a huge gain in just four weeks. 00:01:30.870 --> 00:01:32.590 And what's really clear from the analysis is 00:01:32.590 --> 00:01:34.590 this wasn't really about one model suddenly becoming, 00:01:34.730 --> 00:01:37.049 you know, way better than the other. The reason 00:01:37.049 --> 00:01:39.790 Gemini gained so much so fast was just pure distribution. 00:01:40.129 --> 00:01:42.310 Distribution wins. It's a classic lesson, but 00:01:42.310 --> 00:01:45.489 the scale of it here is... Something else. Gemini 00:01:45.489 --> 00:01:46.989 is just showing up where people already are. 00:01:47.049 --> 00:01:49.069 There's no friction. Exactly. It's embedded right 00:01:49.069 --> 00:01:51.069 inside Google search. It's on every Android phone. 00:01:51.310 --> 00:01:53.750 It's in Chrome. The user doesn't have to consciously 00:01:53.750 --> 00:01:55.750 think, OK, I'm going to go to the AI now. It's 00:01:55.750 --> 00:01:58.349 just there, ambient. We should add some perspective 00:01:58.349 --> 00:02:01.230 here, though. ChatGPT is still the giant in the 00:02:01.230 --> 00:02:03.870 room. The denominator, I mean, the total number 00:02:03.870 --> 00:02:07.310 of people using generative AI is exploding. So 00:02:07.310 --> 00:02:10.050 losing market share doesn't automatically mean 00:02:10.050 --> 00:02:12.580 they're losing users. But it's definitely some 00:02:12.580 --> 00:02:15.060 serious pressure. Oh, for sure. And you can see 00:02:15.060 --> 00:02:17.020 that pressure elsewhere, too. Look at Microsoft 00:02:17.020 --> 00:02:20.620 Copilot. Its public web traffic actually dipped 00:02:20.620 --> 00:02:24.819 a little from 1 .5 % to 1 .2%. basically flat. 00:02:25.060 --> 00:02:28.159 Right. But the sources are pretty clear that 00:02:28.159 --> 00:02:31.139 the web data for Microsoft is misleading. That's 00:02:31.139 --> 00:02:34.240 the key. Most copilot use isn't on a public website. 00:02:34.439 --> 00:02:37.000 It's native. It's running inside Word, inside 00:02:37.000 --> 00:02:39.939 Excel, inside Teams. And that activity, you just 00:02:39.939 --> 00:02:41.819 can't see it in public metrics. It's real successes 00:02:41.819 --> 00:02:44.889 in the enterprise. So the path forward for ChatGPT 00:02:44.889 --> 00:02:47.030 seems pretty clear. They have to integrate more 00:02:47.030 --> 00:02:48.889 deeply. They have to live where their users are. 00:02:49.030 --> 00:02:51.389 Does this really mean that raw model power is 00:02:51.389 --> 00:02:53.430 starting to matter less than just being accessible? 00:02:53.710 --> 00:02:56.280 Utility often beats purity. Getting the model 00:02:56.280 --> 00:02:58.060 in front of the user is the real challenge now. 00:02:58.180 --> 00:03:00.379 Okay, let's shift from the market landscape to 00:03:00.379 --> 00:03:02.500 what the builders are actually doing, this pivot 00:03:02.500 --> 00:03:04.860 to agents. Yeah, we're seeing a big change in 00:03:04.860 --> 00:03:07.580 strategy. The sources point out that a lot of 00:03:07.580 --> 00:03:10.080 AI automation agencies are actually struggling. 00:03:10.379 --> 00:03:12.879 And it's because simple automation is basically 00:03:12.879 --> 00:03:15.439 a commodity now. It's cheap. Anyone can do it. 00:03:15.520 --> 00:03:17.520 So they're pivoting. They're moving up the value 00:03:17.520 --> 00:03:19.900 chain. Instead of just building simple bots, 00:03:20.219 --> 00:03:22.919 they're selling outcomes. They're doing AI audits. 00:03:22.919 --> 00:03:24.860 They're focusing on enterprise adoption. And 00:03:24.860 --> 00:03:27.699 that kind of pivot demands incredible speed from 00:03:27.699 --> 00:03:31.259 the top. AI CEOs today have to iterate constantly. 00:03:31.460 --> 00:03:34.280 They have to watch user signals like a hawk and 00:03:34.280 --> 00:03:37.680 build systems that compound in value. It's why 00:03:37.680 --> 00:03:39.639 the tech itself has to change. We're moving beyond 00:03:39.639 --> 00:03:42.300 that simple if this, then that logic. We're talking 00:03:42.300 --> 00:03:45.419 about real AI agents, digital workers that don't 00:03:45.419 --> 00:03:47.580 just follow a script. They make decisions. And 00:03:47.580 --> 00:03:49.439 this is where the old tools, you know, the Zapiers 00:03:49.439 --> 00:03:51.280 and makes of the world, they start to fall apart. 00:03:51.400 --> 00:03:54.199 They get really expensive, really fast, and they 00:03:54.199 --> 00:03:56.819 just weren't built for complex, multi -step AI 00:03:56.819 --> 00:03:59.840 reasoning. The sources highlight tools like NEN 00:03:59.840 --> 00:04:02.759 as being the choice for pros who need more power. 00:04:02.919 --> 00:04:06.960 The end goal is pretty wild. Building systems 00:04:06.960 --> 00:04:09.840 that run while you're asleep. Automating an entire 00:04:09.840 --> 00:04:12.620 YouTube content pipeline. Or managing customer 00:04:12.620 --> 00:04:14.979 support across five different channels. It's 00:04:14.979 --> 00:04:17.180 the next level of efficiency. And the big players 00:04:17.180 --> 00:04:19.579 know it. I mean, look at Meta. They just acquired 00:04:19.579 --> 00:04:22.060 Manus AI, a company known for building agents 00:04:22.060 --> 00:04:24.660 that were outperforming some of OpenAI's own 00:04:24.660 --> 00:04:26.860 research models. That tells you everything. The 00:04:26.860 --> 00:04:29.540 race is on for smarter agents. It's a huge step. 00:04:29.639 --> 00:04:33.120 Although, I'll admit, even with these new frameworks 00:04:33.120 --> 00:04:35.980 and tools, I still find myself wrestling with 00:04:35.980 --> 00:04:38.379 prompt design. You know, just getting an agent 00:04:38.379 --> 00:04:41.240 to maintain focus on a long task without getting 00:04:41.240 --> 00:04:43.759 sidetracked. It's a real challenge. That's a 00:04:43.759 --> 00:04:45.980 really important point. If we're struggling with 00:04:45.980 --> 00:04:47.819 it, what does that mean for the average user? 00:04:47.899 --> 00:04:49.540 Is that why we need a better way to think about 00:04:49.540 --> 00:04:52.199 memory? It suggests the basic structure for deep 00:04:52.199 --> 00:04:54.579 thinking just isn't there in the old tools. We 00:04:54.579 --> 00:04:56.860 need a better brain. Before we get into that 00:04:56.860 --> 00:04:58.620 new brain, let's just quickly touch on a few 00:04:58.620 --> 00:05:00.980 other market signals that show how intense this 00:05:00.980 --> 00:05:03.680 is. Sure. One thing that jumped out was a job 00:05:03.680 --> 00:05:07.879 posting from Sam Altman, a role with a $555 ,000 00:05:07.879 --> 00:05:10.699 salary. And the job was just to plan for advanced 00:05:10.699 --> 00:05:13.740 AI. It shows the scale of thinking required. 00:05:14.000 --> 00:05:17.279 Wow. That's not a developer role. That's a strategist. 00:05:17.399 --> 00:05:20.000 Whoa. Yeah. Imagine trying to scale a development 00:05:20.000 --> 00:05:23.319 team to handle a billion new complex AI queries 00:05:23.319 --> 00:05:26.120 every single day. The job is about infrastructure. 00:05:26.199 --> 00:05:28.910 It's about ethics. It's about risk. And on the 00:05:28.910 --> 00:05:30.829 creative side, we saw that seven -minute movie 00:05:30.829 --> 00:05:33.389 made entirely by one person with AI. It just 00:05:33.389 --> 00:05:35.470 went viral. It's blurring the lines completely 00:05:35.470 --> 00:05:38.449 between what a human can create versus a machine. 00:05:38.790 --> 00:05:40.930 Then you have the regulatory side, which is sending 00:05:40.930 --> 00:05:43.269 these mixed signals. China put out draft rules 00:05:43.269 --> 00:05:45.870 that would force AI apps to intervene if a user 00:05:45.870 --> 00:05:48.170 seems addicted. It's a huge signal that regulators 00:05:48.170 --> 00:05:50.689 are worried about agents becoming, well, too 00:05:50.689 --> 00:05:52.970 human. Yeah. That psychological pull is already 00:05:52.970 --> 00:05:55.199 on their radar. That sets the stage perfectly 00:05:55.199 --> 00:05:56.839 for the technical breakthrough we saw in the 00:05:56.839 --> 00:05:59.759 sources, the agent's brain. Right. And this is 00:05:59.759 --> 00:06:01.839 where it gets a little more academic. But it's 00:06:01.839 --> 00:06:04.060 so important for anyone building in this space. 00:06:04.439 --> 00:06:06.959 For years, we've thought about memory in simple 00:06:06.959 --> 00:06:09.800 terms, short term, long term, stuffing things 00:06:09.800 --> 00:06:12.860 into a context window or using RAG. Let's define 00:06:12.860 --> 00:06:16.120 AG quickly. It stands for Retrieval Augmented 00:06:16.120 --> 00:06:18.750 Generation. It's basically just a lookup system. 00:06:19.009 --> 00:06:21.790 The A .I. gets a question. It finds the relevant 00:06:21.790 --> 00:06:24.810 info in a database and uses that to form an answer. 00:06:24.930 --> 00:06:27.629 It keeps it grounded in facts. But the consensus 00:06:27.629 --> 00:06:30.709 now is that our rag, while it's useful, is just 00:06:30.709 --> 00:06:33.970 not enough. It's too passive for a true agent 00:06:33.970 --> 00:06:36.149 that needs to make decisions. And so this new 00:06:36.149 --> 00:06:39.470 paper lays out a full taxonomy, a kind of. a 00:06:39.470 --> 00:06:41.970 builder's checklist for agent memory it treats 00:06:41.970 --> 00:06:45.230 memory as its own complex system exactly they 00:06:45.230 --> 00:06:47.550 break it down into three lenses lens one is about 00:06:47.550 --> 00:06:50.050 the forms of memory what memory actually is okay 00:06:50.050 --> 00:06:51.769 so you have token level memory which is just 00:06:51.769 --> 00:06:53.910 that temporary space for the current chat then 00:06:53.910 --> 00:06:56.050 you have parametric memory that's the knowledge 00:06:56.050 --> 00:06:57.730 that's actually baked into the model's weights 00:06:57.730 --> 00:07:00.370 and the third form is what they call latent memory 00:07:00.370 --> 00:07:03.230 these are sort of hidden objects like embeddings 00:07:03.230 --> 00:07:05.310 that are created on the fly to help the agent 00:07:05.310 --> 00:07:08.089 keep track of things then you have lens two which 00:07:08.089 --> 00:07:10.120 is about functions, what the memory actually 00:07:10.120 --> 00:07:13.180 does. This starts with factual memory, ground 00:07:13.180 --> 00:07:15.819 truths, things that don't change. But the most 00:07:15.819 --> 00:07:18.339 important one here is experiential memory. This 00:07:18.339 --> 00:07:20.779 is where the agent actually learns. It's a log 00:07:20.779 --> 00:07:23.740 of everything it's done, its successes, its failures. 00:07:23.980 --> 00:07:26.459 It's how it gets better over time. And rounding 00:07:26.459 --> 00:07:29.439 it out is working memory, which is just a temporary 00:07:29.439 --> 00:07:31.420 scratch pad. It's how the agent keeps track of 00:07:31.420 --> 00:07:33.620 what it's doing right now in a long task. Now, 00:07:33.639 --> 00:07:35.800 the third lens is the big one, the real mental 00:07:35.800 --> 00:07:39.339 leap. Lens three is dynamics. This is about how 00:07:39.339 --> 00:07:41.920 memory changes and grows, and the paper calls 00:07:41.920 --> 00:07:44.360 it a control problem. And that's a really important 00:07:44.360 --> 00:07:46.600 phrase. It's not just about finding data. It's 00:07:46.600 --> 00:07:48.939 about actively managing memory, deciding what 00:07:48.939 --> 00:07:52.120 to keep, what to forget, how to learn. It's a 00:07:52.120 --> 00:07:55.300 strategy. It changes everything. It means we 00:07:55.300 --> 00:07:57.439 have to build memory systems like we're stacking 00:07:57.439 --> 00:08:00.699 intricate Lego blocks of data, not just pouring 00:08:00.699 --> 00:08:03.620 it all into one big bucket. So why is calling 00:08:03.620 --> 00:08:06.779 memory dynamics a control problem such a big 00:08:06.779 --> 00:08:08.720 cognitive leap here? What makes it so different? 00:08:08.860 --> 00:08:11.519 It forces builders to actively manage how agents 00:08:11.519 --> 00:08:14.480 learn and adapt. We're moving way beyond simple 00:08:14.480 --> 00:08:16.560 retrieval. So let's bring this all together. 00:08:16.660 --> 00:08:18.500 The big ideas from what we've looked at today. 00:08:18.699 --> 00:08:21.139 That market share shift proves one thing above 00:08:21.139 --> 00:08:24.569 all else. Distribution is everything. If you're 00:08:24.569 --> 00:08:26.250 not where the user is, you're going to lose. 00:08:26.449 --> 00:08:28.709 And that pressure is what's driving the need 00:08:28.709 --> 00:08:31.750 for better agents. Simple automation tools are 00:08:31.750 --> 00:08:34.889 out. The future demands agents with these sophisticated, 00:08:35.230 --> 00:08:38.029 structured memory systems built using that three 00:08:38.029 --> 00:08:40.509 lens taxonomy we just talked about. To stay relevant 00:08:40.509 --> 00:08:42.409 or to build the kind of services that command 00:08:42.409 --> 00:08:45.090 those half a million dollar salaries, AI needs 00:08:45.090 --> 00:08:47.230 a much better way to remember and learn from 00:08:47.230 --> 00:08:49.210 experience. And this all comes back around to 00:08:49.210 --> 00:08:52.169 the human side of things. get better and better 00:08:52.169 --> 00:08:54.889 as they feel more human, we see regulators starting 00:08:54.889 --> 00:08:57.049 to step in, like with that China report. Which 00:08:57.049 --> 00:08:59.450 leaves a final provocative thought for you to 00:08:59.450 --> 00:09:02.289 think about. If AI agents are rapidly evolving 00:09:02.289 --> 00:09:04.970 their memory to mimic our own complexity, especially 00:09:04.970 --> 00:09:08.889 that experiential memory, what happens when that 00:09:08.889 --> 00:09:10.970 becomes the global standard? What does it mean 00:09:10.970 --> 00:09:13.330 for you when you interact with an AI that remembers 00:09:13.330 --> 00:09:15.970 every conversation, learns from it, and adapts 00:09:15.970 --> 00:09:18.029 its behavior just for you? We really appreciate 00:09:18.029 --> 00:09:20.230 you sharing your sources for this deep dive. 00:09:20.759 --> 00:09:21.379 Until next time.