WEBVTT 00:00:00.000 --> 00:00:02.620 Imagine you walk into a Ferrari dealership. You 00:00:02.620 --> 00:00:04.459 point at their top of the line million dollar 00:00:04.459 --> 00:00:06.919 hypercar. The one on the pedestal, yeah. Exactly. 00:00:07.639 --> 00:00:10.599 And the dealer just looks at you and says, actually, 00:00:10.800 --> 00:00:13.640 we're going to give you that same engine, but 00:00:13.640 --> 00:00:16.039 we're putting in a sedan and the gas is 67 % 00:00:16.039 --> 00:00:18.640 cheaper. And the sedan somehow gets better mileage. 00:00:19.000 --> 00:00:21.660 That is, I mean, effectively what just happened 00:00:21.660 --> 00:00:24.179 this week. Welcome back to the Deep Dive. It's 00:00:24.179 --> 00:00:26.140 good to be here. Today, we're not just talking 00:00:26.140 --> 00:00:28.539 about new features or specs. We are tracking 00:00:28.539 --> 00:00:31.780 a really massive shift in the industry. This 00:00:31.780 --> 00:00:34.899 pivot from just raw power to what you might call 00:00:34.899 --> 00:00:37.219 radical efficiency. It's the moment the tech 00:00:37.219 --> 00:00:39.759 really matures, you know. We're moving from the 00:00:39.759 --> 00:00:43.020 wow phase to the ROI phase. Right. Can it do 00:00:43.020 --> 00:00:45.659 the job cheaply? And can it do it without hallucinating 00:00:45.659 --> 00:00:49.039 halfway through? That's the key. Lay out the 00:00:49.039 --> 00:00:51.259 map for us today. How are we going to tackle 00:00:51.259 --> 00:00:53.479 this? All right. We've got a heavy lineup. First, 00:00:53.579 --> 00:00:56.359 we're starting with Anthropic's new clawed updates 00:00:56.359 --> 00:01:00.159 on it, 4 .6. It feels less like an update and 00:01:00.159 --> 00:01:03.340 more like a declaration of war on the whole flagship 00:01:03.340 --> 00:01:06.739 model category. Then we've got a lightning round 00:01:06.739 --> 00:01:10.299 NPR suing Google over a voice that sounds a little 00:01:10.299 --> 00:01:13.879 too familiar. Apple's hardware roadmap for 2027. 00:01:13.959 --> 00:01:16.920 All the big stuff. And we end on the underdog 00:01:16.920 --> 00:01:19.909 story. The biggest story of the week, in my opinion. 00:01:20.030 --> 00:01:23.670 A recruiting startup in China, not a tech giant, 00:01:23.790 --> 00:01:27.769 a recruiting firm, just released a tiny AI model 00:01:27.769 --> 00:01:30.769 that is absolutely beating the giants at their 00:01:30.769 --> 00:01:33.150 own game. I saw those numbers. It doesn't seem 00:01:33.150 --> 00:01:35.230 mathematically possible, but we'll get there. 00:01:35.349 --> 00:01:38.730 So let's start with the big dog. Anthropic, Claude 00:01:38.730 --> 00:01:42.329 Son at 4 .6. Right. Normally a .6 release is, 00:01:42.390 --> 00:01:45.189 you know, a patch, some bug fixes. But these 00:01:45.189 --> 00:01:48.140 benchmarks feel... Very aggressive. The whole 00:01:48.140 --> 00:01:50.140 headline here is the price to performance ratio. 00:01:50.519 --> 00:01:52.560 Anthropic is saying this new Sonnet delivers 00:01:52.560 --> 00:01:55.480 near Opus performance. And Opus is their Einstein 00:01:55.480 --> 00:01:58.500 model, the expensive heavy lifter. Sonnet's the 00:01:58.500 --> 00:02:00.560 mid -tier. Correct. But the pricing didn't change 00:02:00.560 --> 00:02:02.299 at all. So the pricing didn't move up. No, it 00:02:02.299 --> 00:02:04.760 stayed flat. $3 per million input tokens, 15 00:02:04.760 --> 00:02:08.180 for output. Opus is 5 and 25. OK, let's just 00:02:08.180 --> 00:02:10.180 sit with that for a second. If I'm getting near 00:02:10.180 --> 00:02:13.340 flagship intelligence for those prices, doing 00:02:13.340 --> 00:02:18.050 the math. That's roughly 67 % more tokens for 00:02:18.050 --> 00:02:20.889 every dollar I spend. Which is huge. But think 00:02:20.889 --> 00:02:23.030 about what that lets you do. As a developer, 00:02:23.189 --> 00:02:26.870 you can now afford to have the AI think longer, 00:02:27.150 --> 00:02:30.009 maybe check its own work three times for the 00:02:30.009 --> 00:02:32.550 same cost as one answer from the old flagship. 00:02:32.930 --> 00:02:34.710 It totally changes the economics of reliability. 00:02:35.229 --> 00:02:38.169 It does. It turns intelligence into a commodity 00:02:38.169 --> 00:02:40.590 you can afford to waste a little bit of. And 00:02:40.590 --> 00:02:42.210 it seems like the users are really noticing. 00:02:42.289 --> 00:02:44.900 I was looking at the sentiment analysis. 70 % 00:02:44.900 --> 00:02:47.159 of users prefer this new Sonnet over the last 00:02:47.159 --> 00:02:50.319 one. Right. But the stat that blew my mind, 59 00:02:50.319 --> 00:02:53.460 % prefer it over the previous Opus. That's the 00:02:53.460 --> 00:02:55.780 cannibalization metric right there. When your 00:02:55.780 --> 00:02:58.099 mid -tier model beats your old flagship, you've 00:02:58.099 --> 00:03:00.199 successfully raised the floor for the entire 00:03:00.199 --> 00:03:02.960 industry. But why the preference? Is it just 00:03:02.960 --> 00:03:06.639 about speed? Because usually better means smarter, 00:03:06.819 --> 00:03:09.620 not just faster. It's bloat reduction. That's 00:03:09.620 --> 00:03:12.580 what users are saying. Fewer hallucinations and, 00:03:12.740 --> 00:03:15.479 this is interesting, less over -engineering in 00:03:15.479 --> 00:03:17.500 the answers. Oh, I know exactly what that means. 00:03:17.620 --> 00:03:20.020 You ask for a simple two -line Python script 00:03:20.020 --> 00:03:22.379 and it gives you a lecture on coding ethics and 00:03:22.379 --> 00:03:24.319 a five -paragraph history of the language first. 00:03:24.560 --> 00:03:28.520 Exactly that. Sonnet 4 .6 just cuts the waffle. 00:03:29.000 --> 00:03:31.900 It's pragmatic. In fact, on the honest world 00:03:31.900 --> 00:03:35.360 benchmark, which tests how well an AI can navigate 00:03:35.360 --> 00:03:38.379 computer apps. Like a human would. Yes. It's 00:03:38.379 --> 00:03:40.939 matching the current opus. It's a worker bee 00:03:40.939 --> 00:03:43.659 with the brain of a queen. Speaking of worker 00:03:43.659 --> 00:03:45.979 bees, there's a new feature they rolled out called 00:03:45.979 --> 00:03:48.919 context compaction. This sounds technical, but 00:03:48.919 --> 00:03:50.439 I have a feeling it's going to matter. It really 00:03:50.439 --> 00:03:52.800 might. Do you ever have that issue where you're 00:03:52.800 --> 00:03:54.819 in a long chat with an AI, maybe working on a 00:03:54.819 --> 00:03:57.280 project, and by message 50, it starts forgetting 00:03:57.280 --> 00:03:59.740 what you told it in message one? Prompt drift. 00:04:00.740 --> 00:04:02.560 I'll be honest. I struggle with this constantly. 00:04:02.699 --> 00:04:05.840 I spend half my time reminding the bot, hey, 00:04:05.879 --> 00:04:08.620 remember, we aren't using that library or don't 00:04:08.620 --> 00:04:10.819 use that tone. It feels like the model gets dementia. 00:04:11.400 --> 00:04:14.620 Right. And that happens because the context window, 00:04:14.759 --> 00:04:17.600 the AI's short -term memory, it gets filled up 00:04:17.600 --> 00:04:19.459 with every single word. It's trying to remember 00:04:19.459 --> 00:04:23.699 every typo from the last four hours. And eventually 00:04:23.699 --> 00:04:25.699 it just collapses under its own weight. Exactly. 00:04:25.759 --> 00:04:28.360 So compaction fixes that. Think of it like a 00:04:28.360 --> 00:04:31.180 court stenographer versus a really good executive 00:04:31.180 --> 00:04:33.379 assistant. Okay. The stenographer writes down 00:04:33.379 --> 00:04:36.550 every single word verbatim. The assistant listens 00:04:36.550 --> 00:04:39.610 for 30 minutes, shreds the full transcript, and 00:04:39.610 --> 00:04:41.629 then writes a perfect three -bullet summary of 00:04:41.629 --> 00:04:44.389 what was actually decided. So the AI is editing 00:04:44.389 --> 00:04:47.829 its own memory in real time. Yes. It summarizes 00:04:47.829 --> 00:04:50.029 the older parts of the conversation. It keeps 00:04:50.029 --> 00:04:52.250 the intent and the decisions, but it throws away 00:04:52.250 --> 00:04:55.089 all the verbatim fluff. This means you can have 00:04:55.089 --> 00:04:57.629 a never -ending conversation. That is huge for 00:04:57.629 --> 00:05:00.540 autonomous agents. If I have a bat monitoring 00:05:00.540 --> 00:05:03.379 my emails for a week, it has to remember the 00:05:03.379 --> 00:05:05.740 rules I set on Monday without crashing on Friday. 00:05:05.879 --> 00:05:08.060 Exactly. It keeps the context clean. So let me 00:05:08.060 --> 00:05:10.839 ask you this. If Sonnet is this good and this 00:05:10.839 --> 00:05:13.500 cheap, does the flagship model category just 00:05:13.500 --> 00:05:17.040 die? Why would I ever pay for Opus or GPT -5? 00:05:17.259 --> 00:05:20.740 It doesn't die, but it forces flagships to justify 00:05:20.740 --> 00:05:23.160 their premium price. They can't just be good 00:05:23.160 --> 00:05:25.459 anymore. They have to be geniuses. Raising the 00:05:25.459 --> 00:05:28.889 floor forces the ceiling up. Okay, let's pivot. 00:05:29.230 --> 00:05:32.410 The industry never sleeps. Let's hit the lightning 00:05:32.410 --> 00:05:35.589 round. Let's do it. First up, a legal battle 00:05:35.589 --> 00:05:37.529 that feels like it's straight out of a sci -fi 00:05:37.529 --> 00:05:41.449 novel. NPR versus Google. David Green, the former 00:05:41.449 --> 00:05:44.300 host, is suing. He says Google's notebook LM 00:05:44.300 --> 00:05:46.920 tool is ripping off his voice. This is fascinating 00:05:46.920 --> 00:05:49.300 because Google's defense is pretty standard. 00:05:49.519 --> 00:05:52.040 We hired a paid actor, and they probably did, 00:05:52.240 --> 00:05:54.339 but Green isn't arguing about the voice print. 00:05:54.579 --> 00:05:56.579 He's arguing about the cadence. Exactly, the 00:05:56.579 --> 00:05:59.839 public radio voice. The pauses, the intonation, 00:05:59.939 --> 00:06:02.060 the way they sort of lean into the microphone. 00:06:02.160 --> 00:06:05.680 It raises this huge legal gray area. Can you 00:06:05.680 --> 00:06:07.560 copyright a pause? Can you copyright a vibe? 00:06:07.899 --> 00:06:10.199 Right, but I think there's a real identity theft 00:06:10.199 --> 00:06:12.180 angle here. If I listen to it and think it's... 00:06:12.199 --> 00:06:14.160 David Green, isn't that the real problem? That's 00:06:14.160 --> 00:06:17.120 the danger zone. Legally, Cadence isn't copyrightable, 00:06:17.120 --> 00:06:19.579 not yet. But if a jury believes Google told the 00:06:19.579 --> 00:06:22.740 AI to sound like David Green, that gets into 00:06:22.740 --> 00:06:25.439 right of publicity. It's the same thing Scarlett 00:06:25.439 --> 00:06:27.800 Johansson dealt with. It's not about the raw 00:06:27.800 --> 00:06:30.899 data. It's about the intent to mimic. So regarding 00:06:30.899 --> 00:06:34.220 the NPR lawsuit, where is the line between inspiration 00:06:34.220 --> 00:06:37.959 and theft? in AI voices. Its blurry cadence isn't 00:06:37.959 --> 00:06:41.139 copyrightable, but identity theft is a real legal 00:06:41.139 --> 00:06:44.879 risk. Okay, moving from legal battles to actual 00:06:44.879 --> 00:06:48.300 tools. A new tool called Manos just launched 00:06:48.300 --> 00:06:51.790 agents inside Telegram. This is wild. It's OpenClaw 00:06:51.790 --> 00:06:54.170 -style agents, basically autonomous research 00:06:54.170 --> 00:06:56.449 bots, but they live inside your chat app. So 00:06:56.449 --> 00:06:58.889 no coding needed? None. You just ask it to research 00:06:58.889 --> 00:07:01.490 a topic or process some data, create a PDF, and 00:07:01.490 --> 00:07:03.870 it just does it right there in the chat. It's 00:07:03.870 --> 00:07:05.930 putting a research assistant in the app we use 00:07:05.930 --> 00:07:07.750 for memes. Okay, let's look further down the 00:07:07.750 --> 00:07:10.560 timeline. Apple. Rumors are swirling about an 00:07:10.560 --> 00:07:13.439 acceleration on three wearables. Yes, the roadmap 00:07:13.439 --> 00:07:16.240 seems to be pointing toward late 2027. We're 00:07:16.240 --> 00:07:19.259 hearing about an AI pin, smart glasses, and this 00:07:19.259 --> 00:07:21.660 is the big one, AirPods with really deep Siri 00:07:21.660 --> 00:07:24.639 integration. 2027 sounds so far away, but in 00:07:24.639 --> 00:07:26.860 hardware terms, that's basically tomorrow. It 00:07:26.860 --> 00:07:30.420 is. If production starts in December 2027, we 00:07:30.420 --> 00:07:32.879 are looking at a very, very interesting holiday 00:07:32.879 --> 00:07:36.110 season that year. Apple is clearly betting the 00:07:36.110 --> 00:07:38.689 future isn't just a phone it's AI whispering 00:07:38.689 --> 00:07:42.240 in your ear and overlaying your vision meanwhile 00:07:42.240 --> 00:07:45.800 on the infrastructure side open ai just introduced 00:07:45.800 --> 00:07:48.560 lockdown mode this is for the enterprise crowd 00:07:48.560 --> 00:07:52.240 right totally the security focus teams it blocks 00:07:52.240 --> 00:07:55.220 risky tools cuts off live web access it's like 00:07:55.220 --> 00:07:57.839 putting the ai in a clean room so it can't leak 00:07:57.839 --> 00:08:00.560 secrets or download malware in meta they're just 00:08:00.560 --> 00:08:02.519 buying up all the chips all of them millions 00:08:02.519 --> 00:08:06.019 of nvidia chips grace cpus the next gen viewer 00:08:06.019 --> 00:08:08.899 rubin systems they are building a massive war 00:08:09.040 --> 00:08:12.199 chest of compute. Speaking of war chests, Blackstone 00:08:12.199 --> 00:08:15.920 just led a $1 .2 billion round for a company 00:08:15.920 --> 00:08:19.060 called Nasa. An Indian AI data center startup. 00:08:19.160 --> 00:08:21.240 This is a huge signal that the demand for AI 00:08:21.240 --> 00:08:23.220 infrastructure is just exploding in emerging 00:08:23.220 --> 00:08:25.240 markets. It's not just a Silicon Valley game 00:08:25.240 --> 00:08:27.480 anymore. Not at all. The physical backbone of 00:08:27.480 --> 00:08:30.199 AI is going global. So when you zoom out on all 00:08:30.199 --> 00:08:33.440 this, lawsuits over voices, agents in Telegram, 00:08:33.740 --> 00:08:37.539 AI in our glasses, data centers in India, what's 00:08:37.539 --> 00:08:40.129 the big picture? The big picture is AI becoming 00:08:40.129 --> 00:08:44.149 more human, more accessible, and just more ubiquitous. 00:08:44.389 --> 00:08:47.509 The gap between computer and partner is really 00:08:47.509 --> 00:08:49.909 starting to vanish. And that transition needs 00:08:49.909 --> 00:08:52.750 new tools. It's not just about the big models. 00:08:52.809 --> 00:08:54.950 It's about that layer in between. For sure. We've 00:08:54.950 --> 00:08:56.429 seen a few interesting ones pop up. We can do 00:08:56.429 --> 00:08:58.840 a quick fire on these. All right. First, figure 00:08:58.840 --> 00:09:01.320 AI. This one's for the designers. It maps out 00:09:01.320 --> 00:09:04.279 user flows, it spots edge cases, and it can even 00:09:04.279 --> 00:09:06.860 build out A -B variations for you. It's automating 00:09:06.860 --> 00:09:08.759 the logic part of design. Okay, then there's 00:09:08.759 --> 00:09:11.159 layers. That's a growth planning tool. It generates 00:09:11.159 --> 00:09:14.500 content, social posts, ads, basically a marketing 00:09:14.500 --> 00:09:17.860 agency in a box. And boost .spacev5. This one 00:09:17.860 --> 00:09:20.580 calls itself a persistent context layer. Think 00:09:20.580 --> 00:09:22.360 of it as the glue that holds all your different 00:09:22.360 --> 00:09:24.220 AIs together so they can actually talk to each 00:09:24.220 --> 00:09:28.360 other. And finally... QN3 .5, an open -weight 00:09:28.360 --> 00:09:31.179 vision language model. The key stat here is just 00:09:31.179 --> 00:09:33.720 fascinating. It delivers the capabilities of 00:09:33.720 --> 00:09:37.340 a nearly 400 billion parameter model. It's huge. 00:09:37.679 --> 00:09:40.139 But with the inference speed of a 17 billion 00:09:40.139 --> 00:09:43.460 parameter model. Okay, hold on. Inference speed. 00:09:43.539 --> 00:09:46.080 Let's define that. Simply put, it's how fast 00:09:46.080 --> 00:09:49.299 the AI thinks and replies. It's the delay between 00:09:49.299 --> 00:09:52.320 you hitting enter and getting an answer. Usually, 00:09:52.360 --> 00:09:55.960 smart models are slow and heavy. When 3 .5 is 00:09:55.960 --> 00:09:58.679 smart, but fast. So with tools like figure and 00:09:58.679 --> 00:10:01.259 layers, are we automating the creative director? 00:10:01.500 --> 00:10:03.200 We're automating the grunt work. The director 00:10:03.200 --> 00:10:06.279 still needs to choose the vision. Sponsor. Okay, 00:10:06.340 --> 00:10:08.379 we've talked about the giant centropic, Google, 00:10:08.620 --> 00:10:10.519 Apple. But this is where the story gets really 00:10:10.519 --> 00:10:12.779 interesting. The biggest surprise of the week, 00:10:12.779 --> 00:10:15.539 for me, didn't come from Silicon Valley. It came 00:10:15.539 --> 00:10:18.039 from a recruiting startup in China. This is the 00:10:18.039 --> 00:10:19.740 story everyone should be paying attention to. 00:10:20.000 --> 00:10:23.659 The lab is called Nanbije LLM Lab, and they released 00:10:23.659 --> 00:10:27.820 a model called Nanbije 4 .13b. 3b, as in 3 billion 00:10:27.820 --> 00:10:31.000 parameters. In the world of LLMs, that is. Tiny. 00:10:31.059 --> 00:10:32.860 That's pocket -sized. It's incredibly small. 00:10:33.259 --> 00:10:36.539 For context, GPT -4 is rumored to be in the trillions 00:10:36.539 --> 00:10:41.080 of parameters. Lama 3 is 70 or 400 billion. 3 00:10:41.080 --> 00:10:42.440 billion is something you could theoretically 00:10:42.440 --> 00:10:45.159 run on a high -end phone. But it's crushing benchmarks, 00:10:45.340 --> 00:10:47.460 despite being so small. It is playing David to 00:10:47.460 --> 00:10:51.750 the industry's Goliath. On Arena Hard V2, a really 00:10:51.750 --> 00:10:55.470 tough stress test, it scored a 73 .2. Okay, numbers 00:10:55.470 --> 00:10:57.049 are numbers. Let's talk about coding. That's 00:10:57.049 --> 00:10:59.210 where the rubber meets the road. This is the 00:10:59.210 --> 00:11:02.340 shocker. On LeetCode Weekly Contests... These 00:11:02.340 --> 00:11:04.700 are real coding challenges that actual humans 00:11:04.700 --> 00:11:09.100 take to get jobs. Nan Beige reported an 85 .0 00:11:09.100 --> 00:11:13.659 % pass rate. 85%. Now compare that to QEM332B, 00:11:13.919 --> 00:11:17.279 a model 10 times its size, which scored 50%. 00:11:17.279 --> 00:11:20.740 Wait, hold on. A 3 billion parameter model? beat 00:11:20.740 --> 00:11:24.519 a 32 billion parameter model by 35 percentage 00:11:24.519 --> 00:11:26.860 points yes that shouldn't be possible that's 00:11:26.860 --> 00:11:29.080 like a go -kart beating a formula one car imagine 00:11:29.080 --> 00:11:31.860 fitting a grandmaster chess player in your pocket 00:11:31.860 --> 00:11:34.299 that is what this really represents it just defies 00:11:34.299 --> 00:11:36.500 the logic that bigger is always better so how 00:11:36.500 --> 00:11:38.500 did they do it what's the secret sauce here it's 00:11:38.500 --> 00:11:40.600 all about the methodology they didn't just throw 00:11:40.600 --> 00:11:42.620 more data at it they trained it smarter they 00:11:42.620 --> 00:11:45.159 used an upgraded fine tuning mix but the real 00:11:45.159 --> 00:11:48.139 magic is in the reinforcement learning or rl 00:11:48.139 --> 00:11:50.580 okay break that down for us They used a two -stage 00:11:50.580 --> 00:11:54.669 process. First is pointwise RL. The model generates 00:11:54.669 --> 00:11:56.809 eight different answers for a single prompt, 00:11:56.929 --> 00:11:59.529 and a reward model scores each one. Right. And 00:11:59.529 --> 00:12:02.570 they optimize this using something called GRPO 00:12:02.570 --> 00:12:05.970 group relative policy optimization. Jargon alert. 00:12:06.690 --> 00:12:09.509 GRPO. Think of it as a way to reduce repetitive 00:12:09.509 --> 00:12:12.309 errors. It's like a teacher not just marking 00:12:12.309 --> 00:12:15.029 an answer wrong, but showing the student specifically 00:12:15.029 --> 00:12:17.610 where they started to ramble or make a mistake 00:12:17.610 --> 00:12:20.190 so they don't do it again. It tightens the logic. 00:12:20.350 --> 00:12:23.639 To the second stage. That's pairwise RL. They 00:12:23.639 --> 00:12:25.799 take a strong response and a weak response and 00:12:25.799 --> 00:12:28.379 force the model to compare them. It uses something 00:12:28.379 --> 00:12:31.620 called a swap consistency regularizer. Okay, 00:12:31.720 --> 00:12:34.059 simpler analogy. It's like wine tasting. It's 00:12:34.059 --> 00:12:35.919 one thing to say this wine is good. It's much 00:12:35.919 --> 00:12:38.840 harder and more educational to taste two wines 00:12:38.840 --> 00:12:41.080 side by side and articulate why one is better. 00:12:41.220 --> 00:12:43.899 It just sharpens the model's judgment. So it's 00:12:43.899 --> 00:12:46.200 teaching the AI to recognize quality, not just 00:12:46.200 --> 00:12:49.220 generate text. Exactly. And because of this intense 00:12:49.220 --> 00:12:52.159 training, this tiny model supports up to 250... 00:12:52.460 --> 00:12:55.980 56 ,000 tokens of context. It can do deep search 00:12:55.980 --> 00:12:58.440 with hundreds of tool calls. It's a real agent. 00:12:58.639 --> 00:13:00.940 Does this prove that training technique matters 00:13:00.940 --> 00:13:04.500 more than raw parameter size? Absolutely. Smart 00:13:04.500 --> 00:13:07.179 training beats brute force computing power every 00:13:07.179 --> 00:13:09.559 time. And this really brings us back to the big 00:13:09.559 --> 00:13:12.480 idea of this deep dive. It all circles back to 00:13:12.480 --> 00:13:15.620 efficiency. We started with Cloud Sonnet, which 00:13:15.620 --> 00:13:18.220 gives us flagship level power. for a fraction 00:13:18.220 --> 00:13:20.940 of the price. Then we saw tools like Quinn and, 00:13:21.039 --> 00:13:23.299 of course, Nan Beige proving that smaller models 00:13:23.299 --> 00:13:25.419 are becoming incredibly smart through better 00:13:25.419 --> 00:13:27.820 architecture, not just bigger data centers. The 00:13:27.820 --> 00:13:30.220 hardware from Apple and Meta is ramping up to 00:13:30.220 --> 00:13:33.080 support this whole ecosystem. But the software 00:13:33.080 --> 00:13:35.159 itself is actually getting leaner. We're not 00:13:35.159 --> 00:13:37.159 just building bigger brains anymore. We're building 00:13:37.159 --> 00:13:40.039 more efficient ones. And that's crucial. If we 00:13:40.039 --> 00:13:42.899 want AI everywhere in our glasses, our chats, 00:13:43.139 --> 00:13:45.320 our phones, it can't rely on a massive server 00:13:45.320 --> 00:13:48.059 farm for every single thought. It has to be efficient. 00:13:48.220 --> 00:13:50.480 It has to be. Before we go, I just want to leave 00:13:50.480 --> 00:13:53.139 you with one final thought. That Nan Beige model, 00:13:53.320 --> 00:13:56.779 the world -beating, code -crushing AI, it came 00:13:56.779 --> 00:13:59.399 from a recruiting startup. Not Google, not OpenAI. 00:13:59.860 --> 00:14:03.100 A company that helps people find jobs. If a non 00:14:03.100 --> 00:14:05.139 -AI native company can build a state -of -the 00:14:05.139 --> 00:14:07.240 -art model to solve their specific problems, 00:14:07.419 --> 00:14:10.620 like coding and hiring, what happens when every 00:14:10.620 --> 00:14:13.289 company starts doing that? A law firm building 00:14:13.289 --> 00:14:15.850 the best legal model. A hospital building the 00:14:15.850 --> 00:14:18.769 best diagnostic model. It means the era of renting 00:14:18.769 --> 00:14:21.850 these giant general purpose brains from a handful 00:14:21.850 --> 00:14:24.649 of companies might be ending. We could be moving 00:14:24.649 --> 00:14:27.289 toward a world of specialized pocket -sized experts 00:14:27.289 --> 00:14:29.669 that live on our own devices. It's a profound 00:14:29.669 --> 00:14:32.389 shift. It's a future I'm very curious to see 00:14:32.389 --> 00:14:34.289 unfold. If you want to check out the links to 00:14:34.289 --> 00:14:36.490 the new Claude features or download that Nambige 00:14:36.490 --> 00:14:38.570 model yourself, check out the show notes. And 00:14:38.570 --> 00:14:40.750 give that context compaction a try. It might 00:14:40.750 --> 00:14:42.909 just save your sanity on a long project. Until 00:14:42.909 --> 00:14:45.429 next time, stay curious. Stay curious.