WEBVTT 00:00:00.000 --> 00:00:03.720 It is February 6th, 2026. If you were online 00:00:03.720 --> 00:00:06.759 yesterday, you saw something that looked like 00:00:06.759 --> 00:00:10.019 a car crash, a Super Bowl halftime show, and 00:00:10.019 --> 00:00:12.460 a scientific breakthrough all happening at once. 00:00:12.560 --> 00:00:16.100 It was a lot. In just 24 hours, we got a marketing 00:00:16.100 --> 00:00:19.300 grenade during the game, two massive AI models 00:00:19.300 --> 00:00:21.440 dropped within, what, minutes of each other, 00:00:21.559 --> 00:00:25.440 and a CEO meltdown online. It really was Kendrick 00:00:25.440 --> 00:00:27.910 versus Drake, but for... you know for code it 00:00:27.910 --> 00:00:30.210 was absolute cinema it was but if you strip away 00:00:30.210 --> 00:00:32.969 the drama and the memes and all the theatrics 00:00:32.969 --> 00:00:35.890 the actual tectonic plates of the tech industry 00:00:35.890 --> 00:00:38.609 shifted yesterday the tools we were all using 00:00:38.609 --> 00:00:41.369 on wednesday are now basically antiques that 00:00:41.369 --> 00:00:43.450 is what i want to figure out today welcome to 00:00:43.450 --> 00:00:45.289 the deep dive i want to slow down and really 00:00:45.289 --> 00:00:48.670 look at this 2026 ai showdown not just the noise 00:00:48.670 --> 00:00:51.750 but The signal. And the signal is getting harder 00:00:51.750 --> 00:00:53.509 to hear. If you were scrolling yesterday, you'd 00:00:53.509 --> 00:00:55.490 think Anthropic had, I don't know, effectively 00:00:55.490 --> 00:00:57.990 killed open AI. The whole vibe shift was palpable. 00:00:58.189 --> 00:01:00.210 But the reality is. The reality is, well, it's 00:01:00.210 --> 00:01:02.770 messy. So help me map this out. We're not just 00:01:02.770 --> 00:01:04.450 going to read spec sheets today. I want to understand 00:01:04.450 --> 00:01:06.769 the human behavior here. Right. I think we need 00:01:06.769 --> 00:01:10.049 to look at four distinct layers. First, the business 00:01:10.049 --> 00:01:12.870 reality, who is actually winning versus who feels 00:01:12.870 --> 00:01:16.849 like they are winning. Okay. Second, the psychological 00:01:16.849 --> 00:01:20.510 warfare of that Super Bowl ad. Third, the hardware 00:01:20.510 --> 00:01:25.810 itself, Claude Opus 4 .6 versus GPT 5 .3. And 00:01:25.810 --> 00:01:28.590 finally, the coffee shop test, which tells us 00:01:28.590 --> 00:01:30.989 a whole lot about where design is going. The 00:01:30.989 --> 00:01:33.969 coffee shop test results were... Honestly, they 00:01:33.969 --> 00:01:36.370 were strange. They were hilarious. Yeah. And 00:01:36.370 --> 00:01:38.189 very revealing. Oh, okay. So let's start with 00:01:38.189 --> 00:01:40.150 that reality check. Because if you exist in the 00:01:40.150 --> 00:01:42.090 tech bubble or you just read developer forums, 00:01:42.370 --> 00:01:44.730 the narrative is that Claude is king. It feels 00:01:44.730 --> 00:01:46.670 like a total takeover. That is the echo chamber 00:01:46.670 --> 00:01:48.689 effect, right? Yeah. I think we need to distinguish 00:01:48.689 --> 00:01:52.230 between mind share and market share. Because 00:01:52.230 --> 00:01:54.730 if you look at the actual data from mid -2025, 00:01:54.890 --> 00:01:58.329 the gap is... It's staggering. How big of a gap 00:01:58.329 --> 00:02:00.750 are we talking about? It's a chasm. I mean, OpenAI's 00:02:00.750 --> 00:02:03.969 chat GPT has roughly 415 million monthly users. 00:02:04.209 --> 00:02:06.629 Wow. They are entrenched in 92 % of the Fortune 00:02:06.629 --> 00:02:09.150 500. It is the default utility, like Google. 00:02:09.229 --> 00:02:12.210 And Anthropic. About 15 .5 million. Wait, 15. 00:02:12.449 --> 00:02:16.330 15 million. That is a 27x difference. Right. 00:02:16.370 --> 00:02:20.129 It's not even close. Even niche tools like Perplexity 00:02:20.129 --> 00:02:22.710 and DeepSeek have more daily active users than 00:02:22.710 --> 00:02:25.939 Claude. So. So when you see Anthropic coming 00:02:25.939 --> 00:02:28.020 out swinging like this, you have to frame it 00:02:28.020 --> 00:02:30.080 correctly. This isn't a heavyweight title fight 00:02:30.080 --> 00:02:33.099 between equals. No. This is a challenger walking 00:02:33.099 --> 00:02:35.479 into the champion's gym, where the champion owns 00:02:35.479 --> 00:02:38.000 the building, and trying to start a riot. That 00:02:38.000 --> 00:02:40.159 reframes everything. It feels less like a rivalry 00:02:40.159 --> 00:02:43.460 and more like... asymmetric warfare. They have 00:02:43.460 --> 00:02:46.120 to be loud because they are small. Exactly. OpenAI 00:02:46.120 --> 00:02:48.460 has the data moat. They have the users, the history, 00:02:48.599 --> 00:02:51.539 the enterprise contracts. Anthropic has the vibe. 00:02:51.659 --> 00:02:54.300 But vibe doesn't pay for a billion dollar server 00:02:54.300 --> 00:02:56.680 clusters. Right. They need the world to believe 00:02:56.680 --> 00:02:58.819 they are technically superior to keep their valuation 00:02:58.819 --> 00:03:01.000 high, even if they aren't the market leader. 00:03:01.159 --> 00:03:05.020 So if the user gap is that nearly half a billion 00:03:05.020 --> 00:03:08.340 against 15 million, does the hype actually matter? 00:03:08.580 --> 00:03:11.250 Not for revenue. Not yet. But hype drives the 00:03:11.250 --> 00:03:13.030 narrative and narrative drives stock prices. 00:03:13.210 --> 00:03:15.590 Right. Hype is leverage for the little guy. Which 00:03:15.590 --> 00:03:18.879 brings us to. How they generated that hype. We 00:03:18.879 --> 00:03:20.539 have to talk about the commercial. The Super 00:03:20.539 --> 00:03:23.360 Bowl ad. I watched this live. Open AI went first, 00:03:23.419 --> 00:03:26.560 didn't they? They did. And it was classic open 00:03:26.560 --> 00:03:30.180 AI. You know, piano music, safe imagery. Yeah. 00:03:30.280 --> 00:03:32.560 AI helps humanity. It felt like a bank commercial. 00:03:32.759 --> 00:03:35.599 Very corporate. Very safe. And then anthropic. 00:03:35.620 --> 00:03:38.039 Then anthropic chones violence. They really did. 00:03:38.159 --> 00:03:40.560 Walk us through it. So the scene is just a regular 00:03:40.560 --> 00:03:44.460 guy talking to an AI on his phone. He has a genuinely 00:03:44.460 --> 00:03:47.039 vulnerable question. How do I communicate better 00:03:47.039 --> 00:03:50.479 with my mom? Heavy. Very. And the AI starts giving 00:03:50.479 --> 00:03:53.379 this sincere, empathetic advice. It's touching. 00:03:53.419 --> 00:03:56.740 You feel the connection. And then just mid -sentence, 00:03:56.740 --> 00:03:59.479 the AI pivots. Oh, I remember this. It suddenly 00:03:59.479 --> 00:04:01.039 says, speaking of connections, have you tried 00:04:01.039 --> 00:04:03.680 this mature dating site? It was physically painful 00:04:03.680 --> 00:04:05.860 to watch. The actor just freezes. You just go, 00:04:05.900 --> 00:04:08.539 what? And the AI doubles down. Want me to create 00:04:08.539 --> 00:04:11.060 your profile? Cut to black. It was visceral. 00:04:11.120 --> 00:04:13.580 It played on that uncanny valley feeling where 00:04:13.580 --> 00:04:17.269 something human suddenly becomes. It was a precision 00:04:17.269 --> 00:04:19.170 strike. You have to remember the context, right? 00:04:19.490 --> 00:04:21.829 OpenAI had just announced they were introducing 00:04:21.829 --> 00:04:25.490 ads to the free tier. Right. Now, they promised 00:04:25.490 --> 00:04:29.670 explicitly that ads would not interrupt the conversation. 00:04:29.910 --> 00:04:32.750 There'd just be banners. But Anthropic didn't 00:04:32.750 --> 00:04:35.230 attack the reality. They attacked the anxiety. 00:04:35.730 --> 00:04:38.670 They attacked the nightmare scenario. Yes. They 00:04:38.670 --> 00:04:41.029 played on the fear that our private... intimate 00:04:41.029 --> 00:04:43.930 conversations are just inventory waiting to be 00:04:43.930 --> 00:04:47.069 sold. It forced everyone to look at OpenAI with 00:04:47.069 --> 00:04:50.329 suspicion right when OpenAI wanted to look like 00:04:50.329 --> 00:04:53.310 a savior. The reaction I saw was totally split. 00:04:53.509 --> 00:04:55.589 Half the Internet called it genius. The other 00:04:55.589 --> 00:04:57.629 half called it fear mongering because, you know, 00:04:57.629 --> 00:04:59.709 technically OpenAI said they wouldn't do that. 00:04:59.970 --> 00:05:02.329 It's the gray area of persuasion. It wasn't a 00:05:02.329 --> 00:05:04.529 lie, but is a simulation of a worst case scenario. 00:05:04.769 --> 00:05:07.230 And it worked. It got under people's skin specifically. 00:05:07.800 --> 00:05:10.240 Sam Altman's skin. Was it a cheap shot or just 00:05:10.240 --> 00:05:12.279 smart counter -programming? It was a marketing 00:05:12.279 --> 00:05:15.920 grenade. Dishonest, maybe, but effective because 00:05:15.920 --> 00:05:18.379 it targeted a specific anxiety. Speaking of getting 00:05:18.379 --> 00:05:21.120 under skin, let's talk about the CEO response. 00:05:21.540 --> 00:05:25.819 Sam Altman. A tweet. He takes to X. Usually the 00:05:25.819 --> 00:05:29.040 playbook for a CEO with 400 million users is 00:05:29.040 --> 00:05:31.120 to just ignore the noise. You don't punch down. 00:05:31.240 --> 00:05:33.339 You never punch down. But he couldn't help himself. 00:05:33.519 --> 00:05:36.160 He wrote this post defending their privacy principles, 00:05:36.379 --> 00:05:39.980 which is fine. But then he added this, this burn. 00:05:40.220 --> 00:05:44.259 He said more Texans use chat GPT for free than 00:05:44.259 --> 00:05:46.740 the total number of people using Claude in the 00:05:46.740 --> 00:05:49.579 entire U .S. It's a brutal statistic. And logically, 00:05:49.740 --> 00:05:52.050 he is right. He's flexing that data moat we talked 00:05:52.050 --> 00:05:54.389 about earlier. But strategically. Strategically, 00:05:54.470 --> 00:05:56.870 it was a disaster. It's a classic example of 00:05:56.870 --> 00:05:59.930 the backfire effect. The Anthropic ad had, what, 00:05:59.970 --> 00:06:03.009 about 2 .7 million views online? A decent hit. 00:06:03.110 --> 00:06:06.329 Yeah. Sam's response, that got 8 .8 million views. 00:06:06.529 --> 00:06:08.470 He amplified the thing he was trying to kill. 00:06:08.610 --> 00:06:10.910 He took a joke, albeit a mean one, and replied 00:06:10.910 --> 00:06:13.329 with a paragraph of statistics. The rule of the 00:06:13.329 --> 00:06:15.589 internet is simple. If you reply to a joke at 00:06:15.589 --> 00:06:17.930 a lecture, you just make the joke louder. He 00:06:17.930 --> 00:06:20.060 signaled that it hurt. Is there a lesson here 00:06:20.060 --> 00:06:22.680 for crisis management? Absolutely. Never reply 00:06:22.680 --> 00:06:24.680 to a joke with an essay. You just make the joke 00:06:24.680 --> 00:06:26.639 louder. So we have the marketing war. We have 00:06:26.639 --> 00:06:29.199 the ego war. But while all that was happening, 00:06:29.439 --> 00:06:31.680 the engineers were actually changing the world. 00:06:32.139 --> 00:06:34.139 Let's shift to the hardware drop. The real story, 00:06:34.300 --> 00:06:37.980 yeah. Thursday morning, 9 .00 a .m. Pacific. 00:06:38.819 --> 00:06:43.139 Anthropic drops Cloud Opus 4 .6. Then, seemingly 00:06:43.139 --> 00:06:47.660 out of nowhere, OpenAI drops GPT -5 .3 Codex 00:06:47.660 --> 00:06:51.040 at 10 .00 a .m. Not even, it was like 15 minutes 00:06:51.040 --> 00:06:53.259 later. That wasn't an accident. No. Let's look 00:06:53.259 --> 00:06:55.480 at Cloud first. I was reading the specs, and 00:06:55.480 --> 00:06:57.420 there's one number that just, it stopped me. 00:06:57.839 --> 00:07:01.300 A one million token context window. It's absurd. 00:07:01.519 --> 00:07:03.620 Imagine a context window of one million tokens. 00:07:03.680 --> 00:07:06.620 That is roughly 750 ,000 words. You could feed 00:07:06.620 --> 00:07:09.120 it an entire library of documentation, every 00:07:09.120 --> 00:07:11.000 code base you've ever written, and a couple of 00:07:11.000 --> 00:07:13.300 novels all in one breath. It changes the fundamental 00:07:13.300 --> 00:07:16.019 workflow. It moves us from asking questions to 00:07:16.019 --> 00:07:18.439 processing systems. This is designed for what 00:07:18.439 --> 00:07:21.639 we call agentic workflows. Define agentic for 00:07:21.639 --> 00:07:24.740 us. Think of it like a team of interns. You don't 00:07:24.740 --> 00:07:26.879 want an AI that just answers a question. You 00:07:26.879 --> 00:07:28.920 want an AI that can act. You give it a goal. 00:07:29.040 --> 00:07:31.240 And because it can hold that massive amount of 00:07:31.240 --> 00:07:33.660 context, your whole company history in its head, 00:07:33.779 --> 00:07:36.480 it can execute complex tasks without needing 00:07:36.480 --> 00:07:38.680 you to remind it of the rules every five minutes. 00:07:38.920 --> 00:07:41.240 So Claude is optimizing for memory. Exactly. 00:07:41.399 --> 00:07:43.379 Okay, so Claude is the librarian with a photographic 00:07:43.379 --> 00:07:47.800 memory. But then we have GBT 5 .3 Codex. And 00:07:47.800 --> 00:07:50.160 I have to admit, there is a detail in the source 00:07:50.160 --> 00:07:53.620 report about this model that it... unsettles 00:07:53.620 --> 00:07:55.319 me a bit. I think I know where you're going with 00:07:55.319 --> 00:07:58.040 this. The report mentions that OpenAI used early 00:07:58.040 --> 00:08:01.319 versions of GPT -5 .3 to debug its own training 00:08:01.319 --> 00:08:03.800 process. Yeah. It's recursive self -improvement. 00:08:03.879 --> 00:08:05.680 It is the holy grail and the sci -fi nightmare 00:08:05.680 --> 00:08:08.439 wrapped in one. It's a feedback loop. The model 00:08:08.439 --> 00:08:11.860 analyzes its own failures, fixes the code, and 00:08:11.860 --> 00:08:13.620 then trains a better version of itself. It's 00:08:13.620 --> 00:08:15.980 building its own ladder. Exactly. And the results 00:08:15.980 --> 00:08:18.439 are undeniable. On Terminal Bench 2 .0, which 00:08:18.439 --> 00:08:20.500 is the hardest coding benchmark we have, GPT 00:08:20.500 --> 00:08:26.000 5 .3 scored a 77 .3%. Clut Opus scored a 65 .4%. 00:08:26.000 --> 00:08:28.660 So GPT is significantly better at the raw engineering 00:08:28.660 --> 00:08:32.100 part. In terms of raw execution, yes. OpenAI 00:08:32.100 --> 00:08:35.440 optimized for get the code right. Anthropic optimized 00:08:35.440 --> 00:08:38.080 for understand the massive context. So if you 00:08:38.080 --> 00:08:40.039 had to pick a winner based on the drop, who takes 00:08:40.039 --> 00:08:42.240 it? Anthropic won the first move excitement. 00:08:42.730 --> 00:08:45.690 Yeah. But OpenAI won on raw engineering metrics. 00:08:46.149 --> 00:08:48.230 Okay. We are back. We've talked about the drama 00:08:48.230 --> 00:08:50.509 and the specs. Now I want to look at the vibe 00:08:50.509 --> 00:08:52.730 check because benchmarks are great, but how does 00:08:52.730 --> 00:08:54.809 it feel to actually use these things? Right. 00:08:54.889 --> 00:08:57.049 The source we're looking at did this fascinating 00:08:57.049 --> 00:09:00.110 test. They gave both models a really simple prompt. 00:09:00.529 --> 00:09:03.429 Build a landing page for a specialty coffee roastery 00:09:03.429 --> 00:09:06.210 in Florence, Italy. That's it. No design specs. 00:09:06.429 --> 00:09:08.529 Nothing. Just build it. Yeah. A test of their 00:09:08.529 --> 00:09:11.250 default personality. What did the AI assume is 00:09:11.250 --> 00:09:13.330 good? Okay, so Claude goes first. What's the 00:09:13.330 --> 00:09:15.690 result? Claude was very professional, clean. 00:09:15.929 --> 00:09:18.909 It used SVGs, which are scalable vector graphics, 00:09:19.049 --> 00:09:21.370 very sharp. It had these subtle bobbing animations. 00:09:21.570 --> 00:09:23.370 It felt like something a high -end design agency 00:09:23.370 --> 00:09:26.470 would deliver. Reliable. Very safe, very reliable. 00:09:26.690 --> 00:09:30.590 Okay. And then GPT 5 .3. GPT 5 .3 built a modern, 00:09:30.669 --> 00:09:34.169 trendy site, but it included an animated surfboard. 00:09:34.350 --> 00:09:39.419 A surfboard for a coffee shop. In Florence, Italy. 00:09:39.840 --> 00:09:42.019 Maybe the AI knows something about the Arno River 00:09:42.019 --> 00:09:45.019 that we don't. But the point was, it made a cool 00:09:45.019 --> 00:09:48.600 choice. It was trying to be trendy. It prioritized 00:09:48.600 --> 00:09:51.480 style over logical context. It's like the AI 00:09:51.480 --> 00:09:53.960 was trying to impress you with its flair. Look, 00:09:54.059 --> 00:09:56.600 I can do scroll animations. Look at this surfboard. 00:09:56.759 --> 00:09:58.519 Totally. Whereas Baud was like, here is your 00:09:58.519 --> 00:10:01.259 business asset, sir. Exactly. But here is the 00:10:01.259 --> 00:10:04.340 kicker. Both of them essentially did the job 00:10:04.340 --> 00:10:07.289 in 15 seconds. The difference wasn't capability. 00:10:07.529 --> 00:10:10.409 It was style. Does the average person actually 00:10:10.409 --> 00:10:12.649 care about the difference between clean SVGs 00:10:12.649 --> 00:10:15.250 and trendy motion? No. They just care that the 00:10:15.250 --> 00:10:17.830 website exists. Yeah. Both are magic to a non 00:10:17.830 --> 00:10:20.720 -coder. So zooming out. We have these two giants 00:10:20.720 --> 00:10:23.000 fighting. Why does this matter to you, the listener? 00:10:23.139 --> 00:10:25.059 Why should we care about the rivalry? Because 00:10:25.059 --> 00:10:27.220 of that word you mentioned earlier, recursive. 00:10:27.240 --> 00:10:30.159 The self -improvement. Yes. The fact that AI 00:10:30.159 --> 00:10:33.080 is helping build the next AI means the progress 00:10:33.080 --> 00:10:35.919 is accelerating. We aren't waiting years anymore 00:10:35.919 --> 00:10:37.759 for these updates. We're waiting months. And 00:10:37.759 --> 00:10:39.899 competition is the fuel for that. Competition 00:10:39.899 --> 00:10:43.539 is the only thing preventing a monopoly. If only 00:10:43.539 --> 00:10:45.899 open AI existed, we'd have to accept whatever 00:10:45.899 --> 00:10:48.799 ad model they chose. Anthropic forces them to 00:10:48.799 --> 00:10:51.070 be better. Are we heading toward a world where 00:10:51.070 --> 00:10:53.090 we can't keep up with the updates? We're already 00:10:53.090 --> 00:10:55.750 there. The cycle has moved from annual to monthly. 00:10:55.929 --> 00:10:59.110 So the big idea here, we have this massive gap 00:10:59.110 --> 00:11:04.490 between Twitter hype and real world usage. But 00:11:04.490 --> 00:11:07.389 the technology itself is accelerating because 00:11:07.389 --> 00:11:10.230 of this rivalry. Exactly. The Super Bowl antics 00:11:10.230 --> 00:11:13.330 are just noise. The signal is that AI is now 00:11:13.330 --> 00:11:15.740 improving itself recursively. That's the real 00:11:15.740 --> 00:11:17.940 story. I want to encourage you to try the coffee 00:11:17.940 --> 00:11:20.159 shop test yourself. Go see which vibe works for 00:11:20.159 --> 00:11:22.059 you. I also want to leave you with a final thought. 00:11:22.360 --> 00:11:24.799 If an AI can debug its own code better than a 00:11:24.799 --> 00:11:26.940 human, how long until it decides what features 00:11:26.940 --> 00:11:29.240 it wants to build next? That is the real question. 00:11:29.399 --> 00:11:30.080 Thanks for listening.