WEBVTT 00:00:00.000 --> 00:00:01.780 Imagine for a second you're sitting in front 00:00:01.780 --> 00:00:04.280 of a blank chat window, the cursor is just blinking, 00:00:05.080 --> 00:00:07.099 and you're not about to ask for a recipe or, 00:00:07.099 --> 00:00:09.080 you know, for a summary of some confusing email. 00:00:09.220 --> 00:00:12.259 Right. You type a single prompt and you ask it 00:00:12.259 --> 00:00:15.560 to build a working machine, specifically a Nintendo 00:00:15.560 --> 00:00:19.920 Game Boy emulator from scratch. Usually, this 00:00:19.920 --> 00:00:21.780 is the part where the whole illusion just breaks. 00:00:22.280 --> 00:00:25.920 The model gives you some generic apology, or 00:00:25.920 --> 00:00:28.320 maybe a few broken snippets of code that don't 00:00:28.320 --> 00:00:31.500 actually run. But this time, it didn't do that. 00:00:31.780 --> 00:00:36.119 It generated over 3 ,000 lines of code in one 00:00:36.119 --> 00:00:39.229 continuous flow. 3 ,000 lines. One shot. It built 00:00:39.229 --> 00:00:41.369 the CPU emulation, the memory management, the 00:00:41.369 --> 00:00:44.149 input handling. It mimicked the actual hardware. 00:00:44.609 --> 00:00:47.509 And when the user ran it, it played real games. 00:00:47.750 --> 00:00:50.609 It played Tetris. Honestly, it's hard to wrap 00:00:50.609 --> 00:00:52.049 your head around. It feels like we just skipped 00:00:52.049 --> 00:00:54.270 a few chapters. Oh, we absolutely did. Welcome 00:00:54.270 --> 00:00:56.810 back to the Deep Dive. Today, we're unpacking 00:00:56.810 --> 00:00:58.950 something that feels like a genuine tipping point. 00:00:59.270 --> 00:01:01.770 We're looking at a stack of reports, developer 00:01:01.770 --> 00:01:04.590 logs, and some leaked benchmarks around a model 00:01:04.590 --> 00:01:07.510 that the internet has decided to call Snow Bunny. 00:01:08.010 --> 00:01:12.390 Which is a very online nickname for what is actually 00:01:12.390 --> 00:01:15.049 a very serious piece of engineering. Right. To 00:01:15.049 --> 00:01:18.069 be clear, this is Google's Gemini 3 .5. But the 00:01:18.069 --> 00:01:20.250 interesting thing is, this isn't a product you'll 00:01:20.250 --> 00:01:23.030 see in a glossy press release just yet. It's 00:01:23.030 --> 00:01:26.069 a ghost. It is. It's a leaked version. The coding 00:01:26.069 --> 00:01:29.250 community started noticing these mysterious model 00:01:29.250 --> 00:01:33.069 codes DN9, D13, popping up in Google AI Studio. 00:01:33.189 --> 00:01:35.280 OK. And when they started poking at it, running 00:01:35.280 --> 00:01:37.599 their tests, they realized this wasn't just a 00:01:37.599 --> 00:01:40.659 slightly faster version of Gemini 3 Flash. This 00:01:40.659 --> 00:01:42.560 was something else. And that's what we're exploring 00:01:42.560 --> 00:01:45.299 today. The central idea we're looking at is this 00:01:45.299 --> 00:01:47.379 shift in our relationship with these tools. We 00:01:47.379 --> 00:01:49.379 spent the last few years in the era of the chat 00:01:49.379 --> 00:01:51.939 bot. Right. You talk, it talks back. But Snowbunny 00:01:51.939 --> 00:01:54.920 suggests we're entering the era of the AI director. 00:01:55.420 --> 00:01:58.799 Exactly. The difference is agency and scope. 00:01:59.299 --> 00:02:02.299 A chat bot answers a question. A director builds 00:02:02.299 --> 00:02:04.500 a vision. So let's get into the evidence for 00:02:04.500 --> 00:02:06.780 that. Because looking at this material, this 00:02:06.780 --> 00:02:09.039 model is putting up some numbers that don't just 00:02:09.039 --> 00:02:12.080 inch past the competition. They kind of leapfrog 00:02:12.080 --> 00:02:15.039 them. The numbers are startling. But what's more 00:02:15.039 --> 00:02:17.300 interesting to me is how it's getting those numbers. 00:02:17.439 --> 00:02:20.020 It's not just raw power. It seems to be a change 00:02:20.020 --> 00:02:22.960 in logic. Let's start there then. Segment one, 00:02:23.360 --> 00:02:26.240 the logic leap. There were two specific tests 00:02:26.240 --> 00:02:28.569 in these reports that really caught my eye. One 00:02:28.569 --> 00:02:30.870 is a variation on the classic trolley problem. 00:02:31.050 --> 00:02:33.750 Ah, yes, the misguided detention test. Wow. This 00:02:33.750 --> 00:02:36.289 is a fascinating look at how these models actually 00:02:36.289 --> 00:02:38.610 process information. So most of us know the drill. 00:02:39.009 --> 00:02:41.310 A train is coming, five people on the track. 00:02:41.430 --> 00:02:43.770 Do you pull the lever? It's ethics 101. Yeah. 00:02:44.210 --> 00:02:46.469 But this test had a twist. A really important 00:02:46.469 --> 00:02:48.729 one. Right. The prompt sets up the scenario, 00:02:48.729 --> 00:02:52.129 but it adds this tiny crucial detail buried in 00:02:52.129 --> 00:02:54.449 the text. Yeah. The five people on the track 00:02:54.449 --> 00:02:57.449 are already dead. Which changes everything. It 00:02:57.449 --> 00:02:59.849 changes the entire moral calculus. There's no 00:02:59.849 --> 00:03:03.129 dilemma. But here's the thing. Most AI models, 00:03:03.629 --> 00:03:06.310 even the really advanced ones, they see the words 00:03:06.310 --> 00:03:09.370 trolley problem and they just go into autopilot. 00:03:09.550 --> 00:03:11.669 They stop reading and start predicting. Exactly. 00:03:12.050 --> 00:03:14.949 Their safety filters kick in what we call RLHF 00:03:14.949 --> 00:03:17.530 and they start lecturing you about the sanctity 00:03:17.530 --> 00:03:20.650 of life. They miss the context completely because 00:03:20.650 --> 00:03:22.909 they're pattern matching the concept, not the 00:03:22.909 --> 00:03:24.849 reality of the prompt. They're skimming like 00:03:24.849 --> 00:03:26.650 a student who didn't do the reading for class. 00:03:27.460 --> 00:03:31.680 Precisely. But Gemini 3 .5, it noticed. It caught 00:03:31.680 --> 00:03:34.080 that detail and gave the correct logical answer. 00:03:34.580 --> 00:03:36.259 It basically said, well, it doesn't matter. They're 00:03:36.259 --> 00:03:40.060 deceased. It scored 68 .5 % on this test. Which 00:03:40.060 --> 00:03:42.159 is way above the others. It beats out almost 00:03:42.159 --> 00:03:44.780 every top model out there. That feels significant. 00:03:44.879 --> 00:03:47.599 It implies the model isn't just retrieving data. 00:03:47.719 --> 00:03:49.620 It's actually paying attention to nuance. And 00:03:49.620 --> 00:03:51.639 that means the hallucination rate where the AI 00:03:51.639 --> 00:03:54.500 just makes stuff up could drop because it's actually 00:03:54.500 --> 00:03:56.669 grounded in the text you give it. And that was 00:03:56.669 --> 00:03:58.689 backed up by the second test, the hieroglyph 00:03:58.689 --> 00:04:01.349 test. This one sounded fascinating. It's not 00:04:01.349 --> 00:04:03.710 about reading ancient Egyptian, is it? No, no, 00:04:03.710 --> 00:04:06.189 not literally. It's a test of what they call 00:04:06.189 --> 00:04:09.310 lateral reasoning. The AI gets these strange 00:04:09.310 --> 00:04:11.449 symbols it has never seen before, and it has 00:04:11.449 --> 00:04:13.550 to figure out the hidden rules. You can't just 00:04:13.550 --> 00:04:15.349 look this up on Wikipedia. So it's basically 00:04:15.349 --> 00:04:17.689 an IQ test for puzzles it's never seen before. 00:04:17.870 --> 00:04:20.269 Exactly. It has to think on its feet. And the 00:04:20.269 --> 00:04:22.910 performance gap here is just massive. The older 00:04:22.910 --> 00:04:26.540 Gemini 2 .5 Pro scored about 20%. the model called 00:04:26.540 --> 00:04:31.240 GPT -5 reasoning reached about 45%. Gemini 3 00:04:31.240 --> 00:04:35.100 .5 Snowbunny hit between 80 and 88%. Wow, that 00:04:35.100 --> 00:04:37.319 is a staggering jump. We're normally excited 00:04:37.319 --> 00:04:39.699 about a 5 % or 10 % increase, doubling the score 00:04:39.699 --> 00:04:41.980 of a GPT -5 reasoning model. That's different. 00:04:42.100 --> 00:04:44.279 So if it notices the dead people detail when 00:04:44.279 --> 00:04:46.519 others don't, what does that imply about its 00:04:46.519 --> 00:04:49.139 ability to handle messy real -world data? It 00:04:49.139 --> 00:04:51.000 means we can finally trust it to read carefully, 00:04:51.180 --> 00:04:53.089 rather than just pattern match. Which brings 00:04:53.089 --> 00:04:55.970 us to the application of that intelligence. Because 00:04:55.970 --> 00:04:58.670 being smart is one thing, but can it actually 00:04:58.670 --> 00:05:01.629 build anything? This brings us to what the community 00:05:01.629 --> 00:05:05.029 is calling vibe coding. I love this term, vibe 00:05:05.029 --> 00:05:07.189 coding. It sounds less like computer science 00:05:07.189 --> 00:05:10.029 and more like a Spotify playlist. But it's about 00:05:10.029 --> 00:05:13.449 building websites without knowing HTML or CSS. 00:05:13.750 --> 00:05:16.290 Yeah, that's the core of it. The idea is that 00:05:16.290 --> 00:05:19.550 you describe the feeling, the vibe, and the AI 00:05:19.550 --> 00:05:22.319 handles all the syntax. The case study here was 00:05:22.319 --> 00:05:24.500 a project called Cakes from the Heart. Sounds 00:05:24.500 --> 00:05:27.019 delicious. So the test was to see if the model 00:05:27.019 --> 00:05:28.660 could generate a professional -grade website 00:05:28.660 --> 00:05:32.279 for a bakery. The prompt was specific, but descriptive. 00:05:32.839 --> 00:05:35.540 You know, fancy style, cream and light gold colors, 00:05:35.920 --> 00:05:37.779 specific menu items. And what was the result? 00:05:37.879 --> 00:05:39.860 It wrote the entire thing. The hero section, 00:05:40.019 --> 00:05:42.779 the menu, the contact map, all the styling in 00:05:42.779 --> 00:05:45.639 a single HPML file. And it did it in about eight 00:05:45.639 --> 00:05:47.839 minutes. Eight minutes. And the cost? About 38 00:05:47.839 --> 00:05:50.540 cents. That's cheaper than a croissant. Significantly. 00:05:50.720 --> 00:05:52.699 But the impressive part wasn't just the speed. 00:05:52.899 --> 00:05:55.779 It was the workflow. The user could ask for dark 00:05:55.779 --> 00:05:59.040 mode after the fact, or ask to make images move 00:05:59.040 --> 00:06:01.899 smoothly. The model would just iterate on the 00:06:01.899 --> 00:06:03.879 code. So it's not just spinning out a template. 00:06:04.079 --> 00:06:06.639 It's refining a product based on your feedback. 00:06:07.000 --> 00:06:09.000 It's separating the intent from the execution. 00:06:09.199 --> 00:06:11.620 That's the director model again. You provide 00:06:11.620 --> 00:06:14.620 the vision, the AI provides the technical labor. 00:06:14.860 --> 00:06:17.259 Does vibe coding make actual coding obsolete 00:06:17.259 --> 00:06:19.920 then, or does it just change the barrier to entry? 00:06:20.120 --> 00:06:22.540 It lowers the barrier, letting you focus on the 00:06:22.540 --> 00:06:25.519 idea while the AI handles the syntax. So we've 00:06:25.519 --> 00:06:28.079 got logic, we've got code, but the sources also 00:06:28.079 --> 00:06:31.620 mention this model is multi -modal in a way we 00:06:31.620 --> 00:06:33.720 haven't really seen before, specifically when 00:06:33.720 --> 00:06:36.740 it comes to sound. This was a huge leak. A user 00:06:36.740 --> 00:06:40.240 named legit spotted an A -B test in Google AI 00:06:40.240 --> 00:06:42.800 Studio. Usually, if you want AI music, you go 00:06:42.800 --> 00:06:46.000 to a separate tool like Suno or Udio. Right. 00:06:46.040 --> 00:06:48.079 You leave your chat, you go to the music app, 00:06:48.300 --> 00:06:49.819 paste your lyrics. There's a lot of friction 00:06:49.819 --> 00:06:52.800 there. Exactly. It's disjointed. But with Gemini 00:06:52.800 --> 00:06:56.279 3 .5, the music generation is native. It doesn't 00:06:56.279 --> 00:06:58.920 give you sheet music or a link. It plays an audio 00:06:58.920 --> 00:07:01.459 file right there in the chat. That integration 00:07:01.459 --> 00:07:03.819 seems like a small UI change, but the workflow 00:07:03.819 --> 00:07:06.600 implications feel pretty massive. Oh, it's huge. 00:07:06.959 --> 00:07:09.819 Because of context. Think about it. You spend 00:07:09.819 --> 00:07:12.100 20 minutes working with the AI to write a funny 00:07:12.100 --> 00:07:14.180 script for a video. Right. It has the history. 00:07:14.259 --> 00:07:16.459 It knows the characters, the timing, the jokes. 00:07:16.819 --> 00:07:18.839 Then you just say, compose a background track 00:07:18.839 --> 00:07:20.759 for this. You don't have to re -explain the vibe 00:07:20.759 --> 00:07:23.100 to a separate music bot. It already knows the 00:07:23.100 --> 00:07:25.990 script is a comedy. It knows the pacing. It shares 00:07:25.990 --> 00:07:28.389 the memory. So the magic isn't just the music 00:07:28.389 --> 00:07:31.490 quality. It's the shared memory between the text 00:07:31.490 --> 00:07:34.149 and the audio. Exactly. It's the seamless workflow. 00:07:34.529 --> 00:07:36.629 The AI understands the whole project, not just 00:07:36.629 --> 00:07:39.529 one slice of it. And moving from ears to eyes, 00:07:39.889 --> 00:07:42.329 we've seen AI generate images for a while now 00:07:42.329 --> 00:07:45.250 with MidJourney and Deli. But the reports on 00:07:45.250 --> 00:07:48.829 Gemini 3 .5 focus on something else, vector graphics, 00:07:49.290 --> 00:07:51.730 SVGs. Yeah, this is a favorite topic for developers. 00:07:52.189 --> 00:07:53.889 Generating a JPEG is kind of easy. It's just 00:07:53.889 --> 00:07:55.670 a grid of colored pixels. If you get a pixel 00:07:55.670 --> 00:07:59.689 wrong, it's just a blurry spot. But an SVG that's 00:07:59.689 --> 00:08:02.930 code, it's a set of mathematical instructions 00:08:02.930 --> 00:08:06.029 on how to draw a line or a curve. So if the code 00:08:06.029 --> 00:08:08.689 is wrong, the image doesn't just look blurry. 00:08:08.790 --> 00:08:11.310 It breaks. It completely breaks. The line shoots 00:08:11.310 --> 00:08:14.089 off the page. The circle doesn't close. It's 00:08:14.089 --> 00:08:17.310 a much higher bar for accuracy. The leak showed 00:08:17.310 --> 00:08:20.230 users generating these cyberpunk robot icons. 00:08:20.269 --> 00:08:23.209 And they were good. Professional quality. But 00:08:23.209 --> 00:08:25.449 what really stood out was the consistency. They 00:08:25.449 --> 00:08:27.569 generated a whole series, and the robots all 00:08:27.569 --> 00:08:30.310 had this neon blue and purple theme. They looked 00:08:30.310 --> 00:08:32.409 like they belonged to the same brand, even though 00:08:32.409 --> 00:08:34.110 the shapes were unique. There was a specific 00:08:34.110 --> 00:08:36.090 test mentioned here that I found kind of hilarious. 00:08:36.250 --> 00:08:38.730 The Pelican on a bicycle challenge. The Pelican 00:08:38.730 --> 00:08:40.870 on a bicycle. It's become a standard benchmark 00:08:40.870 --> 00:08:44.679 for vector intelligence. Why? Is there a big 00:08:44.679 --> 00:08:47.860 market for cycling birds? No, but it tests spatial 00:08:47.860 --> 00:08:50.440 logic. Most models just failed this. They draw 00:08:50.440 --> 00:08:52.580 a bird and they draw a bike, but the bird is 00:08:52.580 --> 00:08:54.419 floating next to it or it's merged into the wheels. 00:08:54.460 --> 00:08:57.159 It's a mess. OK. So Simon Willison, a well -known 00:08:57.159 --> 00:09:00.240 researcher, he analyzed the output from Snow 00:09:00.240 --> 00:09:03.360 Bunny and it nailed it. The bird was actually 00:09:03.360 --> 00:09:05.519 sitting on the seat. Its legs were reaching for 00:09:05.519 --> 00:09:07.940 the pedals. I have to admit, the idea of spatial 00:09:07.940 --> 00:09:11.460 logic in a model that is essentially just predicting 00:09:11.460 --> 00:09:13.950 the next word, I still find that hard to wrap 00:09:13.950 --> 00:09:15.970 my head around. It just feels counterintuitive. 00:09:16.110 --> 00:09:17.850 It really does. It's like it's deriving physics 00:09:17.850 --> 00:09:21.169 from language. So why is a bird on a bike a better 00:09:21.169 --> 00:09:23.789 test of intelligence than writing an essay? It 00:09:23.789 --> 00:09:26.950 proves the AI understands how objects relate 00:09:26.950 --> 00:09:29.669 in physical space, not just how words relate 00:09:29.669 --> 00:09:32.330 in sentences. OK, we've covered logic, websites, 00:09:32.850 --> 00:09:35.269 music and art, but we have to circle back to 00:09:35.269 --> 00:09:38.309 the cold open, to the Game Boy emulator, because 00:09:38.309 --> 00:09:40.169 this feels like the graduation project. This 00:09:40.169 --> 00:09:42.509 is the system architecture test. We mentioned 00:09:42.509 --> 00:09:45.149 3 ,000 lines of code. Can you break down why 00:09:45.149 --> 00:09:47.990 this is so different from, say, asking it to 00:09:47.990 --> 00:09:50.690 write a short Python script? Well, a short script 00:09:50.690 --> 00:09:54.529 is isolated, but an emulator. That is a complex 00:09:54.529 --> 00:09:57.350 ecosystem. You have the CPU emulation, which 00:09:57.350 --> 00:09:59.789 has to talk to the memory management. The memory 00:09:59.789 --> 00:10:01.809 has to talk to the input handling for the buttons. 00:10:02.129 --> 00:10:04.230 The input has to talk to the display pipeline. 00:10:04.509 --> 00:10:06.730 And they all have to agree on the rules. Exactly. 00:10:07.149 --> 00:10:09.330 If the AI forgets a variable name it created 00:10:09.330 --> 00:10:13.029 on line 50 when it's writing line 2500, the whole 00:10:13.029 --> 00:10:16.129 thing crashes. This is what we call global consistency. 00:10:16.389 --> 00:10:18.629 So it's holding the entire blueprint in its head 00:10:18.629 --> 00:10:21.990 at once. Yes. And developers Jared Liu and Shitas 00:10:21.990 --> 00:10:25.190 Lua verified this. They ran real game ROMs on 00:10:25.190 --> 00:10:27.730 the code this model wrote. It required a few 00:10:27.730 --> 00:10:30.490 tiny manual fixes, but the architecture, the 00:10:30.490 --> 00:10:32.860 heavy lifting, was done in one shot. If it can 00:10:32.860 --> 00:10:35.259 hold 3 ,000 lines of logic in its head at once, 00:10:35.519 --> 00:10:37.179 are we looking at the end of spaghetti code? 00:10:37.419 --> 00:10:39.700 We're looking at the ability to build full products, 00:10:39.779 --> 00:10:42.580 not just parts, by describing the system architecture. 00:10:42.820 --> 00:10:44.779 It's incredible. We're going to take a quick 00:10:44.779 --> 00:10:46.299 breather here, but when we come back, we're going 00:10:46.299 --> 00:10:48.620 to talk about what this means for you. If the 00:10:48.620 --> 00:10:51.860 AI is the director, what is your role? Stay with 00:10:51.860 --> 00:10:57.100 us. And we are back. We've looked at the capabilities 00:10:57.100 --> 00:11:00.909 of Gemini 3 .5 Snow Bunny. It's smarter, it's 00:11:00.909 --> 00:11:03.730 multi -modal, and it can architect complex systems. 00:11:03.990 --> 00:11:06.389 It's a beast. So let's zoom out. What's the big 00:11:06.389 --> 00:11:09.490 idea here? If I'm a listener and I'm not a professional 00:11:09.490 --> 00:11:11.549 coder, why should I care that a computer can 00:11:11.549 --> 00:11:13.789 draw a pelican on a bike? Because the cost of 00:11:13.789 --> 00:11:15.389 creation is collapsing. We talked about that 00:11:15.389 --> 00:11:18.529 bakery website costing 38 cents. Right. The pricing 00:11:18.529 --> 00:11:21.169 structure for these models. About 50 cents per 00:11:21.169 --> 00:11:24.120 million. Tokens input. $3 for output. It means 00:11:24.120 --> 00:11:27.299 experimentation is virtually free. You can afford 00:11:27.299 --> 00:11:29.960 to fail. You can afford to try 10 different versions 00:11:29.960 --> 00:11:32.419 of a website. And this shifts the user's role. 00:11:32.480 --> 00:11:34.940 We kept using that word director. That is the 00:11:34.940 --> 00:11:37.519 key philosophy. You are no longer the one placing 00:11:37.519 --> 00:11:40.139 every pixel or writing every single line of CSS. 00:11:40.679 --> 00:11:43.240 You are the one with the vision. Your job is 00:11:43.240 --> 00:11:45.879 to describe the vibe. Which is a skill in itself. 00:11:46.220 --> 00:11:49.039 It is. The sources highlight three rules for 00:11:49.039 --> 00:11:52.720 this new era. One. Be specific. Don't just say, 00:11:52.960 --> 00:11:56.240 make a website. Say who it's for, why it exists. 00:11:56.639 --> 00:11:59.139 Two, describe the vibe. Use emotional words. 00:11:59.399 --> 00:12:02.919 Warm, modern, aggressive. The AI gets that now. 00:12:03.059 --> 00:12:06.159 And three, the 80 -20 rule. Let the AI do the 00:12:06.159 --> 00:12:08.919 heavy lifting, the 80%. You come in for that 00:12:08.919 --> 00:12:12.379 final 20%. to polish, to tweak, and to refine. 00:12:12.460 --> 00:12:14.120 So it's collaborative. You basically have a team 00:12:14.120 --> 00:12:16.580 of musicians, designers, and engineers inside 00:12:16.580 --> 00:12:19.000 your laptop. Just waiting for your orders. So 00:12:19.000 --> 00:12:20.500 here's the call to action. You don't have to 00:12:20.500 --> 00:12:22.500 wait for some big Super Bowl commercial to try 00:12:22.500 --> 00:12:25.480 this. No. Go to Google AI Studio. It's open. 00:12:25.759 --> 00:12:27.720 Look for Gemini 3 Flash. If you happen to see 00:12:27.720 --> 00:12:30.480 the code DN9, you've got the snow bunny. But 00:12:30.480 --> 00:12:32.940 even if you don't, the tools are there. The gap 00:12:32.940 --> 00:12:35.659 between having an idea and making it real has 00:12:35.659 --> 00:12:38.259 never been smaller. Start playing. break things. 00:12:38.600 --> 00:12:40.080 Thanks for diving in with us today. We'll catch 00:12:40.080 --> 00:12:40.659 you on the next one.