WEBVTT 00:00:00.000 --> 00:00:02.580 We watched Meta start the year as the open source 00:00:02.580 --> 00:00:05.259 darling, sharing their llama models everywhere, 00:00:05.480 --> 00:00:08.880 inviting the community in. Right. Now they are 00:00:08.880 --> 00:00:11.119 secretly building something called Avocado and 00:00:11.119 --> 00:00:13.560 they are betting the entire company's future 00:00:13.560 --> 00:00:17.239 on keeping it locked down. That pivot from, you 00:00:17.239 --> 00:00:19.920 know, communal sharing and transparency to suddenly 00:00:19.920 --> 00:00:22.870 locking down their biggest asset. is more than 00:00:22.870 --> 00:00:25.629 just corporate strategy. It's an identity crisis 00:00:25.629 --> 00:00:29.109 for Meta, signaling a massive high stakes reversal 00:00:29.109 --> 00:00:32.530 that affects the entire AI landscape. They feel 00:00:32.530 --> 00:00:34.630 they have to win at the very top tier, whatever 00:00:34.630 --> 00:00:37.250 the cost. Welcome to the Deep Dive. Our mission 00:00:37.250 --> 00:00:39.490 today is to unpack these fundamental shifts. 00:00:39.770 --> 00:00:41.909 Using the recent source material you shared with 00:00:41.909 --> 00:00:44.070 us, we have a lot of tension to work through. 00:00:44.189 --> 00:00:46.649 First, we're going to dig deep into Meta's dramatic 00:00:46.649 --> 00:00:49.509 internal culture shock and the mysterious delay 00:00:49.509 --> 00:00:52.329 of that closed source avocado model. Yeah, there's 00:00:52.329 --> 00:00:54.689 a lot there. Then we'll pivot to AI in the wild, 00:00:54.829 --> 00:00:57.670 looking at practical chat GPT tips, the rapid 00:00:57.670 --> 00:00:59.649 commoditization happening in the voice market, 00:00:59.829 --> 00:01:03.189 and some of the unexpected ways AI is showing 00:01:03.189 --> 00:01:06.439 up. In ads and even late -night chats. Exactly. 00:01:06.659 --> 00:01:10.040 And finally, we'll analyze a truly significant 00:01:10.040 --> 00:01:13.260 scientific breakthrough, Microsoft's open -source 00:01:13.260 --> 00:01:15.879 Gigatime model, which is dramatically reducing 00:01:15.879 --> 00:01:18.840 the cost of cancer research. Down to about $10 00:01:18.840 --> 00:01:21.840 per analysis. It's incredible. It is. So we aren't 00:01:21.840 --> 00:01:23.980 just reading headlines today. We are looking 00:01:23.980 --> 00:01:27.540 for the why. Behind the biggest moves connecting 00:01:27.540 --> 00:01:30.359 the docks between, you know, Silicon Valley drama 00:01:30.359 --> 00:01:32.939 and breakthroughs that fundamentally change science. 00:01:33.180 --> 00:01:35.780 So let's start with Meta. Just months ago, they 00:01:35.780 --> 00:01:38.239 were the hero of the open source community hyping 00:01:38.239 --> 00:01:41.459 Lama. But the sources show Lama is barely mentioned 00:01:41.459 --> 00:01:44.420 now. It's almost as if it's been quietly abandoned. 00:01:44.780 --> 00:01:47.519 What's fascinating here is how immediate and 00:01:47.519 --> 00:01:49.939 radical this shift to secrecy has been. And it's 00:01:49.939 --> 00:01:52.819 all about this new model, Avocado. The secretive 00:01:52.819 --> 00:01:55.859 next gen model is Avocado. It was expected to 00:01:55.859 --> 00:01:58.400 launch by the end of 2025, but that timeline 00:01:58.400 --> 00:02:00.900 is already slipping. Push back. Push back to 00:02:00.900 --> 00:02:03.700 Q1 of 2026. And the biggest fear among researchers 00:02:03.700 --> 00:02:05.879 is that this will be a fully closed source model. 00:02:06.000 --> 00:02:08.020 That's a complete dramatic reversal for them. 00:02:08.180 --> 00:02:10.860 It is. I mean, they built their modern AI reputation 00:02:10.860 --> 00:02:14.000 on being the most generous sharer of foundation 00:02:14.000 --> 00:02:17.740 models. So why the sudden loss of faith in their 00:02:17.740 --> 00:02:20.680 own strategy? Well, the core hypothesis outlined 00:02:20.680 --> 00:02:23.419 in your sources is simple, if a little brutal. 00:02:23.520 --> 00:02:26.120 Okay. Meta concluded that open source simply 00:02:26.120 --> 00:02:29.139 didn't pay off. They poured billions in, but 00:02:29.139 --> 00:02:31.020 they didn't get the revenue or the market lead 00:02:31.020 --> 00:02:33.569 they wanted. So they feel they need to own the 00:02:33.569 --> 00:02:36.789 next GPT level model to stay relevant. They feel 00:02:36.789 --> 00:02:39.030 they absolutely have to. That urgency is translating 00:02:39.030 --> 00:02:41.389 into some pretty brutal internal changes. We're 00:02:41.389 --> 00:02:43.509 hearing the culture has really fractured. It 00:02:43.509 --> 00:02:47.870 has. The new mantra inside key AI teams is demo, 00:02:47.969 --> 00:02:50.270 don't memo. No more lengthy workplace updates 00:02:50.270 --> 00:02:52.710 then. Exactly. They're forcing teams to operate 00:02:52.710 --> 00:02:55.669 like these intense, scrappy startups, pushing 00:02:55.669 --> 00:02:58.129 for immediate results. And that pressure cooker 00:02:58.129 --> 00:03:01.030 environment is leading to, what was it, 70 hours? 00:03:01.069 --> 00:03:03.229 Workweeks. Reported 70 -hour workweeks. Yeah. 00:03:03.389 --> 00:03:05.650 Coupled with targeted layoffs that are specifically 00:03:05.650 --> 00:03:07.969 hitting the FAIR division. And FAIR is their 00:03:07.969 --> 00:03:10.409 fundamental AI research group. That's where all 00:03:10.409 --> 00:03:12.909 the academic, open -minded research used to happen. 00:03:13.090 --> 00:03:15.289 That was the heart of it. And that culture is 00:03:15.289 --> 00:03:18.270 being dismantled. It seems less about sustainability 00:03:18.270 --> 00:03:22.800 and more about desperation. Deliver or exit. 00:03:22.939 --> 00:03:24.960 And we're seeing that at the top, too. The sources 00:03:24.960 --> 00:03:28.060 highlight the replacement of longtime meta executives. 00:03:28.199 --> 00:03:30.780 Right. And the departure of chief scientist Jan 00:03:30.780 --> 00:03:34.020 LeCun back in October when a pioneer like LeCun, 00:03:34.120 --> 00:03:35.960 who basically is the champion of open research, 00:03:36.219 --> 00:03:39.520 leaves. Or is pushed out. Or is pushed out. Yeah. 00:03:39.620 --> 00:03:41.780 It signals a fundamental housecleaning. It's 00:03:41.780 --> 00:03:44.120 all about securing results, period. And that 00:03:44.120 --> 00:03:46.240 pressure is also being intensified by some hard 00:03:46.240 --> 00:03:48.199 infrastructure challenges, right? Precisely. 00:03:49.179 --> 00:03:52.180 Meta's massive, what, $27 billion data center 00:03:52.180 --> 00:03:54.919 still isn't fully ready. So they have to rely 00:03:54.919 --> 00:03:57.060 on third -party clouds just to train their own 00:03:57.060 --> 00:03:59.560 models. Which drives up costs and adds latency. 00:03:59.819 --> 00:04:01.520 They're playing catch -up, and they're doing 00:04:01.520 --> 00:04:03.759 it by changing their entire identity. Okay, so 00:04:03.759 --> 00:04:06.280 let's unpack this. The main reason for Meta's 00:04:06.280 --> 00:04:09.039 dramatic cultural overhaul. They believe open 00:04:09.039 --> 00:04:12.659 source failed to pay off, forcing this high -stakes 00:04:12.659 --> 00:04:15.240 closed model push. So that's the heavy corporate 00:04:15.240 --> 00:04:18.689 drama. Let's pivot now. What ties all this together 00:04:18.689 --> 00:04:21.790 is how fast the technology itself is moving outside 00:04:21.790 --> 00:04:24.050 of these big labs. Yeah, let's look at what that 00:04:24.050 --> 00:04:26.089 means for you, the user. Okay. We can start with 00:04:26.089 --> 00:04:29.230 some practical nuggets. OpenAI recently dropped 00:04:29.230 --> 00:04:32.649 six official chat GPT tips for getting better 00:04:32.649 --> 00:04:35.360 output. I saw that. The concepts aren't entirely 00:04:35.360 --> 00:04:38.180 new, but it's a useful summary. It is. Things 00:04:38.180 --> 00:04:40.500 like starting with a clear role for the model. 00:04:40.620 --> 00:04:43.560 And, you know, I'll admit I still wrestle with 00:04:43.560 --> 00:04:46.540 Trump drift myself. Oh, me too. It's that frustrating 00:04:46.540 --> 00:04:48.579 moment when the model just starts losing the 00:04:48.579 --> 00:04:50.800 original context of your conversation. Exactly. 00:04:51.310 --> 00:04:53.250 So having these codified tips is actually pretty 00:04:53.250 --> 00:04:55.930 helpful. And speaking of utility, the Adobe integration 00:04:55.930 --> 00:04:59.709 sounds like a real game changer. Oh, it is. Photoshop 00:04:59.709 --> 00:05:02.430 and Express are finally accessible just using 00:05:02.430 --> 00:05:04.550 natural language prompts. You can literally edit 00:05:04.550 --> 00:05:07.350 images just by talking. Which radically lowers 00:05:07.350 --> 00:05:10.350 the barrier to entry for creative work. No more 00:05:10.350 --> 00:05:13.189 navigating complex menus. Not every deployment 00:05:13.189 --> 00:05:15.410 goes that smoothly. We saw that with McDonald's. 00:05:15.490 --> 00:05:18.750 Right, their AI -generated Christmas ad. The 00:05:18.750 --> 00:05:21.509 sources detail this quick, intense backlash that 00:05:21.509 --> 00:05:23.589 caused them to pull it. People called it creepy 00:05:23.589 --> 00:05:26.050 and god -awful. Yeah. It's a great example of 00:05:26.050 --> 00:05:28.790 how quickly the public rejects AI when it just 00:05:28.790 --> 00:05:32.029 feels unsettling. It's a crucial lesson for marketers, 00:05:32.189 --> 00:05:34.810 for sure. And on the business front, Forbes released 00:05:34.810 --> 00:05:38.230 a surprisingly honest prediction piece for 2026. 00:05:38.430 --> 00:05:41.170 The honesty was the key part. Yes. They explicitly 00:05:41.170 --> 00:05:43.790 admit they got their previous AI timing estimates 00:05:43.790 --> 00:05:47.050 wrong. Now they are focusing on 10 specific bets 00:05:47.050 --> 00:05:49.930 for automation and the future of work. That kind 00:05:49.930 --> 00:05:52.689 of candor is so rare. And we're also seeing the 00:05:52.689 --> 00:05:55.110 competition just heat up globally. Google launched 00:05:55.110 --> 00:05:57.949 a massive move in India. Aggressive. They're 00:05:57.949 --> 00:06:00.449 offering their AI Plus plan with Gemini 3 Pro. 00:06:00.529 --> 00:06:02.649 It starts for new users at about $2 a month. 00:06:02.769 --> 00:06:04.769 Just $2. They are clearly going after market 00:06:04.769 --> 00:06:07.050 share where Meta's open model struggled to get 00:06:07.050 --> 00:06:09.290 a commercial foothold. And all this usage is 00:06:09.290 --> 00:06:12.449 revealing some fascinating things about us. Microsoft's 00:06:12.449 --> 00:06:16.430 end of year report analyzed 37 .5 million chats. 00:06:16.829 --> 00:06:18.970 And they found some highly unexpected patterns. 00:06:19.230 --> 00:06:21.930 Things like 2 a .m. philosophy chats. Philosophy 00:06:21.930 --> 00:06:24.769 chats. That's amazing. Not productivity, but 00:06:24.769 --> 00:06:27.370 deep late night reflection. Exactly. And whoa, 00:06:27.709 --> 00:06:31.750 imagine scaling that analysis to a billion queries 00:06:31.750 --> 00:06:34.430 across different time zones. Right. It paints 00:06:34.430 --> 00:06:36.990 a picture of AI, not just as a tool for drafting 00:06:36.990 --> 00:06:40.160 email. but as this silent late night sounding 00:06:40.160 --> 00:06:43.449 board for You know, human existence. A nonjudgmental 00:06:43.449 --> 00:06:45.550 digital listener. I like that. And this pressure 00:06:45.550 --> 00:06:47.449 is hitting niche companies, too. Eleven Labs, 00:06:47.709 --> 00:06:49.889 the voice synthesis leader, they just raised 00:06:49.889 --> 00:06:53.709 $100 million. But their CEO thinks voice AI itself 00:06:53.709 --> 00:06:56.189 is going to be commoditized. He does. So their 00:06:56.189 --> 00:06:59.149 new bet is on full AI agents, music generation, 00:06:59.389 --> 00:07:02.089 and crucially, deepfake protection tools. So 00:07:02.089 --> 00:07:04.069 what does the Eleven Labs shift, moving away 00:07:04.069 --> 00:07:06.430 from just voice, reveal about the current AI 00:07:06.430 --> 00:07:08.889 market? Even successful niche tech will commoditize 00:07:08.889 --> 00:07:11.819 fast, pushing focus to f***. full AI agents. 00:07:11.980 --> 00:07:13.920 That makes perfect sense. The value moves from 00:07:13.920 --> 00:07:16.399 the parts to the whole system. Okay, let's shift 00:07:16.399 --> 00:07:18.759 gears entirely now. We're moving from commercial 00:07:18.759 --> 00:07:21.660 tension to a highly impactful piece of scientific 00:07:21.660 --> 00:07:24.500 research. This is where that shared knowledge 00:07:24.500 --> 00:07:28.379 model really, really changes lives. We're talking 00:07:28.379 --> 00:07:30.459 about Microsoft's breakthrough with an open source 00:07:30.459 --> 00:07:32.990 model called Gigatime. And what's remarkable 00:07:32.990 --> 00:07:36.750 here is that it democratizes access to high -end 00:07:36.750 --> 00:07:39.170 medical diagnosis. It really does. So how does 00:07:39.170 --> 00:07:42.970 it do that? It takes a basic, low -cost, $10 00:07:42.970 --> 00:07:45.829 tissue slide, the kind any local hospital can 00:07:45.829 --> 00:07:50.269 produce, and transforms it into the rich, detailed 00:07:50.269 --> 00:07:52.930 immune system analysis that previously required 00:07:52.930 --> 00:07:56.089 specialized expensive machines. And human specialists 00:07:56.089 --> 00:07:59.009 and days of work. Right. It essentially turns... 00:07:59.230 --> 00:08:02.529 cheap, widely available data into high -end diagnostic 00:08:02.529 --> 00:08:05.129 data. It maps the immune system and the tumor 00:08:05.129 --> 00:08:07.089 environment. Like creating these high -resolution 00:08:07.089 --> 00:08:10.389 cancer maps. Exactly. To do this, the model was 00:08:10.389 --> 00:08:13.290 trained on a massive data set, 40 million cell 00:08:13.290 --> 00:08:15.889 samples from Providence Health. 40 million! That's 00:08:15.889 --> 00:08:18.509 a huge data scale. It's like stacking Lego blocks 00:08:18.509 --> 00:08:21.370 of data, but for medicine. The depth is unparalleled. 00:08:21.470 --> 00:08:23.329 And this wasn't just a lab exercise. No, it was 00:08:23.329 --> 00:08:26.329 tested on real patients. Over 14 ,000 real patients 00:08:26.329 --> 00:08:28.769 across 24 different cancer types. And the outcome. 00:08:29.029 --> 00:08:31.730 It created a virtual tumor library of more than 00:08:31.730 --> 00:08:35.190 300 ,000 high -resolution images. And it surfaced 00:08:35.190 --> 00:08:39.549 over 1 ,200 entirely new immune patterns. Patterns 00:08:39.549 --> 00:08:41.929 linked directly to cancer stage and survival 00:08:41.929 --> 00:08:45.889 statistics. Yeah. That is profound. And the key 00:08:45.889 --> 00:08:48.269 strategic choice here, again, is that Microsoft 00:08:48.269 --> 00:08:50.820 open -sourced the model. That open source move 00:08:50.820 --> 00:08:53.679 is strategic genius. It accelerates global adoption 00:08:53.679 --> 00:08:57.360 immediately. Creating a massive real world feedback 00:08:57.360 --> 00:09:00.899 loop. Yes. The playbook is clear. Cheaper tools 00:09:00.899 --> 00:09:03.120 mean more hospitals use them, which generates 00:09:03.120 --> 00:09:05.600 better local data, which builds better models 00:09:05.600 --> 00:09:08.820 globally faster. It's a virtuous cycle for science. 00:09:09.000 --> 00:09:10.659 So if we connect this to the bigger picture, 00:09:10.799 --> 00:09:13.519 what is the critical advantage of using Gigatime's 00:09:13.519 --> 00:09:16.330 open source strategy? It builds a massive feedback 00:09:16.330 --> 00:09:19.210 ecosystem, leading to better models faster through 00:09:19.210 --> 00:09:21.669 global use. So what does this all mean for us? 00:09:21.769 --> 00:09:24.350 The AI industry right now is defined by these 00:09:24.350 --> 00:09:26.830 incredibly dramatic internal shifts. The cultural 00:09:26.830 --> 00:09:29.350 demolition inside metaphor avocado. Right. And 00:09:29.350 --> 00:09:32.409 at the same time, this phenomenal, undeniable 00:09:32.409 --> 00:09:35.190 real world utility like Gigatime, we're watching 00:09:35.190 --> 00:09:37.169 two radically different philosophies collide. 00:09:37.470 --> 00:09:40.460 And the key takeaway is this. The race isn't 00:09:40.460 --> 00:09:42.320 simply about building the most powerful model, 00:09:42.440 --> 00:09:45.620 period. It's about choosing the strategy closed 00:09:45.620 --> 00:09:48.500 and proprietary, which is Meta's new bet, or 00:09:48.500 --> 00:09:51.299 open an ecosystem building like Microsoft is 00:09:51.299 --> 00:09:53.539 doing in medicine. Which one wins the long -term 00:09:53.539 --> 00:09:56.840 innovation war? Exactly. And consider this as 00:09:56.840 --> 00:09:59.590 you go about your day. Does the future of critical 00:09:59.590 --> 00:10:02.789 innovation hinge on proprietary control, which 00:10:02.789 --> 00:10:05.289 Meta is betting billions on with a secretive 00:10:05.289 --> 00:10:07.889 model like Avocado? Or does the greatest human 00:10:07.889 --> 00:10:10.490 impact come from shared open knowledge demonstrated 00:10:10.490 --> 00:10:13.730 by life -saving accessible models like Gigatime? 00:10:13.870 --> 00:10:17.370 Which path truly advances society faster? Thank 00:10:17.370 --> 00:10:19.009 you for sharing the sources for this deep dive. 00:10:19.169 --> 00:10:21.250 Keep exploring that intersection of technology 00:10:21.250 --> 00:10:23.250 and human impact because that's where the real 00:10:23.250 --> 00:10:23.789 story lives.