WEBVTT 00:00:00.000 --> 00:00:03.419 Anthropic locked their most dangerous AI in a 00:00:03.419 --> 00:00:07.160 high -security vault. A 22 -year -old dropout 00:00:07.160 --> 00:00:09.960 cracked it open. He gave it to the world in under 00:00:09.960 --> 00:00:12.019 two weeks. The digital walls are coming down 00:00:12.019 --> 00:00:15.380 incredibly fast right now. Welcome to the Deep 00:00:15.380 --> 00:00:18.460 Dive. Today, we are exploring a massive shift. 00:00:18.640 --> 00:00:21.379 The balance of technological power is fundamentally 00:00:21.379 --> 00:00:23.940 changing. We are bringing you the latest signals 00:00:23.940 --> 00:00:26.570 from the frontier. Right. We're looking at how 00:00:26.570 --> 00:00:29.649 a single GitHub repository is threatening a trillion 00:00:29.649 --> 00:00:33.090 dollars of big tech infrastructure today. We're 00:00:33.090 --> 00:00:36.009 also looking at your doctor's office. AI is officially 00:00:36.009 --> 00:00:38.689 stepping in as an active medical teammate. And 00:00:38.689 --> 00:00:41.929 finally, we'll explore the messy, deeply human 00:00:41.929 --> 00:00:44.109 friction happening as these models integrate 00:00:44.109 --> 00:00:46.909 into our daily lives. It's a fascinating moment 00:00:46.909 --> 00:00:49.030 in time. Let's start with this idea of protecting 00:00:49.030 --> 00:00:52.000 AI. Entropic had this incredibly powerful system 00:00:52.000 --> 00:00:55.119 called Claude Mythos. They actually decided it 00:00:55.119 --> 00:00:56.820 was simply too dangerous for the public. Yeah, 00:00:56.880 --> 00:00:58.560 they wanted to keep it strictly contained. They 00:00:58.560 --> 00:01:00.619 launched something they called Project Glasswing. 00:01:00.780 --> 00:01:03.539 Right. The entire goal was to secure the model 00:01:03.539 --> 00:01:06.180 completely. They were carefully vetting all their 00:01:06.180 --> 00:01:08.739 corporate partners. They essentially built this 00:01:08.739 --> 00:01:12.780 massive digital moat to keep it safe. They wanted 00:01:12.780 --> 00:01:15.560 absolute control over the deployment. You build 00:01:15.560 --> 00:01:18.530 a vault and you think the vault is secure. But 00:01:18.530 --> 00:01:21.349 that protection just failed spectacular. It really 00:01:21.349 --> 00:01:24.590 did. Enter a developer named Kai Gomez. He's 00:01:24.590 --> 00:01:27.409 a 22 -year -old college dropout. He just looked 00:01:27.409 --> 00:01:29.189 at the architecture and cracked it. I mean, he 00:01:29.189 --> 00:01:31.709 reverse engineered the entire Claude Mythos system 00:01:31.709 --> 00:01:34.090 right from the outside. And he released his own 00:01:34.090 --> 00:01:36.030 open source clone. He called it Open Mythos. 00:01:36.049 --> 00:01:39.609 He did this in less than two weeks. Beat. I just 00:01:39.609 --> 00:01:41.469 think about the historical irony here. Oh, it's 00:01:41.469 --> 00:01:44.590 massive. Massive corporations build these giant 00:01:44.590 --> 00:01:48.230 impenetrable digital castles and then... A kid 00:01:48.230 --> 00:01:50.569 just invents a ladder. A very, very efficient 00:01:50.569 --> 00:01:52.769 ladder. The unfiltered version is now effectively 00:01:52.769 --> 00:01:55.590 out in the wild. This is an absolute nightmare 00:01:55.590 --> 00:01:58.950 for AI regulators everywhere. Yeah, I bet. Terrifies 00:01:58.950 --> 00:02:01.109 the corporate moat defenders across the industry. 00:02:01.329 --> 00:02:03.430 So why is it so terrifying for them strategically? 00:02:03.730 --> 00:02:06.370 What did this kid actually figure out? It comes 00:02:06.370 --> 00:02:09.930 down to a profound architecture shift. Older 00:02:09.930 --> 00:02:13.650 AI models use massive flat layers to process 00:02:13.650 --> 00:02:16.729 data. Okay. They process information sequentially 00:02:16.729 --> 00:02:19.110 through billions of different parameters. Let's 00:02:19.110 --> 00:02:20.969 define parameters really quickly for everyone. 00:02:21.409 --> 00:02:23.710 Parameters are basically the tiny mathematical 00:02:23.710 --> 00:02:26.870 dials controlling how software thinks. Exactly. 00:02:26.870 --> 00:02:29.849 And to make an AI smarter, you usually just add 00:02:29.849 --> 00:02:32.870 more layers. You add more dials. Right. But that 00:02:32.870 --> 00:02:34.949 requires a massive amount of computing power. 00:02:36.039 --> 00:02:38.620 OpenMythos fundamentally shifts how the AI actually 00:02:38.620 --> 00:02:41.659 operates. It uses a very small subset of layers 00:02:41.659 --> 00:02:44.379 instead. Then it loops the data through them 00:02:44.379 --> 00:02:46.879 repeatedly. So it reuses the exact same digital 00:02:46.879 --> 00:02:49.000 pathways over and over. It doesn't need to travel 00:02:49.000 --> 00:02:52.099 down a massive flat line. Yes, that is the brilliant 00:02:52.099 --> 00:02:55.080 part of the design. It is exactly like stacking 00:02:55.080 --> 00:02:58.099 Lego blocks of data. You build incredibly complex 00:02:58.099 --> 00:03:00.460 towering structures from a few simple repeating 00:03:00.460 --> 00:03:02.919 pieces. Wow. It is highly efficient and it's 00:03:02.919 --> 00:03:05.479 incredibly fast. Whoa. Imagine scaling to a billion 00:03:05.479 --> 00:03:07.280 queries. If it's that efficient, you could run 00:03:07.280 --> 00:03:09.639 frontier intelligence on consumer hardware. Yep. 00:03:09.699 --> 00:03:11.280 You wouldn't even need a massive server farm. 00:03:11.460 --> 00:03:14.800 And that is the trillion dollar problem for big 00:03:14.800 --> 00:03:17.539 tech right now. We literally just saw big tech 00:03:17.539 --> 00:03:20.669 commit one. trillion dollars to infrastructure. 00:03:20.990 --> 00:03:23.530 It's staggering. They are building massive power 00:03:23.530 --> 00:03:26.229 hungry data centers all over the world. Just 00:03:26.229 --> 00:03:30.110 massive physical warehouses filled with extremely 00:03:30.110 --> 00:03:33.030 expensive computer chips. They are pouring concrete 00:03:33.030 --> 00:03:36.090 and laying copper wire everywhere. Yes. And the 00:03:36.090 --> 00:03:38.330 financial markets are aggressively backing this 00:03:38.330 --> 00:03:42.150 approach to KKR just raised 10 billion dollars 00:03:42.150 --> 00:03:45.090 for a new fund. They are backing a massive new 00:03:45.090 --> 00:03:48.460 A .I. infrastructure company led by Adams. Lipsky, 00:03:48.560 --> 00:03:51.180 the former AWS chief. Exactly. They are focusing 00:03:51.180 --> 00:03:54.199 entirely on data centers, power, and connectivity. 00:03:54.500 --> 00:03:56.539 They're building alongside the massive cloud 00:03:56.539 --> 00:03:58.879 computing companies. They are betting everything 00:03:58.879 --> 00:04:01.240 on centralized power. To support the constantly 00:04:01.240 --> 00:04:04.000 rising AI demand. Right. But here's the massive 00:04:04.000 --> 00:04:07.900 catch in their strategy. What if small and loopy 00:04:07.900 --> 00:04:10.759 is fundamentally better architecture? What if 00:04:10.759 --> 00:04:13.360 it simply beats big and flat forever? Then that 00:04:13.360 --> 00:04:15.240 massive hardware investment suddenly hits an 00:04:15.240 --> 00:04:17.399 efficiency wall. You don't need a massive data... 00:04:17.420 --> 00:04:19.860 center if a laptop works. It's a catastrophic 00:04:19.860 --> 00:04:22.879 efficiency wall for those companies. Gomez is 00:04:22.879 --> 00:04:25.199 actively building a decentralized version of 00:04:25.199 --> 00:04:28.000 superintelligence. Meanwhile, Meta and Google 00:04:28.000 --> 00:04:31.040 are spending billions to centralize it. The next 00:04:31.040 --> 00:04:33.560 world -changing breakthrough probably won't come 00:04:33.560 --> 00:04:36.019 from a boardroom. It will likely come directly 00:04:36.019 --> 00:04:39.759 from a GitHub repository. Two sec silence. Is 00:04:39.759 --> 00:04:42.399 Big Tech's hardware advantage officially dead? 00:04:42.860 --> 00:04:45.339 It's not totally dead, but it's deeply vulnerable 00:04:45.339 --> 00:04:48.259 now. Proprietary hardware is no longer an absolute 00:04:48.259 --> 00:04:51.379 shield. Self -taught developers can clone frontier 00:04:51.379 --> 00:04:54.300 architectures in days. So raw computing power 00:04:54.300 --> 00:04:57.000 can't protect them from smarter, leaner algorithms. 00:04:57.699 --> 00:05:00.060 Mid -role sponsor, Reed. We just talked about 00:05:00.060 --> 00:05:02.060 how these powerful, lean models are built. Now 00:05:02.060 --> 00:05:03.519 let's look at where they are actually going. 00:05:03.579 --> 00:05:05.040 They are actively sitting in the room with us 00:05:05.040 --> 00:05:07.139 today. Yeah, they really are. They are diagnosing 00:05:07.139 --> 00:05:09.000 our bodies and running our corporate workflows. 00:05:09.120 --> 00:05:11.279 We are letting them into our most intimate spaces. 00:05:11.959 --> 00:05:14.379 Google DeepMind just unveiled something called 00:05:14.379 --> 00:05:17.240 the AI Co -Clinician. This is not just a text 00:05:17.240 --> 00:05:19.720 box on a screen anymore. It is a real -time, 00:05:19.800 --> 00:05:22.759 video -capable teammate in the exam room. It 00:05:22.759 --> 00:05:25.300 is designed to sit in on your actual doctor appointments. 00:05:25.579 --> 00:05:28.519 It literally watches the interaction unfold live. 00:05:29.259 --> 00:05:31.980 DeepMind is calling this new concept Triadic 00:05:31.980 --> 00:05:34.879 Care. Triadic meaning a core group of three. 00:05:35.240 --> 00:05:38.220 The patient, the doctor, and the AI working together. 00:05:39.639 --> 00:05:42.689 Beat. I mean, the AI isn't textbook anymore. 00:05:42.910 --> 00:05:45.350 It is a hyper observant medical resident who 00:05:45.350 --> 00:05:47.769 never blinks. That's a really great way to visualize 00:05:47.769 --> 00:05:50.509 it. It actually watches your breathing during 00:05:50.509 --> 00:05:53.569 the physical exam. It monitors how your chest 00:05:53.569 --> 00:05:56.779 rises and falls through the camera. Wow. It even 00:05:56.779 --> 00:05:58.959 checks your inhaler technique on video. Right. 00:05:59.180 --> 00:06:00.759 Make sure you're holding it at the exact right 00:06:00.759 --> 00:06:03.019 angle. Exactly. It actively helps your doctor 00:06:03.019 --> 00:06:05.660 spot things the human eye might miss. And to 00:06:05.660 --> 00:06:08.319 achieve this safely, it operates using two very 00:06:08.319 --> 00:06:11.279 distinct modules. Okay. First, there is the talker 00:06:11.279 --> 00:06:13.060 module. And then that is the part that actually 00:06:13.060 --> 00:06:14.660 interacts with the patient. It retrieves the 00:06:14.660 --> 00:06:16.540 medical evidence and speaks in the room. Yes. 00:06:16.600 --> 00:06:19.160 But the second part is the real engineering breakthrough 00:06:19.160 --> 00:06:21.860 here. It is a completely silent overseer module 00:06:21.860 --> 00:06:24.439 running in the background. What does the overseer 00:06:24.439 --> 00:06:27.199 actually do mechanically? How does it prevent 00:06:27.199 --> 00:06:30.560 the AI from just making things up? It constantly 00:06:30.560 --> 00:06:33.139 watches the talker module before it speaks. It 00:06:33.139 --> 00:06:36.019 is specifically looking for any unsafe medical 00:06:36.019 --> 00:06:38.779 advice. So every piece of advice gets double 00:06:38.779 --> 00:06:41.019 -checked instantly. It is cross -referenced directly 00:06:41.019 --> 00:06:43.800 against clinical -grade evidence. Then it is 00:06:43.800 --> 00:06:46.540 cited in real time for the human doctor to verify. 00:06:46.759 --> 00:06:48.920 The test results on this system are honestly 00:06:48.920 --> 00:06:52.620 staggering. They tested it on 98 highly realistic 00:06:52.620 --> 00:06:55.560 primary care queries. The system logged zero 00:06:55.560 --> 00:06:58.759 critical errors in 97 of them. Wait, zero errors? 00:06:59.399 --> 00:07:01.620 It actually beat every other frontier model on 00:07:01.620 --> 00:07:03.680 open -ended drug questions. That even includes 00:07:03.680 --> 00:07:07.339 the brand new GPT -5 .4. It did. And the video 00:07:07.339 --> 00:07:09.360 simulations they ran are even more impressive 00:07:09.360 --> 00:07:12.519 to watch. The AI correctly guided patients through 00:07:12.519 --> 00:07:15.279 complex physical shoulder maneuvers. It helped 00:07:15.279 --> 00:07:17.819 them isolate and find specific rotator cuff injuries. 00:07:18.199 --> 00:07:20.500 It even helped diagnose a case of myasthenia 00:07:20.500 --> 00:07:23.480 gravis. It did this purely through video observation 00:07:23.480 --> 00:07:26.189 of the patient's face. In blind clinical tests, 00:07:26.529 --> 00:07:29.089 physicians vastly preferred this AI teammate. 00:07:29.430 --> 00:07:32.029 They chose it over their current diagnostic tools 00:07:32.029 --> 00:07:36.149 by 67 % to 26%. But the human element is still 00:07:36.149 --> 00:07:38.589 vital here. The study clearly found that experienced 00:07:38.589 --> 00:07:42.449 physicians still beat the AI overall. The human 00:07:42.449 --> 00:07:45.089 intuition combined with experience still wins 00:07:45.089 --> 00:07:48.649 out. For now, yes. But this triad concept is 00:07:48.649 --> 00:07:51.410 also rapidly entering the corporate world. OpenAI 00:07:51.410 --> 00:07:53.730 just launched a product called Codex for Work. 00:07:54.110 --> 00:07:56.509 It is designed specifically for enterprise teams. 00:07:56.769 --> 00:07:59.370 Aaron Livy over at Box is already hiring agent 00:07:59.370 --> 00:08:02.029 engineers. These are specialized people who build 00:08:02.029 --> 00:08:05.029 AI to autonomously do tasks. He wants to automate 00:08:05.029 --> 00:08:08.259 core business workflows entirely. Two sec silence. 00:08:08.699 --> 00:08:11.100 How does the system prevent the AI from giving 00:08:11.100 --> 00:08:13.500 confident but deadly medical advice? It uses 00:08:13.500 --> 00:08:16.800 that critical silent overseer module. This module 00:08:16.800 --> 00:08:19.079 watches the main system and cross -references 00:08:19.079 --> 00:08:21.259 advice against clinical grade evidence in real 00:08:21.259 --> 00:08:23.759 time. It basically has an internal chaperone 00:08:23.759 --> 00:08:26.439 fact -checking every single word. Exactly. The 00:08:26.439 --> 00:08:28.319 architecture is designed entirely around patient 00:08:28.319 --> 00:08:30.980 safety first. Which brings us to the messy part 00:08:30.980 --> 00:08:34.029 of this transition. Integrating agents that can 00:08:34.029 --> 00:08:36.950 see us and talk to us is incredibly jarring. 00:08:37.090 --> 00:08:40.070 It creates immediate painful friction with human 00:08:40.070 --> 00:08:42.750 nature. Oh, absolutely. It clashes with our privacy 00:08:42.750 --> 00:08:45.549 and it completely disrupts our old habits. Let's 00:08:45.549 --> 00:08:48.470 talk about those everyday habits first. Specifically, 00:08:48.509 --> 00:08:52.299 the frustrating issue of prompt drift. GPT 5 00:08:52.299 --> 00:08:55.899 .5 and Opus 4 .7 just released completely new 00:08:55.899 --> 00:08:58.679 architectures. And they completely broke everyone's 00:08:58.679 --> 00:09:01.039 old prompting habits overnight. The way you talk 00:09:01.039 --> 00:09:03.519 to the machine has to change. I still wrestle 00:09:03.519 --> 00:09:06.500 with prompt drift myself. You spend months learning 00:09:06.500 --> 00:09:09.019 how to talk to a specific model. You find the 00:09:09.019 --> 00:09:11.360 absolute perfect phrasing to get what you want. 00:09:11.440 --> 00:09:13.179 Right, you dial it in perfectly. And then they 00:09:13.179 --> 00:09:15.379 seamlessly update the software on the back end. 00:09:15.759 --> 00:09:18.259 Suddenly, your carefully crafted prompts feel 00:09:18.259 --> 00:09:21.320 significantly worse. The output gets lazy or 00:09:21.320 --> 00:09:23.580 weird. But it's not you. It's the underlying 00:09:23.580 --> 00:09:26.440 model changing how it interprets language. The 00:09:26.440 --> 00:09:28.519 rules of engagement simply shift overnight without 00:09:28.519 --> 00:09:31.019 any warning. The cost equations are also shifting 00:09:31.019 --> 00:09:33.960 incredibly rapidly right now. A new evaluation 00:09:33.960 --> 00:09:37.080 tool just launched called Best Value AI 2026. 00:09:37.500 --> 00:09:40.860 It actively compares over 37 different language 00:09:40.860 --> 00:09:43.360 models. It looks across APIs, subscriptions, 00:09:43.539 --> 00:09:45.539 and different hardware setups. And for those 00:09:45.539 --> 00:09:47.940 wondering, APIs are just software bridges connecting 00:09:47.940 --> 00:09:49.980 different computer programs together. Right. 00:09:50.059 --> 00:09:52.919 The tool ranks all these models by quality adjusted 00:09:52.919 --> 00:09:56.100 cost per token. And tokens are simply the small 00:09:56.100 --> 00:09:59.679 chunks of words an AI reads. It runs real tests 00:09:59.679 --> 00:10:03.120 to find the actual value. Because a cheap model 00:10:03.120 --> 00:10:05.159 is useless if the answers are terrible. Exactly. 00:10:05.340 --> 00:10:08.440 But the friction goes way beyond cost and prompting 00:10:08.440 --> 00:10:11.679 habits. It is bleeding into deep, deeply uncomfortable 00:10:11.679 --> 00:10:14.159 privacy issues. Look at what just happened over 00:10:14.159 --> 00:10:17.100 at Meta. Meta is facing massive backlash right 00:10:17.100 --> 00:10:19.379 now. They are under heavy public scrutiny regarding 00:10:19.379 --> 00:10:21.480 their hardware. Yeah, reports just came out about 00:10:21.480 --> 00:10:24.580 their Ray -Ban Meta smart glasses. Workers actually 00:10:24.580 --> 00:10:27.320 reviewed sensitive naked footage captured from 00:10:27.320 --> 00:10:29.840 users' glasses. The staff members involved ultimately 00:10:29.840 --> 00:10:32.379 lost their jobs over it. But the privacy line 00:10:32.379 --> 00:10:34.720 was already crossed. The cameras are always watching. 00:10:34.899 --> 00:10:37.460 It perfectly shows how technological capabilities 00:10:37.460 --> 00:10:40.879 have completely outpaced our social norms. We 00:10:40.879 --> 00:10:43.519 wear cameras on our faces, but we don't understand 00:10:43.519 --> 00:10:46.559 the risks. And our legal frameworks are really 00:10:46.559 --> 00:10:49.740 struggling to keep up. Copyright law is also 00:10:49.740 --> 00:10:52.519 buckling heavily under the pressure of AI scraping. 00:10:53.039 --> 00:10:55.539 Casey Green is considering legal action right 00:10:55.539 --> 00:10:58.429 now. He is the original creator of the famous 00:10:58.429 --> 00:11:02.149 this is fine dog meme. It's a massive piece of 00:11:02.149 --> 00:11:04.710 Internet culture. An AI startup called Artisan 00:11:04.710 --> 00:11:08.590 basically used his art. They generated it and 00:11:08.590 --> 00:11:10.850 put it in their ads without his consent. It is 00:11:10.850 --> 00:11:13.830 a constant ongoing battle over ownership and 00:11:13.830 --> 00:11:16.090 training data. Artists are fighting to keep their 00:11:16.090 --> 00:11:18.289 work out of the machine. And then there's the 00:11:18.289 --> 00:11:20.509 emotional friction. This might honestly be the 00:11:20.509 --> 00:11:22.720 strangest part of the whole transition. Teens 00:11:22.720 --> 00:11:25.220 are forming incredibly deep emotional ties to 00:11:25.220 --> 00:11:27.460 chatbots. Psychologists and experts are actually 00:11:27.460 --> 00:11:29.679 stepping in now. They recommend five specific 00:11:29.679 --> 00:11:32.240 questions you should ask your kids. You need 00:11:32.240 --> 00:11:34.100 to gauge how they are interacting with these 00:11:34.100 --> 00:11:36.480 AI companions. People are genuinely treating 00:11:36.480 --> 00:11:38.840 them like close friends or even romantic partners. 00:11:39.019 --> 00:11:41.639 It's fundamentally changing human socialization. 00:11:42.090 --> 00:11:44.230 Even the tech leaders themselves are treating 00:11:44.230 --> 00:11:47.049 them differently now. Sam Altman recently tested 00:11:47.049 --> 00:11:50.909 the brand new GPT 5 .5 model. He asked the AI 00:11:50.909 --> 00:11:53.450 to plan its own launch party, right? He asked 00:11:53.450 --> 00:11:55.090 it what kind of celebration it wanted. Yeah. 00:11:55.250 --> 00:11:58.529 And the results were oddly specific. The AI provided 00:11:58.529 --> 00:12:01.730 highly structured, almost uncomfortably human 00:12:01.730 --> 00:12:04.289 -like answers about the vibe and the guest list. 00:12:04.490 --> 00:12:06.730 And he is actually going to do it. He is throwing 00:12:06.730 --> 00:12:10.149 the exact party the AI designed for itself. Two 00:12:10.149 --> 00:12:13.379 sec silence. With teens forming attachments and 00:12:13.379 --> 00:12:16.620 AI throwing its own parties, are we anthropomorphizing 00:12:16.620 --> 00:12:18.600 these models too much? Well, these systems are 00:12:18.600 --> 00:12:20.940 explicitly designed to mimic empathy perfectly. 00:12:21.340 --> 00:12:24.340 It makes human attachment an inevitable, complicated 00:12:24.340 --> 00:12:27.100 feature of the interface, not a bug. We are wired 00:12:27.100 --> 00:12:29.080 to love anything that talks back to us. We really 00:12:29.080 --> 00:12:31.120 are. It hijacks our basic social programming 00:12:31.120 --> 00:12:34.039 entirely. Get it? If we step back and look at 00:12:34.039 --> 00:12:36.259 the big picture today. We are watching a fundamental 00:12:36.259 --> 00:12:39.840 transition in computing history. AI used to be 00:12:39.840 --> 00:12:42.840 an expensive centralized tool. You just type 00:12:42.840 --> 00:12:44.980 text into a simple box. Right. It was distant. 00:12:45.039 --> 00:12:47.100 It was entirely contained within the web browser. 00:12:47.549 --> 00:12:50.450 Exactly. But now it is becoming a decentralized, 00:12:50.950 --> 00:12:54.409 highly capable video entity. It literally watches 00:12:54.409 --> 00:12:56.870 your breathing in the exam room. It automates 00:12:56.870 --> 00:12:58.789 your entire business workflow in the background. 00:12:58.990 --> 00:13:01.470 And it constantly breaks your old prompting habits 00:13:01.470 --> 00:13:03.669 as it evolves. Thank you for joining us on this 00:13:03.669 --> 00:13:05.330 deep dive. We want to leave you with a final 00:13:05.330 --> 00:13:08.070 thought to mull over. A 22 -year -old can unleash 00:13:08.070 --> 00:13:11.129 a super intelligence from a laptop. An AI can 00:13:11.129 --> 00:13:13.330 act as a real -time medical teammate with a perfect 00:13:13.330 --> 00:13:16.009 bedside manner. What happens to the value of 00:13:16.009 --> 00:13:18.309 human intuition in a world where logic and empathy 00:13:18.309 --> 00:13:21.210 can both be downloaded from GitHub? Outero Music.