WEBVTT 00:00:00.000 --> 00:00:02.339 Just two years ago, human beings wrote software 00:00:02.339 --> 00:00:05.339 code manually. We typed every single logical 00:00:05.339 --> 00:00:08.199 instruction by hand. Yeah, line by tedious line. 00:00:08.339 --> 00:00:11.460 Right. Today, we are building autonomous AI managers 00:00:11.460 --> 00:00:14.699 instead. Those managers hire other AI workers 00:00:14.699 --> 00:00:17.519 to do it. Welcome to our latest deep dive. I 00:00:17.519 --> 00:00:20.320 am truly glad you are here with us. Okay, let's 00:00:20.320 --> 00:00:23.660 unpack this. Today, we explore a profound technological 00:00:23.660 --> 00:00:27.440 shift. We are moving from human prompts to continuous 00:00:27.440 --> 00:00:30.140 agentic loops. It is a massive leap forward. 00:00:30.260 --> 00:00:32.780 It really is. We will trace the staggering hardware 00:00:32.780 --> 00:00:35.719 costs of this autonomy. We will see how it enters 00:00:35.719 --> 00:00:38.259 Hollywood and the beauty industry. Finally, we 00:00:38.259 --> 00:00:40.460 examine a plan to shield the Internet's foundation. 00:00:40.880 --> 00:00:43.600 We are crossing into completely uncharted territory 00:00:43.600 --> 00:00:46.600 today. The fundamental architecture of digital 00:00:46.600 --> 00:00:49.140 creation is transforming. It is happening much 00:00:49.140 --> 00:00:51.560 faster than anyone originally predicted. We should 00:00:51.560 --> 00:00:54.259 start with how work actually gets done now. AI 00:00:54.259 --> 00:00:57.060 is operating autonomously in continuous, invisible 00:00:57.060 --> 00:01:00.399 loops. Beat? At Meta's At Scale conference, we 00:01:00.399 --> 00:01:02.100 heard a fascinating announcement. Right, from 00:01:02.100 --> 00:01:05.079 Boris Cherny. He created Cloud Code. Exactly. 00:01:05.260 --> 00:01:08.620 He officially declared the era of agents prompting 00:01:08.620 --> 00:01:11.540 other agents. This changes the baseline of software 00:01:11.540 --> 00:01:13.980 development entirely. The evolution of software 00:01:13.980 --> 00:01:16.689 engineering is actually quite striking. First, 00:01:16.810 --> 00:01:19.069 we relied on humans writing handwritten code. 00:01:19.189 --> 00:01:22.409 Which, you know, took thousands of hours. Then 00:01:22.409 --> 00:01:25.349 we transitioned to asking AI agents to write 00:01:25.349 --> 00:01:28.349 code. That required complex prompt engineering 00:01:28.349 --> 00:01:31.049 and constant human oversight. Prompt engineering 00:01:31.049 --> 00:01:33.670 is rapidly becoming an obsolete skill. Yeah. 00:01:34.010 --> 00:01:37.349 It really is. We used to spend hours perfectly 00:01:37.349 --> 00:01:40.409 phrasing our technical requests. Now the AI agent 00:01:40.409 --> 00:01:42.629 optimizes its own internal prompt structure. 00:01:43.090 --> 00:01:45.650 Human engineers just set up managers to handle 00:01:45.650 --> 00:01:48.750 tasks. Those managers then prompt other sub -agents 00:01:48.750 --> 00:01:51.829 to complete specific jobs. These autonomous processes 00:01:51.829 --> 00:01:54.670 are formally known as agentic loops. They rely 00:01:54.670 --> 00:01:57.269 entirely on complex, non -deterministic logic. 00:01:57.930 --> 00:01:59.829 Let me define that specific technical term for 00:01:59.829 --> 00:02:02.390 you quickly. An AI making its own choices instead 00:02:02.390 --> 00:02:04.650 of following fixed rules. Right. It doesn't follow 00:02:04.650 --> 00:02:06.909 rigid tracks like a train anymore. It is more 00:02:06.909 --> 00:02:09.330 like a self -driving car encountering a detour. 00:02:09.509 --> 00:02:12.550 It calculates a brand new route dynamically in 00:02:12.550 --> 00:02:15.650 real time. It evaluates the physical obstacles. 00:02:16.069 --> 00:02:18.810 Right. And adjusts its path continuously. Czerny 00:02:18.810 --> 00:02:20.430 noted this is already happening professionally. 00:02:20.710 --> 00:02:23.090 He has agents constantly running in his system's 00:02:23.090 --> 00:02:25.430 background. Just constantly working. Exactly. 00:02:26.060 --> 00:02:28.620 They scan his entire code base for outdated legacy 00:02:28.620 --> 00:02:32.039 code. They unify duplicated abstractions and 00:02:32.039 --> 00:02:34.919 improve code architecture endlessly. They do 00:02:34.919 --> 00:02:37.060 this tedious maintenance work without any human 00:02:37.060 --> 00:02:39.400 prompting. I still wrestle with prompt drift 00:02:39.400 --> 00:02:42.960 myself. My AI forgets instructions during a long 00:02:42.960 --> 00:02:45.840 chat session. Goal. It loses the original context 00:02:45.840 --> 00:02:49.000 after 20 or 30 messages. How exactly do they 00:02:49.000 --> 00:02:51.360 keep these advanced models on track? Well, developers 00:02:51.360 --> 00:02:53.800 now use clever engineering tricks like the Ralph 00:02:53.800 --> 00:02:56.219 Loop. It is a brilliant mechanism for maintaining 00:02:56.219 --> 00:02:58.780 digital focus. How does it work? It forces the 00:02:58.780 --> 00:03:02.000 AI to output its reasoning aloud first. By reading 00:03:02.000 --> 00:03:04.639 its own logic step by step, the model self -corrects. 00:03:04.639 --> 00:03:06.840 It feeds its previous output back as the next 00:03:06.840 --> 00:03:09.719 input. So it prevents the autonomous model from 00:03:09.719 --> 00:03:13.159 hallucinating a wrong turn. It constantly grounds 00:03:13.159 --> 00:03:17.219 the AI in its own verified logic. Exactly. That 00:03:17.219 --> 00:03:20.139 is crucial for what engineers call hill -climbing 00:03:20.139 --> 00:03:23.379 problems. Meaning an AI making incremental software 00:03:23.379 --> 00:03:26.580 improvements endlessly. Right. It optimizes complex 00:03:26.580 --> 00:03:29.759 architecture as long as your budget allows. Anthropic 00:03:29.759 --> 00:03:32.439 recently showed a truly powerful demo of this 00:03:32.439 --> 00:03:36.500 concept. Their engineers used AI to build software 00:03:36.500 --> 00:03:39.270 entirely from scratch. Not just asking it to 00:03:39.270 --> 00:03:42.090 build a whole app lazily. Exactly. They guide 00:03:42.090 --> 00:03:43.969 the agents through the complex development process 00:03:43.969 --> 00:03:46.849 iteratively. There is an emerging tool called 00:03:46.849 --> 00:03:50.110 SkyBridge doing exactly this. It handles the 00:03:50.110 --> 00:03:52.629 complex development loop entirely on its own. 00:03:52.909 --> 00:03:55.030 Meaning it spins up a virtual network and tests 00:03:55.030 --> 00:03:57.050 the written code? Yeah, it establishes secure 00:03:57.050 --> 00:03:59.449 testing tunnels to evaluate external webhooks 00:03:59.449 --> 00:04:02.150 safely. It manages view rendering and client 00:04:02.150 --> 00:04:04.569 compatibility automatically behind scenes. So 00:04:04.569 --> 00:04:07.110 human engineers just focus on designing the core 00:04:07.110 --> 00:04:09.289 product features. It's like stacking Lego blocks 00:04:09.289 --> 00:04:11.770 of data. That is a great analogy. But how does 00:04:11.770 --> 00:04:13.990 a non -deterministic loop actually know when 00:04:13.990 --> 00:04:16.029 to stop? If it runs in the background endlessly, 00:04:16.310 --> 00:04:18.490 what kills it? It seems like it would just optimize 00:04:18.490 --> 00:04:21.279 code forever. Well, developers program a separate 00:04:21.279 --> 00:04:23.759 independent subagent to monitor the process. 00:04:24.180 --> 00:04:26.759 This subagent compares mathematical confidence 00:04:26.759 --> 00:04:30.120 scores against a predefined threshold. Ah, I 00:04:30.120 --> 00:04:32.319 see. Yeah, it constantly evaluates the current 00:04:32.319 --> 00:04:35.040 output against the primary goal. When the score 00:04:35.040 --> 00:04:38.439 passes the success metric, it cuts power. It 00:04:38.439 --> 00:04:40.959 acts as an objective independent auditor for 00:04:40.959 --> 00:04:43.879 the working agent. So an AI manager simply decides 00:04:43.879 --> 00:04:46.259 when the job is finally done. Yeah, it becomes 00:04:46.259 --> 00:04:48.920 a completely self -regulating digital ecosystem. 00:04:49.629 --> 00:04:52.149 The human becomes a supervisor rather than a 00:04:52.149 --> 00:04:54.750 direct creator. Infinite loops are undeniably 00:04:54.750 --> 00:04:57.269 amazing in theory, but they require infinite 00:04:57.269 --> 00:04:59.930 guardrails to remain perfectly safe. Two secs 00:04:59.930 --> 00:05:02.250 silence. They also require practically unlimited 00:05:02.250 --> 00:05:04.930 computing power to function properly. This brings 00:05:04.930 --> 00:05:07.009 us to the immense financial cost of autonomy. 00:05:07.290 --> 00:05:09.529 These continuous loops burn through processing 00:05:09.529 --> 00:05:12.949 tokens insanely fast. Every single loop cycle 00:05:12.949 --> 00:05:15.790 costs real money to execute. There is practically 00:05:15.790 --> 00:05:18.449 no theoretical compute ceiling here at all. This 00:05:18.449 --> 00:05:20.910 raises an urgent question about evaluation and 00:05:20.910 --> 00:05:23.310 software testing. Before you run infinite loops, 00:05:23.529 --> 00:05:26.110 you need strong safety nets. You cannot afford 00:05:26.110 --> 00:05:29.149 to let an agent hallucinate endlessly. There 00:05:29.149 --> 00:05:32.490 is a new tool called AgentX for this. It evaluates 00:05:32.490 --> 00:05:34.670 these agents and creates comprehensive test suites. 00:05:34.810 --> 00:05:36.490 Before they fail in production, right? Exactly. 00:05:36.750 --> 00:05:39.310 It runs intense evaluations before agents go 00:05:39.310 --> 00:05:42.709 live. It simulates thousands of random user interactions 00:05:42.709 --> 00:05:45.949 simultaneously. This helps pinpoint catastrophic 00:05:45.949 --> 00:05:50.879 edge case failures. incredibly early. But better 00:05:50.879 --> 00:05:52.639 software testing does not solve the hardware 00:05:52.639 --> 00:05:55.639 bottleneck. If loops just burn tokens indefinitely, 00:05:55.879 --> 00:05:58.540 what happens next? Aren't we just building a 00:05:58.540 --> 00:06:01.019 machine designed to bankrupt developers? It sounds 00:06:01.019 --> 00:06:03.279 like a severe financial drain for smaller teams. 00:06:03.540 --> 00:06:06.860 Whoa, imagine scaling to a billion queries. The 00:06:06.860 --> 00:06:09.379 hardware compute costs become absolutely astronomical 00:06:09.379 --> 00:06:12.639 very quickly. This is precisely why Grok recently 00:06:12.639 --> 00:06:15.839 raised $650 million. They were challenging the 00:06:15.839 --> 00:06:18.060 current graphical processing paradigms directly 00:06:18.060 --> 00:06:21.180 now. NVIDIA also completed a $20 billion deal 00:06:21.180 --> 00:06:23.720 recently. Right. A massive not -acquire -hire 00:06:23.720 --> 00:06:27.180 deal. They absorbed key engineering talent without 00:06:27.180 --> 00:06:30.180 triggering antitrust monopoly laws. That deal 00:06:30.180 --> 00:06:33.620 also absorbed highly specific chip architecture 00:06:33.620 --> 00:06:36.170 intellectual property. Traditional AI models 00:06:36.170 --> 00:06:39.350 rely heavily on graphics processing units. Those 00:06:39.350 --> 00:06:42.290 chips process thousands of math problems simultaneously 00:06:42.290 --> 00:06:45.670 in parallel, but they suffer from severe memory 00:06:45.670 --> 00:06:49.110 bandwidth limitations during generation. Why 00:06:49.110 --> 00:06:51.730 is Grok's alternative hardware so crucial for 00:06:51.730 --> 00:06:54.129 infinite loops? They engineered what they call 00:06:54.129 --> 00:06:57.199 language processing units instead. These specific 00:06:57.199 --> 00:07:00.000 chips process sequential data without traditional 00:07:00.000 --> 00:07:02.240 memory bottlenecks. So they eliminate the memory 00:07:02.240 --> 00:07:04.459 bandwidth limitations that plague traditional 00:07:04.459 --> 00:07:06.959 graphics cards. Exactly. They skip the heavy 00:07:06.959 --> 00:07:08.980 graphics processing overhead entirely during 00:07:08.980 --> 00:07:11.920 operation. This makes running text -based AI 00:07:11.920 --> 00:07:15.120 models significantly faster and cheaper. Grok 00:07:15.120 --> 00:07:17.139 is spreading this specific architecture across 00:07:17.139 --> 00:07:20.279 13 data centers. They clearly see an enormous, 00:07:20.459 --> 00:07:22.899 unending demand for continuous compute. Will 00:07:22.899 --> 00:07:24.600 these severe hardware bottlenecks eventually 00:07:24.600 --> 00:07:26.600 kill the infinite loop dream? Are we hitting 00:07:26.600 --> 00:07:28.180 the physical limits of our digital ambitions? 00:07:28.660 --> 00:07:31.720 It certainly will, unless alternative clouds 00:07:31.720 --> 00:07:34.759 challenge the chip monopolies. Companies like 00:07:34.759 --> 00:07:37.899 Grok must successfully provide affordable, scalable 00:07:37.899 --> 00:07:40.540 computing power. They need to drive the cost 00:07:40.540 --> 00:07:43.370 of inference down dramatically. Otherwise, only 00:07:43.370 --> 00:07:46.209 tech giants can afford to run these models. Basically, 00:07:46.370 --> 00:07:49.509 endless AI needs endless chips or the whole thing 00:07:49.509 --> 00:07:53.529 stalls. Physical reality firmly limits our theoretical 00:07:53.529 --> 00:07:56.009 digital potential. To fuel these highly capable 00:07:56.009 --> 00:07:58.970 models, they need enormous memory. They also 00:07:58.970 --> 00:08:01.629 require constant ongoing observation of complex 00:08:01.629 --> 00:08:04.660 human behavior. beat. This collision directly 00:08:04.660 --> 00:08:07.240 impacts our foundational expectations of digital 00:08:07.240 --> 00:08:11.040 privacy. Google's Gemini 3 .5 Pro nears a June 00:08:11.040 --> 00:08:13.720 launch. It features deeply advanced reasoning 00:08:13.720 --> 00:08:16.319 capabilities for complex tasks. It also features 00:08:16.319 --> 00:08:18.959 a staggering 2 million token context window. 00:08:19.240 --> 00:08:21.480 Let me define that specific context window for 00:08:21.480 --> 00:08:23.519 you quickly. The amount of text the AI can remember 00:08:23.519 --> 00:08:25.930 in one conversation. A 2 million token window 00:08:25.930 --> 00:08:28.529 is practically an entire personal archive. Yeah, 00:08:28.610 --> 00:08:30.769 and pro and ultra tier users get this advanced 00:08:30.769 --> 00:08:33.649 access first. To capture proper human context, 00:08:33.889 --> 00:08:36.610 new software tracks everything constantly. A 00:08:36.610 --> 00:08:39.490 tool called ReadyWhen monitors your daily workplace 00:08:39.490 --> 00:08:42.210 decisions continuously. It indexes your daily 00:08:42.210 --> 00:08:45.250 Slack messages, your email, and meetings. Right. 00:08:45.330 --> 00:08:47.990 It builds a comprehensive vector database of 00:08:47.990 --> 00:08:50.690 your professional life. It drafts your necessary 00:08:50.690 --> 00:08:53.580 next steps automatically. based on behavior. 00:08:53.779 --> 00:08:56.500 Here's where it gets really interesting. Meta 00:08:56.500 --> 00:08:59.559 paused an internal AI training program recently. 00:09:00.159 --> 00:09:02.899 Employee personal data was exposed during the 00:09:02.899 --> 00:09:05.419 initial training process. The internal program 00:09:05.419 --> 00:09:07.620 tracked daily keystrokes and mouse movements 00:09:07.620 --> 00:09:10.000 continuously. They wanted to train models on 00:09:10.000 --> 00:09:12.100 how employees actually work. They wanted the 00:09:12.100 --> 00:09:15.340 AI to learn our natural human hesitations. Anthropic 00:09:15.340 --> 00:09:17.740 updated Claude's official privacy policy in a 00:09:17.740 --> 00:09:20.200 similar vein. They now strictly require identity 00:09:20.200 --> 00:09:22.919 checks for specific flagged users. This verification 00:09:22.919 --> 00:09:25.360 process includes scanning passports and official 00:09:25.360 --> 00:09:27.299 driver's licenses. It even requires uploading 00:09:27.299 --> 00:09:29.960 selfies and complex face geometry data. They 00:09:29.960 --> 00:09:32.379 map the physical topology of your face for security. 00:09:32.580 --> 00:09:34.779 If we connect this to the bigger picture, it 00:09:34.779 --> 00:09:38.980 feels invasive. Models clearly need deep, intimate 00:09:38.980 --> 00:09:42.399 access to be truly helpful. But that deep access 00:09:42.399 --> 00:09:46.120 creates enormous, terrifying new privacy vulnerabilities. 00:09:46.759 --> 00:09:49.240 There is a playful new tool online called In 00:09:49.240 --> 00:09:52.059 the Weights. It lets you check if large models 00:09:52.059 --> 00:09:54.379 learned your actual name. You can literally see 00:09:54.379 --> 00:09:56.639 where you rank in their training data. How does 00:09:56.639 --> 00:09:58.600 that tool actually work mathematically behind 00:09:58.600 --> 00:10:00.980 the scenes? Well, it queries the model's latent 00:10:00.980 --> 00:10:04.200 space for specific token associations. It measures 00:10:04.200 --> 00:10:06.360 how strongly your name... connects to specific 00:10:06.360 --> 00:10:09.000 data vectors. If your name appears frequently, 00:10:09.200 --> 00:10:11.220 the mathematical connection is stronger. Are 00:10:11.220 --> 00:10:13.720 we trading our digital anonymity just for better 00:10:13.720 --> 00:10:16.100 autocomplete? Why do they need something as intimate 00:10:16.100 --> 00:10:19.379 as face geometry? As AI capabilities grow dangerously 00:10:19.379 --> 00:10:22.299 powerful, accountability becomes a break. Companies 00:10:22.299 --> 00:10:24.720 cannot allow anonymous users to run autonomous 00:10:24.720 --> 00:10:27.279 infinite loops. Because a malicious infinite 00:10:27.279 --> 00:10:30.100 loop could launch devastating cyber attacks continuously. 00:10:30.500 --> 00:10:32.879 Exactly. They need to know exactly who deployed 00:10:32.879 --> 00:10:35.860 the destructive code. Verified human accountability 00:10:35.860 --> 00:10:38.740 is the ultimate emergency stop button. Bigger 00:10:38.740 --> 00:10:41.740 memory means bigger risks. Forcing companies 00:10:41.740 --> 00:10:44.779 to demand your actual face. Right. And the glorious 00:10:44.779 --> 00:10:47.720 anonymity of the early Internet disappears. All 00:10:47.720 --> 00:10:50.340 this highly personalized data isn't staying confined 00:10:50.340 --> 00:10:54.580 to dashboards. Beat. It is actively moving into 00:10:54.580 --> 00:10:57.759 our creative and physical realities. It is rapidly 00:10:57.759 --> 00:11:00.059 moving from the back office to the movie set. 00:11:00.179 --> 00:11:03.399 That is great analogy. Google is currently investing 00:11:03.399 --> 00:11:07.799 $75 million into A24. That is the brilliant studio 00:11:07.799 --> 00:11:10.320 behind everything everywhere all at once. Right. 00:11:10.399 --> 00:11:12.620 They are famous for their incredibly specific, 00:11:13.000 --> 00:11:16.779 surreal visual aesthetics. Google wants to build 00:11:16.779 --> 00:11:19.809 powerful new AI filmmaking tools together. Because 00:11:19.809 --> 00:11:21.990 algorithms are excellent in generation, but they 00:11:21.990 --> 00:11:24.370 lack actual taste. Yeah, they tend to create 00:11:24.370 --> 00:11:27.509 sterile, plastic -looking visual content by default. 00:11:27.750 --> 00:11:30.110 So Google is using direct feedback from real 00:11:30.110 --> 00:11:32.269 Hollywood artists. They want to ensure the tools 00:11:32.269 --> 00:11:34.929 serve the true creative process. The beauty industry 00:11:34.929 --> 00:11:37.049 has historically relied entirely on physical 00:11:37.049 --> 00:11:39.330 products. Now they are transforming into massive 00:11:39.330 --> 00:11:42.070 digital technology companies, essentially. L 00:11:42.070 --> 00:11:44.549 'Oreal recently partnered with OpenAI for a highly 00:11:44.549 --> 00:11:47.730 visual project. They revealed an incredible Maybelline 00:11:47.730 --> 00:11:50.940 virtual makeup. try -on tool recently. Which 00:11:50.940 --> 00:11:53.700 they showcased at VivaTech 2026, right? Yeah. 00:11:53.799 --> 00:11:56.919 You will soon use this tool directly inside the 00:11:56.919 --> 00:11:59.940 chat GPT interface. They understand that digital 00:11:59.940 --> 00:12:02.279 appearance is becoming increasingly important 00:12:02.279 --> 00:12:05.179 today. The Maybelline tool analyzes the unique 00:12:05.179 --> 00:12:08.059 geometry of your face perfectly. It applies digital 00:12:08.059 --> 00:12:10.440 makeup that reacts perfectly to virtual lighting. 00:12:10.639 --> 00:12:13.220 You can test makeup virtually before buying the 00:12:13.220 --> 00:12:15.480 physical product. There's also a marketing tool 00:12:15.480 --> 00:12:19.850 called Align 2 .0. It captures exact micro -brand 00:12:19.850 --> 00:12:22.570 details for modern targeted marketing campaign. 00:12:22.850 --> 00:12:25.250 It learns the specific typography and hex codes 00:12:25.250 --> 00:12:27.830 of a brand. It generates perfectly on -brand 00:12:27.830 --> 00:12:30.250 social assets and digital advertisements easily. 00:12:30.450 --> 00:12:32.950 It is kind of incredibly efficient for large 00:12:32.950 --> 00:12:36.269 creative teams. Does Google paying A24 mean AI 00:12:36.269 --> 00:12:38.889 officially replaces human artists? Or does it 00:12:38.889 --> 00:12:41.149 prove algorithms desperately need our aesthetic 00:12:41.149 --> 00:12:43.350 judgment? It seems like they're admitting their 00:12:43.350 --> 00:12:45.850 models lack artistic soul. Developers finally 00:12:45.850 --> 00:12:48.210 realize algorithms lack true human aesthetic 00:12:48.210 --> 00:12:51.129 taste entirely. They're essentially buying access 00:12:51.129 --> 00:12:53.990 to artists to train aesthetics. They need human 00:12:53.990 --> 00:12:56.509 creativity to guide the enormous computing power. 00:12:56.750 --> 00:12:59.909 Right. Without human taste, the generated output 00:12:59.909 --> 00:13:02.649 remains completely sterile. They're paying artists 00:13:02.649 --> 00:13:05.409 to help build the exact tools they'll use, sponsor. 00:13:06.139 --> 00:13:08.820 If we are going to rely on AI for our movies 00:13:08.820 --> 00:13:11.480 and makeup, we desperately need to secure the 00:13:11.480 --> 00:13:14.159 open source code running underneath. Two sec 00:13:14.159 --> 00:13:16.940 silence. Open source software projects are the 00:13:16.940 --> 00:13:19.440 fundamental bedrock of commercial tech. They 00:13:19.440 --> 00:13:22.039 run our servers, our phones, and our global banking 00:13:22.039 --> 00:13:24.779 networks. But they are mostly run by unpaid, 00:13:24.960 --> 00:13:28.100 deeply passionate volunteers. This widespread 00:13:28.100 --> 00:13:31.039 decentralization creates huge, terrifying security 00:13:31.039 --> 00:13:34.059 gaps worldwide. OpenAI just announced a major 00:13:34.059 --> 00:13:36.399 new program called Patch the Planet. Which is 00:13:36.399 --> 00:13:39.899 a geeky nod to the 1995 movie hackers. OpenAI 00:13:39.899 --> 00:13:42.259 officially teamed up with a cybersecurity firm 00:13:42.259 --> 00:13:44.980 called Trail of Bits. They are actively defending 00:13:44.980 --> 00:13:47.879 major open source projects using codec security. 00:13:48.279 --> 00:13:50.779 This is a fascinating approach to a deeply decentralized 00:13:50.779 --> 00:13:54.700 problem. The AI scans vast open source repositories 00:13:54.700 --> 00:13:57.220 for deeply hidden vulnerabilities. It understands 00:13:57.220 --> 00:13:59.320 the underlying logic of the complex software 00:13:59.320 --> 00:14:02.120 architecture. It flags potential security flaws 00:14:02.120 --> 00:14:05.059 that human reviewers easily miss. Do you remember 00:14:05.059 --> 00:14:08.320 the massive log4j security crisis a few years 00:14:08.320 --> 00:14:12.120 ago? A tiny unnoticed piece of open source code 00:14:12.120 --> 00:14:15.220 broke half the internet. So what does this all 00:14:15.220 --> 00:14:18.210 mean? That is exactly the kind of nightmare scenario 00:14:18.210 --> 00:14:21.649 Trail of Bits targets. Human reviewers cannot 00:14:21.649 --> 00:14:24.210 possibly audit millions of lines of volunteer 00:14:24.210 --> 00:14:27.769 code. The AI acts as an untiring, incredibly 00:14:27.769 --> 00:14:30.250 thorough security guard. But the human engineers 00:14:30.250 --> 00:14:32.669 at Trail of Bits carefully review the AI findings. 00:14:32.909 --> 00:14:34.970 Right. They handle all the difficult heavy lifting 00:14:34.970 --> 00:14:37.870 for the community. They translate the AI's complex 00:14:37.870 --> 00:14:40.929 math into human readable security patches. They 00:14:40.929 --> 00:14:43.250 actively work with volunteer projects to deploy 00:14:43.250 --> 00:14:46.259 these essential fixes. important software tests 00:14:46.259 --> 00:14:48.720 and build reusable workflows securely. And they 00:14:48.720 --> 00:14:51.039 do all this before project maintainers ever see 00:14:51.039 --> 00:14:53.779 the bug. It is an incredible technological irony 00:14:53.779 --> 00:14:56.279 when you really think about it. We deploy autonomous 00:14:56.279 --> 00:14:58.779 AI systems to shield our heavily decentralized 00:14:58.779 --> 00:15:02.220 internet. We use it to fix dangerous vulnerabilities 00:15:02.220 --> 00:15:04.440 we created ourselves. It is a wild cycle for 00:15:04.440 --> 00:15:07.659 sure. Does relying on open AI to find bugs create 00:15:07.659 --> 00:15:11.080 a single point of failure? Is it bad for internet 00:15:11.080 --> 00:15:14.690 security to rely on one corporate entity? Human 00:15:14.690 --> 00:15:17.629 engineers at Trail of Bits act as a vital buffer. 00:15:17.889 --> 00:15:21.330 They carefully review abstract syntax trees before 00:15:21.330 --> 00:15:24.149 pushing patches live. They run the suggested 00:15:24.149 --> 00:15:26.789 fixes in isolated sandbox environments first. 00:15:27.279 --> 00:15:29.379 This ensures the AI model does not accidentally 00:15:29.379 --> 00:15:32.240 introduce new backdoors. We cannot trust a machine 00:15:32.240 --> 00:15:34.700 to rewrite our security protocols blindly. Exactly. 00:15:34.700 --> 00:15:37.200 The machine acts as an incredibly powerful digital 00:15:37.200 --> 00:15:40.039 magnifying glass. But the human engineer must 00:15:40.039 --> 00:15:42.480 remain the ultimate decision maker. Humans still 00:15:42.480 --> 00:15:45.379 verify the AI's math before the fix goes live. 00:15:45.480 --> 00:15:48.360 Right. And that expert human oversight is absolutely 00:15:48.360 --> 00:15:50.450 critical currently. We have journeyed through 00:15:50.450 --> 00:15:53.269 an incredible, profound technological shift today. 00:15:53.429 --> 00:15:55.950 We moved from manually writing simple code to 00:15:55.950 --> 00:15:59.029 swarms of agents. Now, autonomous digital managers 00:15:59.029 --> 00:16:01.750 build our complex software and our art. But this 00:16:01.750 --> 00:16:04.470 incredible autonomy requires absolutely unimaginable 00:16:04.470 --> 00:16:07.230 computing power to sustain. It demands a profound, 00:16:07.450 --> 00:16:09.950 unprecedented surrender of our personal biometric 00:16:09.950 --> 00:16:13.129 privacy. And it requires a completely new security 00:16:13.129 --> 00:16:15.070 paradigm for protecting the Internet. We are 00:16:15.070 --> 00:16:17.690 trading fundamental control for unprecedented 00:16:17.690 --> 00:16:21.129 convenience on an enormous scale. Thank you so 00:16:21.129 --> 00:16:23.350 much for joining this deep dive today. We always 00:16:23.350 --> 00:16:25.190 appreciate you spending your valuable time with 00:16:25.190 --> 00:16:27.889 us. I want to leave you with one final lingering 00:16:27.889 --> 00:16:31.009 thought. Two years ago, humans literally wrote 00:16:31.009 --> 00:16:34.730 our basic software manually. Now imagine a powerful 00:16:34.730 --> 00:16:36.870 AI loop running endlessly in the background. 00:16:37.549 --> 00:16:40.009 It subtly rewrites its own complex architecture, 00:16:40.289 --> 00:16:43.289 continuously and silently. Human developers just 00:16:43.289 --> 00:16:45.970 sit back and lazily review the final patch. Beat. 00:16:46.129 --> 00:16:48.470 At what point does the software officially stop 00:16:48.470 --> 00:16:50.370 belonging to us and start belonging entirely 00:16:50.370 --> 00:16:52.970 to the continuously evolving autonomous swarm?