WEBVTT 00:00:00.000 --> 00:00:02.980 We are witnessing the death of a prophecy. Beat. 00:00:03.120 --> 00:00:06.459 The idea of one AI tool to rule them all is dying. 00:00:06.960 --> 00:00:09.599 Meta wants to delete the computer operating system 00:00:09.599 --> 00:00:13.019 entirely. Welcome to the deep dive. Thanks. It 00:00:13.019 --> 00:00:15.699 is genuinely great to be here today. Today's 00:00:15.699 --> 00:00:18.399 stack of sources reveals a really massive shift. 00:00:18.559 --> 00:00:21.239 We are moving far away from single AI tools. 00:00:21.460 --> 00:00:23.940 We're heading into complex, customized tool chains. 00:00:24.059 --> 00:00:26.949 We're scaling to... billions of autonomous agents, 00:00:27.190 --> 00:00:30.230 and we are completely rethinking how AI interacts 00:00:30.230 --> 00:00:32.210 with our computers. What's fascinating here is 00:00:32.210 --> 00:00:34.270 how fast this is moving. We are jumping from 00:00:34.270 --> 00:00:36.869 the current coding trenches today, right straight 00:00:36.869 --> 00:00:38.530 into the future of neural operating systems. 00:00:38.670 --> 00:00:40.289 Yeah, let's start down in those coding trenches. 00:00:40.750 --> 00:00:42.969 Developers are feeling this massive shift right 00:00:42.969 --> 00:00:46.070 now. The dream of a monolithic coding AI is practically 00:00:46.070 --> 00:00:48.369 dead. It's totally fragmenting. I mean, the AI 00:00:48.369 --> 00:00:51.189 coding market is now a three -layer stack. Devs 00:00:51.189 --> 00:00:52.990 are actively composing their own tool chains. 00:00:53.189 --> 00:00:55.509 They're basically combining cursor, quad, and 00:00:55.509 --> 00:00:57.929 codex together. Right. It's like stacking Lego 00:00:57.929 --> 00:01:00.250 blocks of data. You don't want a pre -built house 00:01:00.250 --> 00:01:03.780 anymore. You want specialized blocks to build 00:01:03.780 --> 00:01:07.560 your own custom toolchains. So you want Claude 00:01:07.560 --> 00:01:10.200 as your head chef. He handles the complex architectural 00:01:10.200 --> 00:01:14.040 recipes. And you want Codex just chopping vegetables 00:01:14.040 --> 00:01:16.859 in the background. Yeah, and it fundamentally 00:01:16.859 --> 00:01:19.840 changes the whole workflow. Just look at Cursor 00:01:19.840 --> 00:01:22.819 3, which they're calling Glass. It's actively 00:01:22.819 --> 00:01:25.060 distancing itself from its VS Code roots. It 00:01:25.060 --> 00:01:27.920 focuses on its agents window and agent tabs. 00:01:28.510 --> 00:01:30.489 They call that the manager surface, right? Exactly. 00:01:30.629 --> 00:01:32.170 They officially call it the manager surface. 00:01:32.430 --> 00:01:34.590 A manager surface implies active delegation. 00:01:35.010 --> 00:01:36.890 You know, you're no longer just typing code yourself. 00:01:37.069 --> 00:01:39.329 You're managing digital workers who write it 00:01:39.329 --> 00:01:41.090 for you. Precisely. You're orchestrating the 00:01:41.090 --> 00:01:43.790 build process. And then you have quad code entering 00:01:43.790 --> 00:01:47.349 the mix. It holds 46 % of the most loved market 00:01:47.349 --> 00:01:50.390 share. It uses an MCP -based plug -in system 00:01:50.390 --> 00:01:52.730 to run things. Let's define that really quickly 00:01:52.730 --> 00:01:55.329 for you listening. An MCP -based plugin is a 00:01:55.329 --> 00:01:57.950 universal plug letting AI talk directly to local 00:01:57.950 --> 00:02:00.969 files. Spot on. That system acts as the essential 00:02:00.969 --> 00:02:04.329 glue. It's the primary execution layer for complex 00:02:04.329 --> 00:02:06.590 architectural refactoring. It basically does 00:02:06.590 --> 00:02:09.020 the heavy intellectual lifting. But looking at 00:02:09.020 --> 00:02:11.719 OpenAI's placement in this stack, it feels like 00:02:11.719 --> 00:02:13.759 a downgrade. Are they really just the review 00:02:13.759 --> 00:02:16.219 layer now? Yeah, they repositioned Codex to fit 00:02:16.219 --> 00:02:19.139 exactly that niche. They pragmatically embedded 00:02:19.139 --> 00:02:21.840 it directly into Cloud Code. Using a plugin, 00:02:22.000 --> 00:02:23.939 right? They did, yeah. It's called the Codex 00:02:23.939 --> 00:02:27.000 Plugin CC. They are actively fighting what developers 00:02:27.000 --> 00:02:30.909 call agent fatigue. You use Claude for the big 00:02:30.909 --> 00:02:32.990 picture architectural stuff, but then you swap 00:02:32.990 --> 00:02:35.750 models for the smaller tasks. You hand low complexity 00:02:35.750 --> 00:02:40.090 tasks to Codex or Kimi K2 .5. Just to clarify, 00:02:40.310 --> 00:02:42.590 Kimi K2 .5 for everyone, it's a lightweight model 00:02:42.590 --> 00:02:45.009 for simple background tasks. You do that to manage 00:02:45.009 --> 00:02:47.250 your token burn. Right, because token burn is 00:02:47.250 --> 00:02:49.930 a huge issue. Wasting expensive compute on simple 00:02:49.930 --> 00:02:52.210 tasks will bankrupt you. It's a highly pragmatic 00:02:52.210 --> 00:02:54.509 token economy. I still wrestle with prompt drift 00:02:54.509 --> 00:02:57.460 myself. Oh, we all do. It's exhausting. Yeah, 00:02:57.560 --> 00:02:59.919 the friction of managing these different tools 00:02:59.919 --> 00:03:02.539 is incredibly real. Keeping context perfectly 00:03:02.539 --> 00:03:05.680 aligned across platforms is tough. So let me 00:03:05.680 --> 00:03:09.360 ask you this. Is OpenAI's pragmatic plug -in 00:03:09.360 --> 00:03:12.639 strategy a sign of surrender? Or is it pure genius? 00:03:13.000 --> 00:03:15.599 I think it's pure genius. They willingly surrendered 00:03:15.599 --> 00:03:18.039 the terminal to win the workflow. They embedded 00:03:18.039 --> 00:03:20.000 themselves exactly where the work actually happened. 00:03:20.099 --> 00:03:21.939 So they abandoned the interface to capture the 00:03:21.939 --> 00:03:23.860 actual process. That's the perfect way to phrase 00:03:23.860 --> 00:03:26.340 it. But to run these multilayered tool chains, 00:03:26.500 --> 00:03:29.439 you need serious infrastructure. You need massive 00:03:29.439 --> 00:03:32.000 computing power to handle that load. That brings 00:03:32.000 --> 00:03:35.000 scale, but also very serious risks. Risks are 00:03:35.000 --> 00:03:37.199 getting incredibly massive. And the infrastructure 00:03:37.199 --> 00:03:39.900 is scaling up right now. Cloudflare just announced 00:03:39.900 --> 00:03:42.479 Agents Week. They're rolling out edge -based 00:03:42.479 --> 00:03:44.860 infrastructure designed to run billions of agents 00:03:44.860 --> 00:03:47.819 simultaneously. Billions of active digital workers. 00:03:48.020 --> 00:03:51.419 Imagine scaling to a billion queries. The physical 00:03:51.419 --> 00:03:53.900 compute power required is staggering. to think 00:03:53.900 --> 00:03:56.439 about. And consumer adoption is totally matching 00:03:56.439 --> 00:03:58.879 that scale. CloudCore just moved to full release 00:03:58.879 --> 00:04:02.879 for paid plans. Pro, Max, Team, Enterprise tiers. 00:04:03.319 --> 00:04:06.620 The demand is off the charts. At the Humanex 00:04:06.620 --> 00:04:09.719 AI conference last week, one chatbot came up 00:04:09.719 --> 00:04:13.580 in every panel. Industry leaders dubbed CloudCode 00:04:13.580 --> 00:04:16.199 the absolute must -have tool. But this environment 00:04:16.199 --> 00:04:19.120 is getting genuinely risky. The sheer scale is 00:04:19.120 --> 00:04:21.759 unprecedented. Sam Altman recently broke his 00:04:21.759 --> 00:04:23.879 silence in a New Yorker profile. Yeah, right 00:04:23.879 --> 00:04:25.920 after that frightening attack on his home. Right. 00:04:26.120 --> 00:04:28.500 He admitted to some past leadership mistakes, 00:04:28.800 --> 00:04:32.240 but he also issued a really stark warning. He 00:04:32.240 --> 00:04:34.959 warned that the AGI race is pushing into highly 00:04:34.959 --> 00:04:37.639 risky behaviors. The chaos on the ground is real. 00:04:38.379 --> 00:04:40.899 TechCrunch is even updating their official glossary 00:04:40.899 --> 00:04:43.240 of terms just to help normal people navigate 00:04:43.240 --> 00:04:45.459 the madness. Which you desperately need in this 00:04:45.459 --> 00:04:48.100 space. Take the term hallucination, for example. 00:04:48.459 --> 00:04:51.240 Hallucination. When an AI confidently makes up 00:04:51.240 --> 00:04:53.759 fake information. Exactly. We need simple definitions 00:04:53.759 --> 00:04:56.300 for these complex problems. The vocabulary has 00:04:56.300 --> 00:04:58.399 to keep up. But let me push back here for a second. 00:04:58.480 --> 00:05:00.620 We are scaling up to billions of fully autonomous 00:05:00.620 --> 00:05:03.699 agents. Even Altman is explicitly warning about 00:05:03.699 --> 00:05:06.660 risky behavior. Are we basically building the 00:05:06.660 --> 00:05:08.899 massive engine without installing any brakes? 00:05:09.199 --> 00:05:12.360 It's a super valid concern right now. The infrastructure 00:05:12.360 --> 00:05:15.319 expansion is totally outpacing the safety guardrails. 00:05:16.110 --> 00:05:18.170 We're deploying agents before we understand their 00:05:18.170 --> 00:05:20.750 emergent behaviors. So how do we actually resolve 00:05:20.750 --> 00:05:24.370 this tension between massive digital scale and 00:05:24.370 --> 00:05:27.370 fundamental human safety? Well, we have to build 00:05:27.370 --> 00:05:29.810 robust guardrails at the infrastructure layer 00:05:29.810 --> 00:05:32.750 itself. We just can't rely on the models to police 00:05:32.750 --> 00:05:35.189 themselves anymore. We have to hardwire those 00:05:35.189 --> 00:05:37.170 safety brakes directly into the infrastructure. 00:05:37.430 --> 00:05:39.930 That is the only viable path forward. Okay, let's 00:05:39.930 --> 00:05:43.350 untack this. Beat, we spent the entire last decade 00:05:43.350 --> 00:05:45.769 putting glowing screens on everything. We moved 00:05:45.769 --> 00:05:48.529 all our private files to the distant cloud. And 00:05:48.529 --> 00:05:51.490 now Apple wants to remove screens entirely. LM 00:05:51.490 --> 00:05:53.470 Studio wants to take everything off the cloud. 00:05:53.610 --> 00:05:55.870 It's a massive philosophical pivot. It really 00:05:55.870 --> 00:05:57.750 is. It's a direct reaction to the friction of 00:05:57.750 --> 00:06:00.550 modern computing. People are tired of the costs 00:06:00.550 --> 00:06:03.470 and risks of cloud AI. We're seeing a massive 00:06:03.470 --> 00:06:06.250 pivot toward offline, highly localized tools. 00:06:06.509 --> 00:06:09.230 People want specialized personal AI on their 00:06:09.230 --> 00:06:11.569 own hardware. And LM Studio just made a big strategic 00:06:11.569 --> 00:06:13.949 move there. They certainly did. They acquired 00:06:13.949 --> 00:06:18.019 a fascinating company called Locally AI. Adrienne 00:06:18.019 --> 00:06:21.100 Grondin is joining to lead native AI experiences. 00:06:21.579 --> 00:06:24.480 They're bringing open source models natively 00:06:24.480 --> 00:06:27.860 to your personal devices. iPhone, iPad, Mac integration. 00:06:28.139 --> 00:06:30.819 With no cloud server connection required at all. 00:06:30.899 --> 00:06:33.480 Zero cloud connectivity. No sign -up friction. 00:06:33.680 --> 00:06:36.519 No monthly fee. It's just localized processing 00:06:36.519 --> 00:06:39.170 utilizing your own silicon. Apple is moving in 00:06:39.170 --> 00:06:41.889 a remarkably similar direction too. They're testing 00:06:41.889 --> 00:06:45.529 four distinct smart glasses styles for 2027. 00:06:45.970 --> 00:06:49.149 Crucially, these glasses feature absolutely no 00:06:49.149 --> 00:06:52.750 displays. None. Zero visual displays. It's a 00:06:52.750 --> 00:06:55.269 huge departure from their previous AR headsets. 00:06:55.350 --> 00:06:57.709 It's just a camera, spatial music, phone calls, 00:06:57.709 --> 00:06:59.810 and Siri. It's a much simpler approach to ambient 00:06:59.810 --> 00:07:02.279 computing. The tools are becoming hyper -specialized 00:07:02.279 --> 00:07:04.339 and local. Let's look at a few examples from 00:07:04.339 --> 00:07:06.399 our sources. There's a fascinating new tool called 00:07:06.399 --> 00:07:08.639 Ray. Oh, Ray is incredibly interesting. It acts 00:07:08.639 --> 00:07:11.180 as a terminal -based CFO. It reads your real 00:07:11.180 --> 00:07:13.139 computer transactions locally. It tells you what 00:07:13.139 --> 00:07:15.459 to do, helps plan budgets, and it runs entirely 00:07:15.459 --> 00:07:17.980 on your own private machine. Then there's R0Y, 00:07:18.259 --> 00:07:20.920 spelled with a zero, a natural language financial 00:07:20.920 --> 00:07:24.339 studio. It builds full investing dashboards in 00:07:24.339 --> 00:07:27.060 seconds, tailored to you. And Google's Gemini 00:07:27.060 --> 00:07:29.360 just introduced interactive visual simulations. 00:07:29.600 --> 00:07:32.319 You don't just read text. You ask Gemini to show 00:07:32.319 --> 00:07:34.759 me or help me visualize. You can actually play 00:07:34.759 --> 00:07:37.019 with abstract concepts locally. We also have 00:07:37.019 --> 00:07:40.259 the Eleven Labs music marketplace. Creators generate 00:07:40.259 --> 00:07:42.839 tracks and publish them seamlessly. They earn 00:07:42.839 --> 00:07:45.560 real money on downloads. Eleven Labs has already 00:07:45.560 --> 00:07:48.259 paid out $11 million to voice creators. It's 00:07:48.259 --> 00:07:51.420 birthing a totally new creator economy. The localized 00:07:51.420 --> 00:07:54.379 empowerment gives individuals studio -level production 00:07:54.379 --> 00:07:57.300 capabilities. Two sec silence. So looking at 00:07:57.300 --> 00:08:00.300 all this screenless, highly specific tech, is 00:08:00.300 --> 00:08:03.699 the true future of AI actually completely invisible 00:08:03.699 --> 00:08:06.240 and entirely offline? History shows the best 00:08:06.240 --> 00:08:09.060 technology always fades away. It disappears completely 00:08:09.060 --> 00:08:11.259 into the ambient background of our daily lives. 00:08:11.420 --> 00:08:13.459 So the best technology simply becomes an invisible 00:08:13.459 --> 00:08:15.339 part of our environment. It just becomes the 00:08:15.339 --> 00:08:18.560 silent air we breathe. We talked about powerful 00:08:18.560 --> 00:08:21.379 local tools. We discussed the shift toward invisible, 00:08:21.579 --> 00:08:24.230 display -free hardware. Here's where it gets 00:08:24.230 --> 00:08:28.910 really interesting. Beat, Meta is trying to leapfrog 00:08:28.910 --> 00:08:31.889 this entire localized paradigm. They want the 00:08:31.889 --> 00:08:34.490 AI to literally become the operating system. 00:08:34.750 --> 00:08:36.970 Yeah, they're trying to fundamentally delete 00:08:36.970 --> 00:08:40.350 the traditional OS entirely. Meta is moving away 00:08:40.350 --> 00:08:43.190 from models that simply use computers. They're 00:08:43.190 --> 00:08:45.549 developing models that actually act as the computers 00:08:45.549 --> 00:08:47.710 themselves. It's a monumental shift in architecture. 00:08:48.029 --> 00:08:51.049 Their new neural computer prototype executes 00:08:51.049 --> 00:08:53.470 desktop tasks directly via learned behavior. 00:08:53.909 --> 00:08:57.269 There are no APIs involved, no complex orchestration 00:08:57.269 --> 00:08:59.509 layers to translate commands. This is radically 00:08:59.509 --> 00:09:01.610 different from traditional AI tool use, isn't 00:09:01.610 --> 00:09:03.750 it? Completely different. When ChatGPT calls 00:09:03.750 --> 00:09:05.970 a Python interpreter, that's basic tool use. 00:09:06.149 --> 00:09:08.970 It uses a programmatic bridge. Meta's prototype 00:09:08.970 --> 00:09:12.570 is a unified, end -to -end, learned system. It's 00:09:12.570 --> 00:09:14.350 like teaching someone to drive by showing them 00:09:14.350 --> 00:09:16.370 thousands of photos of a dashboard. You never 00:09:16.370 --> 00:09:18.549 actually explain what a steering wheel mechanically 00:09:18.549 --> 00:09:20.610 does. You just show the visual results of turning 00:09:20.610 --> 00:09:23.279 it. That's a perfect visual metaphor. It was 00:09:23.279 --> 00:09:25.860 trained on thousands of hours of raw screen recordings. 00:09:26.200 --> 00:09:29.379 It passively watched human cursor movements and 00:09:29.379 --> 00:09:32.559 watched complex terminal sessions unfold pixel 00:09:32.559 --> 00:09:34.779 by pixel. So the logic isn't programmed with 00:09:34.779 --> 00:09:37.940 standard code. It's encoded directly into the 00:09:37.940 --> 00:09:40.139 model's mathematical weights. Yeah. And that 00:09:40.139 --> 00:09:42.340 changes how it actually operates. It predicts 00:09:42.340 --> 00:09:45.759 the next screen state exactly like GPT predicts 00:09:45.759 --> 00:09:48.639 the next word. It treats human UI interaction 00:09:48.639 --> 00:09:52.470 as raw visual. So if it needs to move a file, 00:09:52.570 --> 00:09:54.990 what does it actually do? It doesn't call a computer 00:09:54.990 --> 00:09:58.399 command or an API. It literally thinks. the specific 00:09:58.399 --> 00:10:01.299 cursor movements required it generates the exact 00:10:01.299 --> 00:10:04.179 keyboard inputs needed to drag and drop that 00:10:04.179 --> 00:10:06.279 file it essentially hallucinates the physical 00:10:06.279 --> 00:10:08.779 mouse moving across the screen mechanically yeah 00:10:08.779 --> 00:10:11.700 it predicts the pixel shifting but uh there are 00:10:11.700 --> 00:10:13.399 some glaring limitations right now it's just 00:10:13.399 --> 00:10:15.740 an early prototype not a finished product what 00:10:15.740 --> 00:10:18.080 happens when it tries a complex multi -step task 00:10:18.080 --> 00:10:20.779 it really struggles with sustained context it 00:10:20.779 --> 00:10:23.039 might remember to open a system folder but three 00:10:23.039 --> 00:10:25.200 steps later it completely forgets why it was 00:10:25.200 --> 00:10:28.250 there it loses the thread it lacks basic digital 00:10:28.250 --> 00:10:30.490 object permanence. That's exactly the problem. 00:10:31.029 --> 00:10:33.909 It hasn't learned the concept of a button. It 00:10:33.909 --> 00:10:36.309 has only learned the visual pattern of a button. 00:10:36.450 --> 00:10:38.830 It recognizes the pattern of an X icon in the 00:10:38.830 --> 00:10:41.429 corner. Right. It sees the X and knows to click 00:10:41.429 --> 00:10:44.330 it. But until the model learns that an X icon 00:10:44.330 --> 00:10:47.669 always means close, it remains limited. It has 00:10:47.669 --> 00:10:50.289 to understand that concept regardless of the 00:10:50.289 --> 00:10:53.559 application. Until then it's basically a highly 00:10:53.559 --> 00:10:56.559 sophisticated macro recorder. Kate, let's dig 00:10:56.559 --> 00:10:59.500 into that limitation. Is the barrier between 00:10:59.500 --> 00:11:01.860 recognizing visual patterns and understanding 00:11:01.860 --> 00:11:04.639 underlying concepts the hardest problem in machine 00:11:04.639 --> 00:11:07.860 learning today? I firmly believe it is. Bridging 00:11:07.860 --> 00:11:10.460 that massive cognitive gap is basically the definition 00:11:10.460 --> 00:11:13.179 of true artificial general intelligence. It's 00:11:13.179 --> 00:11:14.500 the difference between mimicking intelligence 00:11:14.500 --> 00:11:17.779 and possessing actual comprehension. So scaling 00:11:17.779 --> 00:11:20.320 up visual patterns might not organically lead 00:11:20.320 --> 00:11:22.820 to true conceptual understanding. That remains 00:11:22.820 --> 00:11:25.340 the absolute biggest unanswered question in AI 00:11:25.340 --> 00:11:27.659 research. Scaling visual patterns just doesn't 00:11:27.659 --> 00:11:29.960 guarantee true cognitive reasoning. You're finding 00:11:29.960 --> 00:11:31.879 that out the hard way, yeah. Let's zoom out for 00:11:31.879 --> 00:11:33.899 a second. We've covered a massive amount of technical 00:11:33.899 --> 00:11:35.860 ground today. If we connect this to the bigger 00:11:35.860 --> 00:11:38.610 picture. We're witnessing a total restructuring 00:11:38.610 --> 00:11:41.549 of how humanity computes. Devs are hacking together 00:11:41.549 --> 00:11:44.250 specialized tool chains today. We're scaling 00:11:44.250 --> 00:11:46.350 infrastructure to support billions of agents 00:11:46.350 --> 00:11:49.129 tomorrow. We're bringing open source models locally 00:11:49.129 --> 00:11:52.590 to our pockets. And Meta is aggressively trying 00:11:52.590 --> 00:11:55.950 to bypass all of it by turning the AI... into 00:11:55.950 --> 00:11:58.590 the computer itself it is a massive unprecedented 00:11:58.590 --> 00:12:01.149 transition thank you for joining us on this deep 00:12:01.149 --> 00:12:03.350 dive take a moment today to look at your own 00:12:03.350 --> 00:12:05.710 digital workflow try to find where you're getting 00:12:05.710 --> 00:12:07.629 your own agent fatigue and see where you can 00:12:07.629 --> 00:12:10.090 specialize your daily tools the landscape is 00:12:10.090 --> 00:12:12.269 shifting incredibly fast beneath our feet it 00:12:12.269 --> 00:12:14.490 really is and it leaves me with one final fascinating 00:12:14.490 --> 00:12:17.990 question to consider beat if meta's neural computer 00:12:17.990 --> 00:12:20.149 eventually succeeds in operating our graphic 00:12:20.149 --> 00:12:22.730 interfaces will the software of the future be 00:12:22.730 --> 00:12:26.090 designed to look good to human eyes Or will it 00:12:26.090 --> 00:12:28.669 be optimized entirely for AI vision models to 00:12:28.669 --> 00:12:30.809 easily read? Out to your own music.