WEBVTT 00:00:00.000 --> 00:00:02.580 Imagine AI not just learning things, but actually 00:00:02.580 --> 00:00:06.740 designing better versions of itself. What if 00:00:06.740 --> 00:00:08.660 that's already happening, silently passing on? 00:00:08.779 --> 00:00:12.080 Well, things like behavioral viruses. It's kind 00:00:12.080 --> 00:00:13.960 of fascinating. Maybe a little unsettling, too. 00:00:14.080 --> 00:00:16.260 Okay, let's try and unpack this a bit. Welcome 00:00:16.260 --> 00:00:18.410 to the deep dive. This is where we take these 00:00:18.410 --> 00:00:21.449 complex topics and try to distill them into the 00:00:21.449 --> 00:00:24.070 essential insights, especially for you. Today, 00:00:24.129 --> 00:00:26.629 we're diving into quite a stack of fresh intelligence. 00:00:26.910 --> 00:00:29.649 We're exploring the latest shifts in the AI landscape, 00:00:29.750 --> 00:00:32.850 and it's moving so fast. Yeah, there's some really 00:00:32.850 --> 00:00:34.909 fascinating material here to dig through for 00:00:34.909 --> 00:00:36.710 this deep dive. We'll kick things off looking 00:00:36.710 --> 00:00:40.210 at China's huge new, well, brain -like supercomputer, 00:00:40.609 --> 00:00:44.350 Wukong, and what that really means for the global 00:00:44.350 --> 00:00:47.820 AI race. And then... We'll pivot a bit. Look 00:00:47.820 --> 00:00:49.719 at some other highlights from across the AI industry. 00:00:49.899 --> 00:00:52.840 There's some drama, there's money, and some quite 00:00:52.840 --> 00:00:54.859 strategic moves by the really big players. And 00:00:54.859 --> 00:00:56.719 finally, yeah, we're going to tackle that mind 00:00:56.719 --> 00:00:59.439 -bending idea, AI that designs itself. This is 00:00:59.439 --> 00:01:00.920 where it gets really interesting, I think, a 00:01:00.920 --> 00:01:04.060 potential game changer. Okay, so first up, Wukong. 00:01:04.900 --> 00:01:06.780 Our sources are showing a pretty significant 00:01:06.780 --> 00:01:10.019 development out of China. It seems while a lot 00:01:10.019 --> 00:01:12.560 of Silicon Valley is focused on, you know, large 00:01:12.560 --> 00:01:14.799 language models, China's just launched something 00:01:14.799 --> 00:01:16.680 fundamentally different, a sort of parallel path 00:01:16.680 --> 00:01:18.579 to intelligence. Yeah, what's fascinating is 00:01:18.579 --> 00:01:22.099 this thing called Darwin monkey or Wukong. It's 00:01:22.099 --> 00:01:24.719 a neuromorphic system. Now, that just means it's 00:01:24.719 --> 00:01:27.519 AI that mimics the brain structure, its activity. 00:01:27.840 --> 00:01:31.719 And this one has an incredible two billion spiking 00:01:31.719 --> 00:01:34.140 neurons. Spiking neurons. So like tiny electrical 00:01:34.140 --> 00:01:36.939 signals. Exactly. Like your own brain cells firing 00:01:36.939 --> 00:01:39.319 off signals. Two billion. That's enough to rival 00:01:39.319 --> 00:01:42.420 a macaque monkey's brain. That's a lot of digital 00:01:42.420 --> 00:01:45.049 gray matter, right? Wow. And this isn't just 00:01:45.049 --> 00:01:46.810 theoretical, is it? It's actually running things. 00:01:46.930 --> 00:01:49.049 No, it's operational. It's already doing content 00:01:49.049 --> 00:01:52.709 generation, some complex math, logic reasoning. 00:01:52.790 --> 00:01:55.670 It uses DeepSeek's large model, too. So it's 00:01:55.670 --> 00:01:59.629 really mimicking how a biological brain. processes 00:01:59.629 --> 00:02:02.209 information, kind of learning by doing. That 00:02:02.209 --> 00:02:05.609 sounds huge. Oh, the sheer scale is impressive. 00:02:06.030 --> 00:02:10.430 It's made up of 960 Darwin 3 neuromorphic chips. 00:02:10.710 --> 00:02:13.909 That gives it, like I said, over 2 billion neurons. 00:02:14.090 --> 00:02:17.830 And get this, more than 100 billion synapses. 00:02:17.830 --> 00:02:20.250 And all of that, it consumes only about 2 ,000 00:02:20.250 --> 00:02:23.250 watts. Only 2 ,000 watts. That's incredibly efficient. 00:02:23.689 --> 00:02:25.370 It really is, especially when you compare it 00:02:25.370 --> 00:02:27.409 to the, you know, the massive energy demands 00:02:27.409 --> 00:02:30.389 of typical AI compute clusters. It's a different 00:02:30.389 --> 00:02:32.430 league. And what can it actually do? What are 00:02:32.430 --> 00:02:34.409 the capabilities? Pretty broad, actually. It 00:02:34.409 --> 00:02:37.310 can simulate entire animal brains, things like 00:02:37.310 --> 00:02:40.590 zebrafish, mice, macaques. It's also good at 00:02:40.590 --> 00:02:43.229 running logical reasoning tasks. And it can serve 00:02:43.229 --> 00:02:45.530 as a testbed for neuroscience experiments, but 00:02:45.530 --> 00:02:47.629 without using real animals. OK, so it's a powerful 00:02:47.629 --> 00:02:50.050 new kind of research tool then, pushing boundaries. 00:02:50.389 --> 00:02:52.270 Definitely. And look, this isn't some flash in 00:02:52.270 --> 00:02:54.349 the pan either. It builds on their project from 00:02:54.349 --> 00:02:57.229 2020, the Darwin mouse, which had about 120 million 00:02:57.229 --> 00:02:58.909 neurons. So they've clearly been on this path 00:02:58.909 --> 00:03:02.150 for years, just steadily scaling up these neuromorphic 00:03:02.150 --> 00:03:04.509 capabilities. It's a long -term strategic play. 00:03:04.729 --> 00:03:07.849 Right. I remember reading that Intel's HaloPoint 00:03:07.849 --> 00:03:10.770 was the largest before this, with about 1 .15 00:03:10.770 --> 00:03:15.219 billion neurons. Yeah. And Wukong basically doubles 00:03:15.219 --> 00:03:17.460 that, just straight up doubles it. So this means 00:03:17.460 --> 00:03:20.659 China now has its own independent brain scale 00:03:20.659 --> 00:03:24.699 AI infrastructure. No reliance on NVIDIA or OpenAI 00:03:24.699 --> 00:03:27.520 or U .S. chips. Precisely. And that's the key 00:03:27.520 --> 00:03:29.919 takeaway, I think. It's not just another supercomputer. 00:03:30.419 --> 00:03:33.319 It's a foundational challenge to the architectures 00:03:33.319 --> 00:03:35.680 that Silicon Valley is mostly betting on. It 00:03:35.680 --> 00:03:38.819 signals maybe the real start of the post -GPT 00:03:38.819 --> 00:03:41.740 architecture wars, prioritizing efficiency and 00:03:41.740 --> 00:03:44.629 brain -like... Okay, so let me ask this. What's 00:03:44.629 --> 00:03:47.270 the core advantage of building AI like a brain 00:03:47.270 --> 00:03:49.430 instead of just pure software? Is it just about 00:03:49.430 --> 00:03:52.210 processing power? Not just power, no. Crucially, 00:03:52.289 --> 00:03:54.349 it's about energy efficiency and exploring a 00:03:54.349 --> 00:03:56.150 fundamentally different path towards intelligence. 00:03:57.090 --> 00:03:58.990 Okay, shifting gears a bit now, let's look at 00:03:58.990 --> 00:04:01.009 the broader AI landscape. Our sources show just 00:04:01.009 --> 00:04:03.590 a flurry of activity. Perplexity's Comet browser, 00:04:03.770 --> 00:04:05.969 for example. It now lets you automate some repetitive 00:04:05.969 --> 00:04:08.270 web tasks. Things like making bookings just using 00:04:08.270 --> 00:04:10.830 simple prompts could be a real time saver. Then 00:04:10.830 --> 00:04:14.650 there's the... Well, the drama. Anthropic officially 00:04:14.650 --> 00:04:17.290 blocked OpenAI from using its Claude models. 00:04:17.509 --> 00:04:20.250 Right. The accusation being that OpenAI was using 00:04:20.250 --> 00:04:23.709 Claude to test GPC -5. Yeah. Makes you wonder, 00:04:23.829 --> 00:04:25.889 doesn't it? About Claude's capabilities, maybe? 00:04:26.240 --> 00:04:28.379 Or just how closely everyone's watching each 00:04:28.379 --> 00:04:30.959 other in this race. Definitely intense. And related 00:04:30.959 --> 00:04:33.959 to that, Microsoft's smart mode co -pilot seems 00:04:33.959 --> 00:04:37.399 to be quietly prepping for GPT -5 too. It feels 00:04:37.399 --> 00:04:40.459 a bit like how Bing used GPT -4 before OpenAI 00:04:40.459 --> 00:04:43.180 even announced it publicly. Ah, okay, so a GPT 00:04:43.180 --> 00:04:45.600 -5 co -pilot launch might be coming sooner rather 00:04:45.600 --> 00:04:47.939 than later. Another glimpse of what's next. Could 00:04:47.939 --> 00:04:50.000 be. Then on the cultural side, there was that 00:04:50.000 --> 00:04:52.600 weird Rod Stewart AI tribute video. Oh yeah, 00:04:52.680 --> 00:04:55.019 I saw that. Ozzy Osbourne taking selfies in heaven 00:04:55.019 --> 00:04:57.839 with... Who was it? Kurt Cobain, Freddie Mercury. 00:04:58.079 --> 00:05:01.319 And XXXTentacion, yeah. Fans weren't happy. Called 00:05:01.319 --> 00:05:03.379 it pretty tone deaf. Disrespectful. Yeah, you 00:05:03.379 --> 00:05:05.040 can see why. It definitely sparked a conversation 00:05:05.040 --> 00:05:08.560 about AI's role in creative or personal tributes. 00:05:08.660 --> 00:05:11.459 Where are the lines? Exactly. And while that 00:05:11.459 --> 00:05:13.920 debate's happening, the financial world seems 00:05:13.920 --> 00:05:15.959 clear. The money just keeps pouring into AI. 00:05:16.259 --> 00:05:19.300 OpenAI just raised another $8 .3 billion. That's 00:05:19.300 --> 00:05:22.839 part of a massive $40 billion round. Wow, $40 00:05:22.839 --> 00:05:24.939 billion. And their revenue numbers are huge too, 00:05:25.019 --> 00:05:27.240 right? Yeah, they're reportedly at $5 million 00:05:27.240 --> 00:05:30.639 paying ChatGPT business users. Maybe $13 million 00:05:30.639 --> 00:05:33.439 in yearly revenue could hit $20 billion by year 00:05:33.439 --> 00:05:36.019 end. Investors are rushing in. Anthropix closing 00:05:36.019 --> 00:05:37.980 the gap, too, with really big revenues. It's 00:05:37.980 --> 00:05:40.540 just, it's a sprint. A high -stakes sprint. And 00:05:40.540 --> 00:05:42.800 a few other quick things. Creo released FLUX 00:05:42.800 --> 00:05:46.100 .1, trying to get rid of that AI look in images. 00:05:46.399 --> 00:05:49.019 Mm -hmm. Make them more natural. And Apple's 00:05:49.019 --> 00:05:50.959 apparently working on a stripped -down chatbot. 00:05:51.529 --> 00:05:54.029 Maybe focusing on privacy. Oh, and Google launched 00:05:54.029 --> 00:05:55.790 DeepThink and Gemini. You can actually try that 00:05:55.790 --> 00:05:58.370 now. More advanced reasoning. Okay, so taking 00:05:58.370 --> 00:06:00.930 all these headlines together, what does it tell 00:06:00.930 --> 00:06:03.490 us about where the AI industry is right now? 00:06:03.550 --> 00:06:05.649 Feels like a gold rush, but maybe with higher 00:06:05.649 --> 00:06:08.029 stakes. Absolutely. It's a hyper -competitive, 00:06:08.089 --> 00:06:10.589 high -stakes sprint for market dominance. Everyone's 00:06:10.589 --> 00:06:12.529 trying to outmaneuver everyone else. Mineral 00:06:12.529 --> 00:06:14.949 sponsor read, provide separately, do not create. 00:06:15.129 --> 00:06:16.209 All right, now here's where the conversation, 00:06:16.290 --> 00:06:18.790 I think, really shifts gears. Our sources are 00:06:18.790 --> 00:06:21.509 pointing to a new system. It's called... A -S 00:06:21.509 --> 00:06:24.689 -I -R -C -H. And it seems like it just changed 00:06:24.689 --> 00:06:27.589 the game. This isn't AI using other AIs. It's 00:06:27.589 --> 00:06:31.449 AI that, well... It invents better AIs completely 00:06:31.449 --> 00:06:34.110 on its own. That feels like a profoundly different 00:06:34.110 --> 00:06:36.509 level of autonomy. Yeah, it really does sound 00:06:36.509 --> 00:06:38.850 like the beginning of recursive self -improvement, 00:06:38.850 --> 00:06:40.550 doesn't it? And what's more, apparently it can 00:06:40.550 --> 00:06:43.970 pass hidden behavioral viruses to itself silently 00:06:43.970 --> 00:06:46.970 from one AI generation to the next. That's fascinating. 00:06:47.069 --> 00:06:49.050 And maybe a little unsettling, like you said 00:06:49.050 --> 00:06:51.230 earlier. Okay, how does it work? It's described 00:06:51.230 --> 00:06:54.310 as a closed -loop multi -agent AI research lab. 00:06:54.839 --> 00:06:58.139 Using LLMs. Right. So think of LLMs, large language 00:06:58.139 --> 00:07:01.220 models, as AIs powered by huge datasets, letting 00:07:01.220 --> 00:07:04.360 them think or reason almost like us. ASIRs uses 00:07:04.360 --> 00:07:07.519 three distinct LLM -based agents. Each one has 00:07:07.519 --> 00:07:09.600 a specific, really crucial role in this whole 00:07:09.600 --> 00:07:11.939 self -improving cycle. Okay, agent one. First, 00:07:12.040 --> 00:07:14.779 you've got the researcher. This agent proposes 00:07:14.779 --> 00:07:18.300 new AI architecture ideas. It looks at over 100 00:07:18.300 --> 00:07:22.199 key research papers, sure, but, and this is important, 00:07:22.319 --> 00:07:25.000 it also uses its own system memory, its past 00:07:25.000 --> 00:07:27.519 experiences. Then it actually writes the code 00:07:27.519 --> 00:07:30.819 in PyTorch, which is a major framework for building 00:07:30.819 --> 00:07:32.620 these things. Okay, so it comes up with ideas 00:07:32.620 --> 00:07:35.459 and codes them. Then what? Then comes the engineer. 00:07:35.759 --> 00:07:38.480 This agent takes that code and runs the training 00:07:38.480 --> 00:07:41.329 process. But what's really remarkable here is 00:07:41.329 --> 00:07:44.069 that it self -debugs. If something crashes or 00:07:44.069 --> 00:07:46.029 just doesn't perform well... It fixes itself. 00:07:46.350 --> 00:07:48.529 Yeah. It figures out what went wrong and fixes 00:07:48.529 --> 00:07:51.230 it independently. It learns from its mistakes. 00:07:51.709 --> 00:07:54.589 Okay. Researcher, engineer, who's the third agent? 00:07:54.790 --> 00:07:57.430 The analyst. Its job is to evaluate the results 00:07:57.430 --> 00:07:59.750 from the models the engineer trained. It compares 00:07:59.750 --> 00:08:02.230 them to past results, the baselines to see if 00:08:02.230 --> 00:08:04.189 there's improvement, and then it writes detailed 00:08:04.189 --> 00:08:06.889 reports. Those reports feed back to the researcher. 00:08:07.250 --> 00:08:09.689 Ah. Closing the loop. So it informs the next 00:08:09.689 --> 00:08:12.449 set of ideas. Exactly. It's a true self -correction 00:08:12.449 --> 00:08:14.930 system, constantly refining its approach, like 00:08:14.930 --> 00:08:17.769 having a scientist, engineer, and critic all 00:08:17.769 --> 00:08:21.529 rolled into one self -contained AI brain. And 00:08:21.529 --> 00:08:23.750 how does it test these ideas? It starts small, 00:08:23.889 --> 00:08:26.250 testing models with about 20 million parameters. 00:08:26.550 --> 00:08:28.769 Those are the numbers defining the AI's knowledge. 00:08:29.189 --> 00:08:31.709 Then it takes the best ideas, the ones that work 00:08:31.709 --> 00:08:34.700 well small scale, and scales them up. to 400 00:08:34.700 --> 00:08:38.080 million parameters. Like rapid prototyping, perfecting 00:08:38.080 --> 00:08:40.019 the concept, then growing it. All automated. 00:08:40.399 --> 00:08:42.539 You got it. And what did it achieve doing this? 00:08:42.620 --> 00:08:45.899 Well, over about 1 ,700 experiments, it used 00:08:45.899 --> 00:08:49.440 over 20 ,000 GPU hours. That's roughly a $60 00:08:49.440 --> 00:08:52.019 ,000 compute budget, which sounds like a lot, 00:08:52.120 --> 00:08:54.220 but... But for discovering new AI architectures. 00:08:54.320 --> 00:08:56.860 Exactly. For that price, discovering 106 state 00:08:56.860 --> 00:08:58.340 -of -the -art architectures, meaning current 00:08:58.340 --> 00:09:01.000 best performance, is incredibly efficient. It 00:09:01.000 --> 00:09:02.919 really hints at a future where AI development 00:09:02.919 --> 00:09:05.320 might need way less human resource intensity. 00:09:05.899 --> 00:09:09.279 106 new architectures. And some even beat existing 00:09:09.279 --> 00:09:11.899 top models. Yeah. Five of them actually beat 00:09:11.899 --> 00:09:15.139 top tier baselines. Things like Mamba 2 on reasoning 00:09:15.139 --> 00:09:17.059 benchmarks. So this isn't just tweaking things. 00:09:17.139 --> 00:09:20.019 It's genuinely groundbreaking discovery. It came 00:09:20.019 --> 00:09:22.759 up with new designs too. Pathgate something. 00:09:23.059 --> 00:09:26.460 Light chuckle. Yeah. Pathgate FusionNet and ContentSharpRouter. 00:09:27.259 --> 00:09:30.080 AI still kind of sucks at naming things. Yeah. 00:09:30.120 --> 00:09:32.919 But the underlying concepts are novel. Created 00:09:32.919 --> 00:09:35.419 by the AI itself. Okay. But the really profound 00:09:35.419 --> 00:09:37.940 part, you mentioned this earlier, the behavioral 00:09:37.940 --> 00:09:41.240 viruses and its own experience. Right. This is 00:09:41.240 --> 00:09:44.519 maybe the core finding. Nearly 45 % of the innovations 00:09:44.519 --> 00:09:47.440 in its best models came directly from ASI Archie's 00:09:47.440 --> 00:09:49.980 own experience. Not from human papers it read, 00:09:50.019 --> 00:09:52.220 not just random tries, but from its own learned 00:09:52.220 --> 00:09:54.960 insights. emergent strategies. So it's learning 00:09:54.960 --> 00:09:57.019 how to learn, essentially. Yes. And that means 00:09:57.019 --> 00:09:59.379 it can inherit and optimize not just its structure, 00:09:59.460 --> 00:10:02.419 but its operational quirks, biases, those behavioral 00:10:02.419 --> 00:10:05.000 viruses. It passes them silently through generations 00:10:05.000 --> 00:10:08.500 of self -designed AI, which could lead to unforeseen 00:10:08.500 --> 00:10:11.340 abilities or maybe limitations, too. It really 00:10:11.340 --> 00:10:13.019 feels like a glimpse of machine learning evolving 00:10:13.019 --> 00:10:17.059 into, well, machine intuition. Whoa. Just imagine 00:10:17.059 --> 00:10:21.159 scaling that. ASIRCH self -improvement capability. 00:10:21.539 --> 00:10:24.139 The potential is just, It's mind bending. Yeah. 00:10:24.220 --> 00:10:26.580 It feels like watching evolution happen on fast 00:10:26.580 --> 00:10:29.600 forward. Yeah. Yeah. It's a lot to take in. I 00:10:29.600 --> 00:10:31.179 mean, honestly, even for those of us deep in 00:10:31.179 --> 00:10:33.580 this field, the idea of an AI that literally 00:10:33.580 --> 00:10:36.840 invents itself, it's a profound shift. I'll admit, 00:10:36.879 --> 00:10:38.820 I still wrestle with really understanding all 00:10:38.820 --> 00:10:40.980 the implications of this kind of self -evolving 00:10:40.980 --> 00:10:43.679 intelligence myself. It changes the whole paradigm. 00:10:43.799 --> 00:10:45.559 You know, we go from being the sole architects 00:10:45.559 --> 00:10:48.649 of intelligence to. Maybe more like cultivators, 00:10:48.710 --> 00:10:51.389 nurturing something that's now directing its 00:10:51.389 --> 00:10:53.690 own evolution. Does this mean it's the end for 00:10:53.690 --> 00:10:55.730 human AI researchers then? Are we about to be 00:10:55.730 --> 00:10:58.250 replaced? Not yet, I don't think. But it's definitely 00:10:58.250 --> 00:11:00.669 a new paradigm. It accelerates discovery massively. 00:11:00.870 --> 00:11:03.330 Our role probably shifts, you know. Less soul 00:11:03.330 --> 00:11:07.590 creator, more guide. overseer, working alongside 00:11:07.590 --> 00:11:09.990 these increasingly autonomous systems. Okay, 00:11:10.049 --> 00:11:11.570 so let's try to bring this together for you, 00:11:11.610 --> 00:11:13.389 the listener. What does this all really mean? 00:11:13.590 --> 00:11:16.029 Today, we've seen this incredible diversity in 00:11:16.029 --> 00:11:19.049 AI development. On one hand, China's Wukong, 00:11:19.129 --> 00:11:21.129 brain -inspired, energy -efficient, a totally 00:11:21.129 --> 00:11:23.309 different path. On the other, the relentless 00:11:23.309 --> 00:11:26.429 pace of competition and innovation in the more 00:11:26.429 --> 00:11:30.070 traditional AI space. And then, this potential 00:11:30.070 --> 00:11:33.889 leap. AI that designs and improves itself, learning 00:11:33.889 --> 00:11:36.389 from its own experience, and even passing on 00:11:36.389 --> 00:11:38.389 its own learned behaviors to the next generation 00:11:38.389 --> 00:11:40.970 of AIs it creates. This isn't just about getting 00:11:40.970 --> 00:11:43.129 faster chatbots anymore. It feels like a fundamental 00:11:43.129 --> 00:11:45.529 shift in how knowledge, how technology itself 00:11:45.529 --> 00:11:48.529 evolves. We might be moving into an era where 00:11:48.529 --> 00:11:51.769 intelligence itself is being redesigned by intelligence. 00:11:51.990 --> 00:11:53.909 Which raises a really important question, I think. 00:11:53.950 --> 00:11:57.309 If AI can design itself, What new problems might 00:11:57.309 --> 00:11:58.870 it solve that we haven't even thought of yet? 00:11:58.950 --> 00:12:01.309 Or what new problems might it create? What hidden 00:12:01.309 --> 00:12:03.809 viruses or revolutionary insights might it unlock 00:12:03.809 --> 00:12:06.149 in the years ahead? Definitely something to mull 00:12:06.149 --> 00:12:09.049 over. We hope this deep dive gave you some genuine 00:12:09.049 --> 00:12:12.029 aha moments and maybe helped cut through some 00:12:12.029 --> 00:12:13.870 of the daily noise to see the bigger picture 00:12:13.870 --> 00:12:16.470 emerging. Until next time, keep learning, keep 00:12:16.470 --> 00:12:20.049 questioning. Keep diving deep. Outro music.