WEBVTT 00:00:00.000 --> 00:00:04.400 We are witnessing a powerful and kind of dizzying 00:00:04.400 --> 00:00:08.500 thing in the AI landscape right now. On one hand, 00:00:08.599 --> 00:00:11.259 you've got these tools getting dramatically cheaper, 00:00:11.439 --> 00:00:14.519 faster, putting efficiency out there for everyone. 00:00:14.640 --> 00:00:17.640 Yeah, exactly. But then at the exact same time, 00:00:17.760 --> 00:00:20.620 you're seeing these deep, really fundamental 00:00:20.620 --> 00:00:23.019 breakthroughs happening in labs. Right. Like 00:00:23.019 --> 00:00:26.780 think about an AI predicting a totally new cancer 00:00:26.780 --> 00:00:29.500 therapy path. And it gets confirmed right away 00:00:29.500 --> 00:00:31.780 in experiments. It's wild. It's both the engine 00:00:31.780 --> 00:00:35.880 for global efficiency and the key to some really 00:00:35.880 --> 00:00:39.079 profound scientific discovery. All at once. Welcome 00:00:39.079 --> 00:00:41.119 to the Deep Dive. Today we're looking at a pretty 00:00:41.119 --> 00:00:43.219 comprehensive update on this whole AI ecosystem. 00:00:43.659 --> 00:00:45.960 Our mission really is to cut through the noise 00:00:45.960 --> 00:00:47.740 and show you where the real leverage points are 00:00:47.740 --> 00:00:50.219 emerging. So we'll kick things off by unpacking 00:00:50.219 --> 00:00:52.119 Anthropic's new model, kind of a sleeper hit. 00:00:52.320 --> 00:00:54.679 We'll talk about why low cost is such a game 00:00:54.679 --> 00:00:57.159 changer for designing applications. Then we're 00:00:57.159 --> 00:00:58.659 going to do a quick run through of the global 00:00:58.659 --> 00:01:01.219 power shifts, the scaling ambitions, some big 00:01:01.219 --> 00:01:03.759 picture stuff. And then finally, we dive into 00:01:03.759 --> 00:01:07.140 the deep science, that bio AI breakthrough that... 00:01:07.359 --> 00:01:09.780 Well, it might just change medicine as we know 00:01:09.780 --> 00:01:13.379 it. Okay, so let's unpack that efficiency story 00:01:13.379 --> 00:01:17.900 first. Anthropic just dropped Cloud Haiku 4 .5. 00:01:18.120 --> 00:01:20.599 And while, you know, the bigger, flashier models 00:01:20.599 --> 00:01:24.000 like Opus grab the headlines, Haiku 4 .5 might 00:01:24.000 --> 00:01:28.200 be the real story here. Because it's really targeting 00:01:28.200 --> 00:01:30.439 high performance, but without that massive price 00:01:30.439 --> 00:01:32.319 tag. Yeah, if you're actually building software 00:01:32.319 --> 00:01:35.180 or shipping AI agents today, the numbers are 00:01:35.180 --> 00:01:38.140 just... Kind of unbelievable. Tell us. This thing 00:01:38.140 --> 00:01:41.519 is two times faster and a third the cost compared 00:01:41.519 --> 00:01:43.560 to Sonnet, its closest relative in the family. 00:01:43.700 --> 00:01:45.939 Wow. We're talking about seriously cutting operating 00:01:45.939 --> 00:01:47.980 budgets for developers. Yeah. Like dramatically. 00:01:47.980 --> 00:01:50.159 Right. That's the real shift, isn't it? Operations. 00:01:50.299 --> 00:01:52.299 And here's where it gets really interesting. 00:01:52.459 --> 00:01:55.920 Okay. Haiku 4 .5's performance. It's actually 00:01:55.920 --> 00:01:58.780 on par with much larger models. Think GPT -5, 00:01:58.959 --> 00:02:01.640 Sonnet 4, even Gemini 2 .5 on key benchmarks. 00:02:01.959 --> 00:02:03.920 So it's a specialized tool, but it... Punches 00:02:03.920 --> 00:02:06.219 way above its weight class. And that speed, that 00:02:06.219 --> 00:02:09.139 cost efficiency. Yeah. It translates directly 00:02:09.139 --> 00:02:11.060 into being ready for production, right? Yeah, 00:02:11.080 --> 00:02:13.819 we heard Zencoder's CEO quoted saying this model 00:02:13.819 --> 00:02:16.599 is, quote, unlocking an entirely new set of use 00:02:16.599 --> 00:02:19.939 cases. And why? Because usually these small, 00:02:19.939 --> 00:02:23.580 fast models, they aren't quite, let's say, serious 00:02:23.580 --> 00:02:26.219 enough for complex, real -world stuff. But Haiku 00:02:26.219 --> 00:02:29.449 4 .5 is. The sources give us some technical specifics 00:02:29.449 --> 00:02:33.330 to back this up. Okay. It scored 73 % on SWE 00:02:33.330 --> 00:02:36.409 Bench, that tests software reasoning, and 41 00:02:36.409 --> 00:02:38.849 % on Terminal Bench, which is for command line 00:02:38.849 --> 00:02:41.030 tasks. Okay. Let's break those down just a bit. 00:02:41.270 --> 00:02:44.990 SWE Bench, 73 % means it can do pretty complex 00:02:44.990 --> 00:02:47.849 things, like take a GitHub repo, find a real 00:02:47.849 --> 00:02:50.430 bug, and actually fix it correctly. So really 00:02:50.430 --> 00:02:52.310 functional for software development, not just 00:02:52.310 --> 00:02:54.490 theory. Exactly. Not just theoretical potential. 00:02:54.530 --> 00:02:57.099 It works. So what does this increased capability 00:02:57.099 --> 00:03:00.219 combine with the lower price? What does it actually 00:03:00.219 --> 00:03:03.539 mean for how we build AI systems? Well, it enables 00:03:03.539 --> 00:03:05.819 something called AI sub -agent orchestration 00:03:05.819 --> 00:03:08.479 at scale. Okay, let's define that. AI sub -agent 00:03:08.479 --> 00:03:11.360 orchestration. It basically means managing lots 00:03:11.360 --> 00:03:14.659 of small, specialized AIs that all work together 00:03:14.659 --> 00:03:17.199 to complete one big, complex mission. Right. 00:03:17.280 --> 00:03:21.120 Think of it like stacking Lego blocks, maybe. 00:03:21.550 --> 00:03:23.810 but blocks of data and decision making. Good 00:03:23.810 --> 00:03:27.349 analogy. And because Haiku 4 .5 is so cheap to 00:03:27.349 --> 00:03:30.129 query, you can run dozens of these little specialized 00:03:30.129 --> 00:03:33.370 agents working in parallel without spending a 00:03:33.370 --> 00:03:36.229 fortune. That capability was just cost prohibitive 00:03:36.229 --> 00:03:39.930 last year for almost everyone. Now, it's kind 00:03:39.930 --> 00:03:42.569 of becoming the new baseline. Imagine an AI running, 00:03:42.689 --> 00:03:45.270 say, 30 different price checking agents at the 00:03:45.270 --> 00:03:47.650 same time in real time. Yeah, maybe bidding on 00:03:47.650 --> 00:03:49.990 some constantly changing inventory auction or 00:03:49.990 --> 00:03:53.270 something. That level of rapid parallel intelligence, 00:03:53.590 --> 00:03:56.289 it was financially impossible for anyone but 00:03:56.289 --> 00:03:58.930 the absolute biggest tech players until recently. 00:03:59.129 --> 00:04:00.990 And now it's accessible, much more accessible. 00:04:01.210 --> 00:04:03.349 So here's the probing question then. Does this 00:04:03.349 --> 00:04:05.590 intense focus on speed and cost fundamentally 00:04:05.590 --> 00:04:07.870 change how developers approach building large 00:04:07.870 --> 00:04:10.330 -scale AI systems? Yeah, absolutely. Low -cost 00:04:10.330 --> 00:04:12.969 models let you bake AI into every corner, even 00:04:12.969 --> 00:04:16.009 in free apps. Okay, shifting gears a bit to the 00:04:16.009 --> 00:04:19.550 wider ecosystem. We're seeing these huge global 00:04:19.550 --> 00:04:23.329 scaling ambitions colliding with some really 00:04:23.329 --> 00:04:26.290 intense market competition. Yeah, let's start 00:04:26.290 --> 00:04:28.350 with some utility applications. On the creative 00:04:28.350 --> 00:04:31.509 side, Google just dropped VO 3 .1. That's their 00:04:31.509 --> 00:04:34.170 video generation model. What's new there? Smoother 00:04:34.170 --> 00:04:37.269 control, which is great. And finally... Native 00:04:37.269 --> 00:04:39.750 audio generation built right into the video. 00:04:39.850 --> 00:04:41.930 That's actually a massive upgrade. OK, yeah. 00:04:41.990 --> 00:04:43.970 Native audio is a big deal. And for just, you 00:04:43.970 --> 00:04:46.290 know, day to day work. Claude Code is becoming 00:04:46.290 --> 00:04:49.449 this quiet powerhouse. The source material detailed 00:04:49.449 --> 00:04:53.069 like 50 creative ways non -technical folks are 00:04:53.069 --> 00:04:55.399 using it. Like who? Marketing managers, analysts, 00:04:55.660 --> 00:04:58.120 people using it to automate complex processes 00:04:58.120 --> 00:05:01.120 without needing to be coders themselves. It really 00:05:01.120 --> 00:05:03.399 shows how coding ability is being democratized. 00:05:03.459 --> 00:05:05.360 That's huge. But, you know, seeing all these 00:05:05.360 --> 00:05:08.199 powerful tools, it doesn't magically make everything 00:05:08.199 --> 00:05:10.519 simple. No, definitely not. The underlying complexity 00:05:10.519 --> 00:05:12.829 is still there. I have to admit, I still wrestle 00:05:12.829 --> 00:05:14.970 with prompt drift myself when I'm trying to build 00:05:14.970 --> 00:05:17.329 complex agents. Yeah, you start with a great 00:05:17.329 --> 00:05:19.410 idea, but the more instructions you layer in, 00:05:19.490 --> 00:05:21.410 the harder it gets sometimes to keep the agent 00:05:21.410 --> 00:05:24.170 really focused on the original goal. Right. Well, 00:05:24.689 --> 00:05:26.910 it's a surprisingly human problem in this AI 00:05:26.910 --> 00:05:29.189 world. That's a really honest admission. Yeah. 00:05:29.310 --> 00:05:31.850 Prompt engineering is still kind of an art beat. 00:05:32.370 --> 00:05:36.910 But stepping back to the big picture, the ambition, 00:05:37.089 --> 00:05:41.500 the scaling efforts are just... extreme how extreme 00:05:41.500 --> 00:05:44.000 well there are reports gemini 3 the upcoming 00:05:44.000 --> 00:05:47.519 model allegedly it cloned an entire windows operating 00:05:47.519 --> 00:05:50.339 system in one go during testing get out really 00:05:50.339 --> 00:05:53.720 that's the rumor It speaks to the raw foundational 00:05:53.720 --> 00:05:55.899 power being built behind the scenes. And that 00:05:55.899 --> 00:05:57.860 kind of engineering takes massive, massive capital. 00:05:57.899 --> 00:06:00.079 Look at OpenAI. They're apparently aiming for 00:06:00.079 --> 00:06:04.259 a $1 trillion valuation and spending $13 billion 00:06:04.259 --> 00:06:07.259 a year right now, mostly funded by user revenue, 00:06:07.379 --> 00:06:09.339 actually. That spending shows how they're constantly 00:06:09.339 --> 00:06:11.740 pushing into new areas beyond just chatbots. 00:06:11.980 --> 00:06:13.699 And you need the physical space for all that 00:06:13.699 --> 00:06:15.920 compute, right? Infrastructure. Meta's putting 00:06:15.920 --> 00:06:19.680 $1 .5 billion into a new AI -optimized data center 00:06:19.680 --> 00:06:22.660 down in El Paso. They're literally pouring concrete 00:06:22.660 --> 00:06:25.100 to build the foundation for these huge models. 00:06:25.300 --> 00:06:27.199 Cementing the infrastructure, yeah. Now let's 00:06:27.199 --> 00:06:29.160 look at the global dynamic. There was this viral 00:06:29.160 --> 00:06:32.579 chart going around, shows a massive shift. What's 00:06:32.579 --> 00:06:34.720 the shift? All the top open weight models, they're 00:06:34.720 --> 00:06:39.959 now Chinese. That's significant. This competition 00:06:39.959 --> 00:06:42.519 isn't just about business. It feels geopolitical 00:06:42.519 --> 00:06:44.779 too. It does. But, you know, the flip side of 00:06:44.779 --> 00:06:46.779 all that competition is often democratization, 00:06:46.879 --> 00:06:50.230 right? more access yeah exactly like open ai 00:06:50.230 --> 00:06:53.209 releasing a no code platform anyone can build 00:06:53.209 --> 00:06:56.209 custom ai agents now no technical skills needed 00:06:56.209 --> 00:06:59.029 oracle did something similar to rolled out 50 00:06:59.029 --> 00:07:02.149 ai agents for automating tasks no extra cost 00:07:02.149 --> 00:07:04.490 the power is definitely being pushed outward 00:07:04.490 --> 00:07:07.870 and policy is trying to keep up yeah anthropic 00:07:07.870 --> 00:07:10.709 shared Was it nine economic policy ideas specifically 00:07:10.709 --> 00:07:12.689 for governments? Yeah, that engagement shows 00:07:12.689 --> 00:07:15.009 the industry knows it has this huge societal 00:07:15.009 --> 00:07:16.670 impact. They have to be part of the conversation. 00:07:16.930 --> 00:07:19.829 OK, so probing question time. Given that rapid 00:07:19.829 --> 00:07:22.589 rise of open weight Chinese models, how does 00:07:22.589 --> 00:07:25.430 this shift the global dynamics of AI power? Well, 00:07:25.490 --> 00:07:27.850 more competition means faster innovation, more 00:07:27.850 --> 00:07:30.829 accessible tools globally. It definitely pushes 00:07:30.829 --> 00:07:34.300 Western development speed. All right. Let's pivot 00:07:34.300 --> 00:07:36.899 now. Away from the commercial side, the efficiency, 00:07:37.259 --> 00:07:40.720 the scale, and towards pure scientific discovery. 00:07:41.079 --> 00:07:44.019 This is where AI might be changing what we even 00:07:44.019 --> 00:07:46.839 thought was possible in biology. We're diving 00:07:46.839 --> 00:07:50.579 into that Google DeepMind and Yale bio AI advancement. 00:07:51.060 --> 00:07:53.819 Yeah, this is genuinely major scientific news. 00:07:53.920 --> 00:07:56.680 Yeah. They released a model called Cell Two Sentence 00:07:56.680 --> 00:08:00.560 Scale 27B. Let's call it C2S scale. C2S scale. 00:08:00.680 --> 00:08:02.720 Got it. It's a large language model. built on 00:08:02.720 --> 00:08:05.240 Google's GEMMA architecture, but it was specifically 00:08:05.240 --> 00:08:07.600 trained to deeply understand how single cells 00:08:07.600 --> 00:08:10.139 behave. So it wasn't just like reading scientific 00:08:10.139 --> 00:08:12.540 papers? No, they put it to work. It simulated 00:08:12.540 --> 00:08:14.879 the effects of over 4 ,000 different drugs across 00:08:14.879 --> 00:08:17.180 two distinct immune settings. Okay, what were 00:08:17.180 --> 00:08:19.339 those settings? Why two? It's important. They 00:08:19.339 --> 00:08:21.540 used immune context positive samples. These come 00:08:21.540 --> 00:08:23.519 from patient cells where the immune signals are 00:08:23.519 --> 00:08:25.740 weak. Think of it as a challenging real world 00:08:25.740 --> 00:08:28.779 scenario. And they also used immune context neutral 00:08:28.779 --> 00:08:31.139 settings. Those are your standard controlled 00:08:31.139 --> 00:08:34.740 lab cell cultures, more like a clean Petri dish 00:08:34.740 --> 00:08:37.799 environment. OK, so testing in both a complex 00:08:37.799 --> 00:08:41.000 patient like setting and a simpler lab setting. 00:08:41.120 --> 00:08:44.659 Exactly. And by simulating across both. The AI, 00:08:44.820 --> 00:08:47.559 well, it predicted a brand new pathway for cancer 00:08:47.559 --> 00:08:49.659 therapy. Something completely novel. Completely 00:08:49.659 --> 00:08:52.059 novel. And it's since been confirmed in actual 00:08:52.059 --> 00:08:54.419 lab experiments. This is real discovery, not 00:08:54.419 --> 00:08:56.840 just crunching existing data. Wow. How specific 00:08:56.840 --> 00:08:59.940 was the finding? Incredibly specific. The C2S 00:08:59.940 --> 00:09:02.159 scale model flagged one particular compound. 00:09:02.519 --> 00:09:05.679 Yeah. a kinase CK2 inhibitor. Its technical name 00:09:05.679 --> 00:09:10.419 is silmitacertib, but let's call it CX4945. CX4945. 00:09:10.460 --> 00:09:12.580 That's the drug target. That's the one. And the 00:09:12.580 --> 00:09:14.379 effect it predicted, it had never been reported 00:09:14.379 --> 00:09:16.559 before anywhere. Okay. So what happened when 00:09:16.559 --> 00:09:18.759 they tested it in the lab? They confirmed the 00:09:18.759 --> 00:09:21.299 AI's prediction. They found roughly a 50 % increase 00:09:21.299 --> 00:09:23.759 in something called antigen presentation. 50 00:09:23.759 --> 00:09:26.320 % increase. That sounds significant. It's massive. 00:09:26.990 --> 00:09:29.909 Antigen presentation is basically how a cancer 00:09:29.909 --> 00:09:32.870 cell signals the immune system. It lets the T 00:09:32.870 --> 00:09:35.990 cells see the tumor and attack it. Ah, okay. 00:09:36.129 --> 00:09:38.990 So boosting that visibility by 50%, that's a 00:09:38.990 --> 00:09:42.210 huge, potentially actionable step towards new 00:09:42.210 --> 00:09:44.730 cancer treatments, specifically immuno -oncology. 00:09:44.950 --> 00:09:47.870 What's really critical here, though, is how this 00:09:47.870 --> 00:09:49.850 discovery happened. You said it wasn't like alpha 00:09:49.850 --> 00:09:52.330 fold, right? Not predicting a protein's shape. 00:09:52.590 --> 00:09:54.330 Exactly. This wasn't about predicting structure. 00:09:54.370 --> 00:09:56.850 It was more like... conversational guidance. 00:09:57.090 --> 00:09:59.789 Conversational guidance with an AI. Yeah, it 00:09:59.789 --> 00:10:01.970 sounds a bit sci -fi, but drug discovery seems 00:10:01.970 --> 00:10:05.690 to be becoming, well, promptable. Promptable 00:10:05.690 --> 00:10:08.009 drug discovery. Researchers are essentially talking 00:10:08.009 --> 00:10:10.830 to the model, guiding it through incredibly complex 00:10:10.830 --> 00:10:13.529 biological data, asking it to flag potential 00:10:13.529 --> 00:10:17.019 targets that fit certain conditions. Whoa. Just 00:10:17.019 --> 00:10:19.620 imagine scaling that, using this kind of bio 00:10:19.620 --> 00:10:22.460 -AI power to analyze data for, I don't know, 00:10:22.519 --> 00:10:25.179 a billion different diseases. The potential speed 00:10:25.179 --> 00:10:28.100 up in discovery is just. It's accelerating faster 00:10:28.100 --> 00:10:30.740 than we've ever seen before, exponentially faster. 00:10:31.139 --> 00:10:33.740 Okay, probing question. How significant is it 00:10:33.740 --> 00:10:35.860 that this process was more about prompting the 00:10:35.860 --> 00:10:38.720 model, this conversational guidance, rather than 00:10:38.720 --> 00:10:41.100 the traditional structure prediction like AlphaFold? 00:10:41.240 --> 00:10:43.360 It means researchers can now conversationally 00:10:43.360 --> 00:10:46.440 guide discovery. speeding up results dramatically. 00:10:46.740 --> 00:10:49.860 It changes the workflow. What a deep dive that 00:10:49.860 --> 00:10:53.639 was. So if we try to synthesize these two huge 00:10:53.639 --> 00:10:56.600 themes we've discussed, efficiency on one side, 00:10:56.700 --> 00:10:59.440 profound discovery on the other, we're seeing 00:10:59.440 --> 00:11:02.259 a real shift in AI architecture, aren't we? Definitely. 00:11:02.539 --> 00:11:05.220 We've moved away from just focusing on raw power, 00:11:05.299 --> 00:11:08.259 like the early days of GPT -4 or Opus, towards 00:11:08.259 --> 00:11:11.259 AI that's much more specialized, optimized. kind 00:11:11.259 --> 00:11:14.139 of right size for the job. Exactly. You've got 00:11:14.139 --> 00:11:17.240 Haiku 4 .5 built purely for scalability and low 00:11:17.240 --> 00:11:18.919 cost orchestration and software development. 00:11:19.080 --> 00:11:22.059 And then you have C2S scale laser focused on 00:11:22.059 --> 00:11:24.879 deep scientific specialization like single cell 00:11:24.879 --> 00:11:27.480 biology. The future isn't just about having the 00:11:27.480 --> 00:11:29.720 biggest model anymore. No, it's about using the 00:11:29.720 --> 00:11:32.000 right size AI for the right specific problem. 00:11:32.139 --> 00:11:34.000 And the barrier to entry for doing that, it's 00:11:34.000 --> 00:11:36.379 dropping fast. Yeah. Whether you want to build 00:11:36.379 --> 00:11:39.340 sophisticated, specialized AI agents thanks to 00:11:39.340 --> 00:11:42.059 models like Haiku. Or you want to conduct fundamental 00:11:42.059 --> 00:11:45.279 scientific discovery using tools like C2S scale. 00:11:45.460 --> 00:11:49.519 The difficulty, the cost, it's just rapidly decreasing. 00:11:49.779 --> 00:11:52.080 Yeah, it really is. Which leads us to a final 00:11:52.080 --> 00:11:54.440 provocative thought for you, the listener, to 00:11:54.440 --> 00:11:57.679 consider. Okay. If the cost and the complexity 00:11:57.679 --> 00:12:00.700 of building these powerful, specialized AIs are 00:12:00.700 --> 00:12:04.399 dropping so quickly, what complex human domain, 00:12:04.519 --> 00:12:06.820 something we previously thought was impenetrable 00:12:06.820 --> 00:12:10.789 by AI, what's going to be the next one to fundamentally 00:12:10.789 --> 00:12:12.690 fall? Yeah, what field are you going to start 00:12:12.690 --> 00:12:14.669 applying this kind of thinking to? Where's the 00:12:14.669 --> 00:12:16.990 next breakthrough going to come from? Thank you 00:12:16.990 --> 00:12:19.269 for joining us for this deep dive into the latest 00:12:19.269 --> 00:12:21.970 in AI efficiency and the incredible biological 00:12:21.970 --> 00:12:24.789 frontiers it's now starting to open up. We really 00:12:24.789 --> 00:12:27.129 encourage you to dig into the links and the concepts 00:12:27.129 --> 00:12:28.649 we talked about today. There's a lot more there. 00:12:29.009 --> 00:12:29.929 Until next time.