WEBVTT 00:00:00.000 --> 00:00:03.100 Imagine a scientific paper. Okay, now imagine 00:00:03.100 --> 00:00:06.599 it has over 3 ,000 authors. Wow. Yeah, more authors 00:00:06.599 --> 00:00:09.380 than some small towns have people. Just think 00:00:09.380 --> 00:00:11.570 about that for a moment. Welcome everyone to 00:00:11.570 --> 00:00:13.990 the Deep Dive. So today we're digging into this 00:00:13.990 --> 00:00:15.730 really fascinating newsletter. We're going to 00:00:15.730 --> 00:00:17.710 unpack what it's telling us about, you know, 00:00:17.710 --> 00:00:20.750 the absolute cutting edge of AI. Our mission 00:00:20.750 --> 00:00:23.789 basically is to explore how AI research itself 00:00:23.789 --> 00:00:26.570 is changing, the incredible new tools it's building, 00:00:26.690 --> 00:00:29.969 and get this, even how it's creating labs that 00:00:29.969 --> 00:00:32.549 drive themselves. Self -driving labs. Yeah, should 00:00:32.549 --> 00:00:37.109 be some real aha moments in here for you. All 00:00:37.109 --> 00:00:38.770 right, let's dive right into that first kind 00:00:38.770 --> 00:00:41.609 of startling fact then. The recent Google Gemini 00:00:41.609 --> 00:00:45.549 2 .5 paper. Yeah. It listed, astonishingly, 3 00:00:45.549 --> 00:00:48.590 ,295 authors. I mean, just let that number sink 00:00:48.590 --> 00:00:50.689 in. Yeah, 3 ,000. And that's not a typo. Seriously. 00:00:51.329 --> 00:00:53.530 To put it in perspective, right, the first Gemini 00:00:53.530 --> 00:00:57.310 paper that had, like, 1 ,250 authors and GPT 00:00:57.310 --> 00:00:59.070 -4, you know, the open AI model. That's smaller. 00:00:59.229 --> 00:01:03.500 Way smaller. Only 417. It seems like... OpenAI 00:01:03.500 --> 00:01:05.939 and Anthropic are, I don't know, maybe a bit 00:01:05.939 --> 00:01:08.420 more selective with credits? Could be. But Google's 00:01:08.420 --> 00:01:11.920 list for Gemini, it jumped, what, 144 % in less 00:01:11.920 --> 00:01:14.120 than two years? It's wild. And what's really 00:01:14.120 --> 00:01:17.420 fascinating is that the paper apparently had 00:01:17.420 --> 00:01:20.540 this hidden message. Oh, yeah. The first initials 00:01:20.540 --> 00:01:22.959 of the, I think the first 43 authors, reportedly 00:01:22.959 --> 00:01:25.640 spelled out something like, AI is a team sport. 00:01:26.670 --> 00:01:29.049 Nice. Which isn't just, you know, a clever little 00:01:29.049 --> 00:01:31.790 Easter egg. Yeah. It feels like Google was intentionally 00:01:31.790 --> 00:01:34.769 highlighting this new reality. That AI development 00:01:34.769 --> 00:01:37.969 is now just this massive collective effort. Exactly. 00:01:37.969 --> 00:01:39.709 These papers are starting to look like phone 00:01:39.709 --> 00:01:42.409 books. Seriously. If this keeps going, I mean, 00:01:42.430 --> 00:01:44.909 just projecting it out, by 2040, we could be 00:01:44.909 --> 00:01:47.950 looking at like 2 .6 million names on one paper. 00:01:48.090 --> 00:01:50.269 You'd need AI just to read the author list. Right. 00:01:50.590 --> 00:01:52.890 We'll need AI to summarize its own author list. 00:01:53.010 --> 00:01:55.750 It's kind of funny, but also a bit... Mind bending 00:01:55.750 --> 00:01:58.150 when you think about the scale. It really makes 00:01:58.150 --> 00:01:59.750 you wonder, though, doesn't it? When you have 00:01:59.750 --> 00:02:02.909 thousands of contributors, likely all over the 00:02:02.909 --> 00:02:06.109 world, how do you even begin to coordinate something 00:02:06.109 --> 00:02:08.169 like that? Yeah. How do you track who did what 00:02:08.169 --> 00:02:10.349 specific piece? Right. It feels like we're genuinely 00:02:10.349 --> 00:02:12.610 stepping into a new era for science collaboration. 00:02:12.770 --> 00:02:17.669 More like particle physics or the human genome 00:02:17.669 --> 00:02:21.080 project, maybe. Yeah. Big science. Yeah, absolutely. 00:02:21.400 --> 00:02:23.919 And this huge growth in authors, it's not just 00:02:23.919 --> 00:02:26.280 more bodies, right? It really signals that AI 00:02:26.280 --> 00:02:28.659 isn't just a coding problem anymore, not by a 00:02:28.659 --> 00:02:32.580 long shot. It's become this grand, like interdisciplinary 00:02:32.580 --> 00:02:35.680 symphony. You've got linguists, ethicists, policy 00:02:35.680 --> 00:02:38.020 people, neuroscientists. All sorts. All working 00:02:38.020 --> 00:02:40.240 together to build something. Well, pretty unprecedented. 00:02:40.419 --> 00:02:42.740 It's not a sprint by a few geniuses anymore. 00:02:42.840 --> 00:02:45.400 It's like a huge expeditionary force. Yeah. So 00:02:45.400 --> 00:02:48.979 stepping back, this huge shift in authorship. 00:02:49.639 --> 00:02:51.759 What's it really telling us about AI development 00:02:51.759 --> 00:02:54.340 right now? Well, I think it clearly shows that 00:02:54.340 --> 00:02:57.800 AI research now demands vast, diverse and interdisciplinary 00:02:57.800 --> 00:03:00.819 collaboration. It's just that complex. OK, that 00:03:00.819 --> 00:03:02.780 makes sense. And here's where it gets really 00:03:02.780 --> 00:03:05.639 interesting beyond just the giant papers. We're 00:03:05.639 --> 00:03:08.740 seeing new AI capabilities pop up super fast. 00:03:08.860 --> 00:03:11.639 Like what? Well, breakthroughs in video generation, 00:03:11.840 --> 00:03:15.569 for one. There's a new real time model. It lets 00:03:15.569 --> 00:03:18.530 you direct a full minute of AI video. A full 00:03:18.530 --> 00:03:20.550 minute. Yeah. And you can actually adjust the 00:03:20.550 --> 00:03:22.509 prompts while it's generating. No more of those 00:03:22.509 --> 00:03:25.389 like eight second clips like VO3. That's a big 00:03:25.389 --> 00:03:29.210 jump. And I saw something about FluxPro combined 00:03:29.210 --> 00:03:32.689 with Seedence producing hyper realistic video. 00:03:32.990 --> 00:03:36.139 Oh, yeah. The results from that are. Honestly, 00:03:36.139 --> 00:03:38.120 pretty remarkable. Getting harder and harder 00:03:38.120 --> 00:03:40.780 to tell what's real anymore. But maybe the biggest 00:03:40.780 --> 00:03:42.840 thing gaining traction right now, this idea of 00:03:42.840 --> 00:03:45.719 agent mode. It's OpenAI's latest thing. Agent 00:03:45.719 --> 00:03:48.680 mode. Think of it like ChatGPT, but it's connected 00:03:48.680 --> 00:03:50.939 to a virtual computer. It can actually do stuff. 00:03:51.180 --> 00:03:53.460 So let me get this straight. An AI agent is basically, 00:03:53.599 --> 00:03:56.599 it's a program that can think and act for you. 00:03:56.680 --> 00:03:58.699 Yeah. By controlling a whole computer. Exactly. 00:03:58.719 --> 00:04:00.840 So it can carry out complex tasks all by itself. 00:04:01.180 --> 00:04:04.199 Autonomously. Precisely. It's like delegating 00:04:04.199 --> 00:04:06.960 to a really, really smart intern who can actually 00:04:06.960 --> 00:04:10.199 use your software, browse the web, write code, 00:04:10.319 --> 00:04:13.680 the whole deal. Wow. And look, Mistral just upgraded 00:04:13.680 --> 00:04:16.899 its assistant, LeChat. It's got a deep research 00:04:16.899 --> 00:04:20.279 mode now, native multilingual reasoning, even 00:04:20.279 --> 00:04:23.100 image editing built in. So it's becoming a serious 00:04:23.100 --> 00:04:25.839 alternative. Yeah, definitely a compelling open 00:04:25.839 --> 00:04:28.079 alternative to something like Microsoft Copilot. 00:04:28.220 --> 00:04:30.750 And speaking of movement. We've seen some pretty 00:04:30.750 --> 00:04:34.069 big talent shuffles too, right? Anthropic hiring 00:04:34.069 --> 00:04:36.910 back key cloud code leaders. Yeah, that was kind 00:04:36.910 --> 00:04:39.290 of an unreversed card on the usual talent poaching, 00:04:39.350 --> 00:04:41.410 wasn't it? Pretty interesting move. Definitely. 00:04:41.730 --> 00:04:44.350 And Perplexity AI, the search company, just got 00:04:44.350 --> 00:04:47.490 a huge valuation boost, like $18 billion. Yep. 00:04:47.829 --> 00:04:50.089 Positioning itself as a major rival to Google 00:04:50.089 --> 00:04:52.310 search using AI. Things are moving incredibly 00:04:52.310 --> 00:04:55.269 fast in that space. It really is. So these new 00:04:55.269 --> 00:04:58.040 agent capabilities. How do they really change 00:04:58.040 --> 00:04:59.800 how we're going to interact with AI, do you think? 00:04:59.939 --> 00:05:02.439 Well, fundamentally, it shifts AI from just being 00:05:02.439 --> 00:05:04.240 a thing you talk to to something that can actually 00:05:04.240 --> 00:05:06.300 do things for you, a digital colleague almost. 00:05:06.810 --> 00:05:09.230 Right. So in short, these agentic capabilities 00:05:09.230 --> 00:05:12.990 mean AI can now take initiative and complete 00:05:12.990 --> 00:05:15.610 complex actions on your behalf. Exactly. It's 00:05:15.610 --> 00:05:17.750 a big step. Okay. So beyond those big headlines 00:05:17.750 --> 00:05:19.910 and these new agents, there's also just this 00:05:19.910 --> 00:05:23.410 flood of new AI tools, right? Designed to automate 00:05:23.410 --> 00:05:25.629 stuff, boost creativity. Oh, yeah. It's like 00:05:25.629 --> 00:05:28.209 a Cambrian explosion of tools right now. Take 00:05:28.209 --> 00:05:30.589 Uncursor, for instance. It lets you build and 00:05:30.589 --> 00:05:34.519 deploy live apps. in literally seconds in seconds 00:05:34.519 --> 00:05:37.060 yeah imagine what that unlocks for people who 00:05:37.060 --> 00:05:39.620 aren't coders the speed of innovation there that's 00:05:39.620 --> 00:05:43.899 huge or fast3d .io you type text or feed it an 00:05:43.899 --> 00:05:45.860 image and boom eight seconds later you have a 00:05:45.860 --> 00:05:49.100 3d model eight seconds That could totally transform 00:05:49.100 --> 00:05:51.180 design workflows. Artists could iterate like 00:05:51.180 --> 00:05:53.240 crazy. Totally. And then there's stuff like force 00:05:53.240 --> 00:05:56.139 equals, which claims to turn just an idea into 00:05:56.139 --> 00:05:59.439 like a full execution ready plan. Interesting. 00:05:59.579 --> 00:06:02.759 And symbol, which can take a long PDF or an article 00:06:02.759 --> 00:06:05.120 and just turn it into a video tutorial for you. 00:06:05.199 --> 00:06:07.480 So it's all about making complex things easier, 00:06:07.579 --> 00:06:10.399 faster, radically supercharging what one person 00:06:10.399 --> 00:06:12.540 can do. Pretty much. And then you have the quick 00:06:12.540 --> 00:06:16.579 hits. Seen some viral AI art coming from Perplexity's 00:06:16.579 --> 00:06:20.639 Ambassador. And an interesting little note that 00:06:20.639 --> 00:06:23.100 Anthropic might be quietly tightening the usage 00:06:23.100 --> 00:06:25.600 limits on Claude for some people. Something to 00:06:25.600 --> 00:06:27.819 keep an eye on. Yeah, definitely. And Adobe's 00:06:27.819 --> 00:06:31.279 new tool, turning silly noises into realistic 00:06:31.279 --> 00:06:33.339 audio effects. That just sounds like pure fun. 00:06:33.500 --> 00:06:36.199 Right, a creative playground. And OpenAI is apparently 00:06:36.199 --> 00:06:39.319 working on a native checkout system. So the AI 00:06:39.319 --> 00:06:42.189 could potentially... complete purchases for you. 00:06:42.250 --> 00:06:45.370 Whoa. Okay, that's interesting implications there. 00:06:45.490 --> 00:06:48.050 And Microsoft Copilot can now apparently see 00:06:48.050 --> 00:06:50.410 whatever's on your screen in real time. Real 00:06:50.410 --> 00:06:52.189 time screen awareness. Yeah, that adds a whole 00:06:52.189 --> 00:06:55.350 new layer of context for the AI. It's amazing, 00:06:55.490 --> 00:06:57.829 really. All these tools, all these rapid developments, 00:06:57.910 --> 00:07:00.410 they're incredible. But honestly. Yeah, I still 00:07:00.410 --> 00:07:03.209 wrestle with just the sheer volume of it all. 00:07:03.470 --> 00:07:05.870 Yeah. Yeah. It's a lot to keep up with. Oh, tell 00:07:05.870 --> 00:07:07.189 me about it. You're definitely not alone there. 00:07:07.230 --> 00:07:08.889 It feels like drinking from a fire hose most 00:07:08.889 --> 00:07:11.350 days. Right. But what's fascinating, I think, 00:07:11.350 --> 00:07:13.550 is that underneath all this rapid fire stuff, 00:07:13.649 --> 00:07:15.629 there's actually a pretty clear pattern emerging, 00:07:15.790 --> 00:07:18.189 isn't there? Yeah, I think so. For all their 00:07:18.189 --> 00:07:20.509 differences, what's the common thread you see 00:07:20.509 --> 00:07:23.550 weaving through these diverse new AI tools? Well, 00:07:23.589 --> 00:07:25.529 they're not just automating tasks, you know. 00:07:25.589 --> 00:07:28.350 They're fundamentally lowering the barrier to 00:07:28.350 --> 00:07:31.389 creating complex things or getting complex things 00:07:31.389 --> 00:07:34.089 done. Lowering the barrier. Yeah. Okay. So the 00:07:34.089 --> 00:07:36.290 common thread is that they significantly simplify 00:07:36.290 --> 00:07:39.189 complex processes and dramatically accelerate 00:07:39.189 --> 00:07:41.389 creative output. Yeah. I think that sums it up 00:07:41.389 --> 00:07:44.050 nicely. They empower the user in a big way. Okay. 00:07:44.110 --> 00:07:47.430 So if we kind of connect that idea of empowerment 00:07:47.430 --> 00:07:51.170 and automation to the really big picture, we're 00:07:51.170 --> 00:07:53.470 starting to see AI not just assist with discovery, 00:07:53.649 --> 00:07:56.129 but actually drive it. Exactly. Which brings 00:07:56.129 --> 00:07:58.350 us to this work at North Carolina State University. 00:07:59.019 --> 00:08:01.399 Researchers there have built a fully autonomous 00:08:01.399 --> 00:08:04.360 AI -powered chemistry lab. Fully autonomous. 00:08:04.620 --> 00:08:07.740 Like no humans involved day to day. Pretty much. 00:08:07.920 --> 00:08:10.560 This lab runs continuously 24 -7. It collects 00:08:10.560 --> 00:08:12.480 something like 10 times more data than traditional 00:08:12.480 --> 00:08:15.959 lab methods. And it's finding promising materials 00:08:15.959 --> 00:08:19.240 for things like clean energy, advanced electronics. 00:08:19.699 --> 00:08:22.160 Finding them in days, not the years it usually 00:08:22.160 --> 00:08:24.339 takes. It's a total game changer for material 00:08:24.339 --> 00:08:26.439 science. So it's not just running one experiment 00:08:26.439 --> 00:08:29.439 and stopping. Nope. It captures a new data point 00:08:29.439 --> 00:08:32.360 literally every half second. It analyzes that 00:08:32.360 --> 00:08:34.700 data, learns, and then adjusts the next step 00:08:34.700 --> 00:08:37.059 of the experiment mid -run. On the fly. On the 00:08:37.059 --> 00:08:40.179 fly. And it never stops testing. It's this constant, 00:08:40.340 --> 00:08:43.940 intelligent loop of discovery. Design, test, 00:08:44.360 --> 00:08:47.139 learn, repeat, nonstop. And you mentioned it's 00:08:47.139 --> 00:08:50.529 efficient, too. Yeah. Big time. Uses fewer chemicals, 00:08:50.690 --> 00:08:53.549 cuts down lab waste drastically. So it's actually 00:08:53.549 --> 00:08:56.129 enabling sustainable science acceleration. That's 00:08:56.129 --> 00:08:58.230 a huge win -win. Whoa. Okay. Just pause there. 00:08:58.529 --> 00:09:01.429 Imagine scaling that up. A self -driving lab 00:09:01.429 --> 00:09:04.070 like this, discovering materials for, I don't 00:09:04.070 --> 00:09:05.730 know, a billion new applications we haven't even 00:09:05.730 --> 00:09:07.830 dreamed of yet. Right. It's like Netflix for 00:09:07.830 --> 00:09:09.889 chemical reactions. Just streaming experiments 00:09:09.889 --> 00:09:13.200 constantly. No pauses. Things that normally take 00:09:13.200 --> 00:09:16.100 researchers weeks or months, done in hours or 00:09:16.100 --> 00:09:19.500 days. My mind kind of explodes thinking about 00:09:19.500 --> 00:09:22.500 what this could mean for medicine, for battery 00:09:22.500 --> 00:09:25.500 technology, for, well, everything. What makes 00:09:25.500 --> 00:09:28.639 this really stand out, though, is, well, we have 00:09:28.639 --> 00:09:30.600 systems like DeepMind's AlphaFold, right? Right. 00:09:31.120 --> 00:09:34.620 Predicting protein structures or NVIDIA's drug 00:09:34.620 --> 00:09:37.039 discovery pipelines. Sure, powerful tools. But 00:09:37.039 --> 00:09:38.799 those are largely computation, right? They don't 00:09:38.799 --> 00:09:40.899 typically operate physical experiments in real 00:09:40.899 --> 00:09:44.120 time in a lab. Exactly. This NC State lab, it 00:09:44.120 --> 00:09:46.659 fully automates the entire loop design, test, 00:09:46.960 --> 00:09:49.279 learn, repeat, all within the physical world. 00:09:49.399 --> 00:09:51.279 It's a closed loop system actually doing the 00:09:51.279 --> 00:09:53.200 chemistry. It's the integration that's key. Totally. 00:09:53.240 --> 00:09:56.460 So now that this kind of autonomous lab is clearly 00:09:56.460 --> 00:09:59.539 possible. It kind of begs the question, right? 00:09:59.600 --> 00:10:01.679 For you, for us, for everyone listening, where's 00:10:01.679 --> 00:10:04.879 the open AI of chemistry or the clod for clean 00:10:04.879 --> 00:10:07.639 energy materials? What's the next step in really 00:10:07.639 --> 00:10:10.100 leveraging this kind of autonomy for scientific 00:10:10.100 --> 00:10:13.080 breakthroughs? It's a huge question. Yeah. But 00:10:13.080 --> 00:10:15.360 it's undeniable how this kind of self -driving 00:10:15.360 --> 00:10:18.000 lab revolutionizes the pace of discovery. Yeah. 00:10:18.059 --> 00:10:20.600 How would you summarize that impact? I'd say 00:10:20.600 --> 00:10:23.840 it enables... continuous rapid experimentation, 00:10:24.419 --> 00:10:27.200 basically compressing years of traditional research 00:10:27.200 --> 00:10:33.159 work into mere days. Speed and scale. Years into 00:10:33.159 --> 00:10:35.759 days. Okay. Sponsor. All right. So let's try 00:10:35.759 --> 00:10:37.259 to wrap our heads around all this. What does 00:10:37.259 --> 00:10:39.740 it all mean for us, for you listening? We've 00:10:39.740 --> 00:10:43.659 seen AI research itself scale dramatically, needing 00:10:43.659 --> 00:10:46.100 thousands of collaborators now. Yeah. The scale 00:10:46.100 --> 00:10:48.820 is next. We've seen this explosion of new AI 00:10:48.820 --> 00:10:51.559 tools and these aging capabilities really empowering 00:10:51.559 --> 00:10:55.919 users to do much more, much faster. Democratizing 00:10:55.919 --> 00:10:58.460 complexity almost. And we've explored the emergence 00:10:58.460 --> 00:11:01.779 of these fully autonomous. AI systems like that 00:11:01.779 --> 00:11:03.840 chemistry lab that can accelerate scientific 00:11:03.840 --> 00:11:06.480 discovery at just an unprecedented rate. Yeah. 00:11:06.559 --> 00:11:08.279 When you put it all together, the sheer scale 00:11:08.279 --> 00:11:10.600 of the research effort, the rise of agents that 00:11:10.600 --> 00:11:13.200 can act for us and now labs that literally think 00:11:13.200 --> 00:11:15.320 and experiment for themselves. Right. It's becoming 00:11:15.320 --> 00:11:17.539 really clear that AI isn't just, you know, assisting 00:11:17.539 --> 00:11:20.240 us anymore. It's evolving into an active, intelligent 00:11:20.240 --> 00:11:22.679 partner. A partner in discovery and creation, 00:11:22.860 --> 00:11:25.799 even in basic science. So here's maybe a final 00:11:25.799 --> 00:11:28.940 thought to leave you with. If AI research papers 00:11:28.940 --> 00:11:32.840 now need thousands of authors and labs can design, 00:11:33.059 --> 00:11:37.080 test, and learn entirely on their own, what happens 00:11:37.080 --> 00:11:40.240 to the role of human curiosity? Human creativity. 00:11:40.720 --> 00:11:43.519 Where do we fit in a world increasingly driven 00:11:43.519 --> 00:11:45.799 by this kind of autonomous intelligence? That's 00:11:45.799 --> 00:11:48.059 the big question, isn't it? A really profound 00:11:48.059 --> 00:11:50.279 shift to think about. It really is. Well, we 00:11:50.279 --> 00:11:52.480 hope this deep dive gave you some valuable insights 00:11:52.480 --> 00:11:54.759 today, maybe sparked some of your own curiosity 00:11:54.759 --> 00:11:56.659 about where all this is heading. Yeah, and if 00:11:56.659 --> 00:11:58.299 something particularly struck you or made you 00:11:58.299 --> 00:12:00.460 think, definitely let us know. We're curious, 00:12:00.519 --> 00:12:02.440 too. Absolutely. Thanks for joining us for this 00:12:02.440 --> 00:12:05.220 deep dive into the latest in AI. Until next time, 00:12:05.259 --> 00:12:06.320 Aotiro Music.