WEBVTT 00:00:00.000 --> 00:00:03.200 Okay, so picture this. AI isn't just your super 00:00:03.200 --> 00:00:05.240 smart friend you ask questions anymore. It's 00:00:05.240 --> 00:00:07.179 actually showing up with, you know, fully formed 00:00:07.179 --> 00:00:10.039 ideas like pitching a startup, but for research 00:00:10.039 --> 00:00:12.039 papers. Yeah, what's really fascinating here 00:00:12.039 --> 00:00:14.660 is just how quickly the AI landscape is evolving, 00:00:14.880 --> 00:00:16.920 right? It's shifting pretty dramatically, you 00:00:16.920 --> 00:00:19.600 know, from just processing information or like 00:00:19.600 --> 00:00:23.230 generating text based on prompts to... actively 00:00:23.230 --> 00:00:25.969 creating and proposing genuinely novel concepts. 00:00:26.190 --> 00:00:28.089 Exactly. It's a whole different ballgame, isn't 00:00:28.089 --> 00:00:29.649 it? And that's what we're really diving into 00:00:29.649 --> 00:00:32.890 today. This deep dive is all about these AI agents 00:00:32.890 --> 00:00:35.049 that are proposing original research ideas, kind 00:00:35.049 --> 00:00:36.689 of like they're pitching a new company concept. 00:00:36.789 --> 00:00:39.109 But yeah, for the world of science and academia. 00:00:39.799 --> 00:00:42.600 And we're pulling insights from this pretty significant 00:00:42.600 --> 00:00:45.060 recent report. It feels like a dispatch straight 00:00:45.060 --> 00:00:47.799 from, I guess, the front lines of AI development. 00:00:47.960 --> 00:00:50.799 It details how these cutting edge idea generating 00:00:50.799 --> 00:00:53.560 agents are actually being put to the test in 00:00:53.560 --> 00:00:55.920 pretty rigorous ways. Yeah, it's like getting 00:00:55.920 --> 00:00:58.479 the inside scoop on what's next. So our mission 00:00:58.479 --> 00:01:00.820 today really is to unpack this source material 00:01:00.820 --> 00:01:03.960 for you. We want to get into what these AI idea 00:01:03.960 --> 00:01:06.760 generators are actually proving capable of right 00:01:06.760 --> 00:01:09.579 now, how they were evaluated in this specific 00:01:09.579 --> 00:01:13.459 benchmark test, and maybe most importantly, what 00:01:13.459 --> 00:01:15.379 the results actually mean for you, whether you're 00:01:15.379 --> 00:01:19.060 using AI for brainstorming, for ideation in your 00:01:19.060 --> 00:01:20.620 own work, you know, whatever creative process 00:01:20.620 --> 00:01:22.319 you're involved in. Let's unpack this. Right. 00:01:22.579 --> 00:01:26.469 It's beyond just hype now. There are concrete 00:01:26.469 --> 00:01:29.730 benchmarks attempting to measure creativity and 00:01:29.730 --> 00:01:32.810 innovation in AI, which is, you know, a big step. 00:01:32.909 --> 00:01:34.790 Right. And the core of this report, you know, 00:01:34.790 --> 00:01:36.769 the main event is this test. They called it the 00:01:36.769 --> 00:01:40.010 AI Idea Bench 2025. It sounds a bit formal, maybe, 00:01:40.109 --> 00:01:42.269 but the premise is pretty cool, I think. They 00:01:42.269 --> 00:01:44.989 essentially put four leading AI idea generator 00:01:44.989 --> 00:01:47.370 agents through their paces like a competition. 00:01:47.670 --> 00:01:49.590 And crucially, they didn't just give them a broad 00:01:49.590 --> 00:01:52.290 topic and say, write something. They designed 00:01:52.290 --> 00:01:56.510 the test framework to really mimic how venture 00:01:56.510 --> 00:01:59.790 capitalists, you know, vet startups. They looked 00:01:59.790 --> 00:02:02.790 at these AI generated ideas based on three key 00:02:02.790 --> 00:02:05.090 criteria that matter in the real world, whether 00:02:05.090 --> 00:02:07.769 it's a business or research idea. Did the idea 00:02:07.769 --> 00:02:10.120 turn? get the right problem? Was it genuinely 00:02:10.120 --> 00:02:13.300 new? That's the novelty aspect. And did it actually 00:02:13.300 --> 00:02:15.360 look like something you could realistically build 00:02:15.360 --> 00:02:17.740 or implement? That's the feasibility side. Oh, 00:02:17.860 --> 00:02:20.379 applying that startup lens to research ideas. 00:02:20.520 --> 00:02:22.479 That makes so much sense. You kind of need all 00:02:22.479 --> 00:02:25.039 three for an idea to really. go anywhere, right? 00:02:25.099 --> 00:02:26.939 But here's where it gets really interesting, 00:02:27.060 --> 00:02:28.860 maybe even a little mind -bending, the ground 00:02:28.860 --> 00:02:31.580 truth they used for comparison. They didn't just 00:02:31.580 --> 00:02:33.699 have human experts guess if it was good. They 00:02:33.699 --> 00:02:36.240 compared the AI ideas against a massive data 00:02:36.240 --> 00:02:40.439 set, 3 ,495 brand new AI papers that were actually 00:02:40.439 --> 00:02:43.120 accepted and presented at top conferences recently. 00:02:43.400 --> 00:02:46.900 This part is absolutely key to the study's validity, 00:02:47.060 --> 00:02:49.580 I think. They were incredibly meticulous about 00:02:49.580 --> 00:02:52.949 the timing. Zero of this data. Zero of these 00:02:52.949 --> 00:02:57.469 3 ,495 research papers existed before the knowledge 00:02:57.469 --> 00:02:59.830 cutoff date for the models they were testing, 00:02:59.969 --> 00:03:03.110 specifically GPT -4io's cutoff date of October 00:03:03.110 --> 00:03:06.770 3, 2023. Whoa, like zero. None of it existed 00:03:06.770 --> 00:03:09.590 anywhere for the AI to have potentially seen 00:03:09.590 --> 00:03:12.530 it during training. Exactly. Zero. This completely 00:03:12.530 --> 00:03:14.729 eliminates the possibility that the AI was just, 00:03:14.789 --> 00:03:16.590 you know, echoing its training data or variations 00:03:16.590 --> 00:03:18.289 of knowledge that already exist in the public 00:03:18.289 --> 00:03:20.490 domain before that date. This was a true test 00:03:20.490 --> 00:03:22.000 of whether it could propose... something that 00:03:22.000 --> 00:03:25.020 was genuinely post -cutoff, something new to 00:03:25.020 --> 00:03:26.939 the world after its knowledge was frozen. Okay. 00:03:27.039 --> 00:03:29.199 That really isolates its ability to generate 00:03:29.199 --> 00:03:31.080 something novel, something new, which is like 00:03:31.080 --> 00:03:33.219 the whole point of ideation, right? Yeah. How 00:03:33.219 --> 00:03:35.740 did they actually score the ideas against this 00:03:35.740 --> 00:03:37.900 unseen data? They used a pretty smart two -step 00:03:37.900 --> 00:03:40.500 scoring process. First, they looked at how well 00:03:40.500 --> 00:03:43.819 the AI's idea aligned with a real paper that 00:03:43.819 --> 00:03:45.800 actually got published in that post -cutoff data 00:03:45.800 --> 00:03:48.719 set. Does it identify the same problem? Did it 00:03:48.719 --> 00:03:50.979 propose similar approaches or experiments? designs 00:03:50.979 --> 00:03:53.300 that measured the idea's relevance and depth. 00:03:53.479 --> 00:03:56.740 OK, so how close was it to a real new paper? 00:03:56.879 --> 00:03:59.400 Right. And second, they calculated a combined 00:03:59.400 --> 00:04:03.189 score for novelty and feasibility. And this part 00:04:03.189 --> 00:04:07.669 used citation data. Citation data. For novelty 00:04:07.669 --> 00:04:10.270 and feasibility, that feels a bit counterintuitive, 00:04:10.270 --> 00:04:12.009 doesn't it? Yeah. How does whether something 00:04:12.009 --> 00:04:14.449 is cited relate to whether it's new or doable? 00:04:14.750 --> 00:04:17.389 Well, it's used as a proxy, right? So for novelty, 00:04:17.509 --> 00:04:20.310 if an AI proposes an idea or a method and there 00:04:20.310 --> 00:04:22.529 are fewer existing papers citing similar work, 00:04:22.649 --> 00:04:25.149 that could indicate a more novel idea. It's less 00:04:25.149 --> 00:04:27.089 connected to the existing body of knowledge, 00:04:27.189 --> 00:04:29.529 you see. Okay. Fewer citations of similar things 00:04:29.529 --> 00:04:32.189 means it's maybe more novel. Got it. Makes sense. 00:04:32.350 --> 00:04:35.610 And for feasibility, looking at citation data 00:04:35.610 --> 00:04:37.949 related to the specific methods proposed in the 00:04:37.949 --> 00:04:41.569 AI idea can be a pretty powerful proxy. If the 00:04:41.569 --> 00:04:44.250 methods an AI suggests are already being adopted 00:04:44.250 --> 00:04:46.829 and cited frequently in recent research, it suggests 00:04:46.829 --> 00:04:49.410 they are more practical, more established, more 00:04:49.410 --> 00:04:52.759 buildable right now. The source notes that this 00:04:52.759 --> 00:04:55.060 citation -weighted feasibility score is actually 00:04:55.060 --> 00:04:58.389 a handy proxy for uh commercial traction too 00:04:58.389 --> 00:05:00.610 it tracks methods that are already demonstrating 00:05:00.610 --> 00:05:02.829 usefulness and adoption in the field kind of 00:05:02.829 --> 00:05:04.790 like market validation for the techniques themselves 00:05:04.790 --> 00:05:07.290 oh i see it's not just about whether the idea 00:05:07.290 --> 00:05:09.430 exists but whether the building blocks the ai 00:05:09.430 --> 00:05:11.850 is suggesting are already proving useful in practice 00:05:11.850 --> 00:05:14.350 like it's taking a pulse check on the technical 00:05:14.350 --> 00:05:16.329 approaches to see which ones are gaining traction 00:05:16.329 --> 00:05:19.550 and seem well practical precisely gives you a 00:05:19.550 --> 00:05:22.329 sense of how grounded in current workable techniques 00:05:22.329 --> 00:05:25.629 the idea is which is you know super useful fascinating 00:05:25.629 --> 00:05:28.949 so So who are these agents and how do they actually 00:05:28.949 --> 00:05:31.129 perform in this benchmark? What did the results 00:05:31.129 --> 00:05:33.490 show? Yeah, so the report specifically highlighted 00:05:33.490 --> 00:05:37.029 two, AI scientist and AI researcher, mainly because 00:05:37.029 --> 00:05:39.189 they showed different strengths, which is interesting 00:05:39.189 --> 00:05:42.930 in itself. AI scientist, this one was frankly 00:05:42.930 --> 00:05:46.829 amazing at alignment. It hit a perfect 5 .0 score 00:05:46.829 --> 00:05:49.889 on both the motivation for the research and the 00:05:49.889 --> 00:05:53.060 experiment design parts. A perfect 5 .0. Wow. 00:05:53.139 --> 00:05:54.759 So it didn't just get the gist. It got the why 00:05:54.759 --> 00:05:58.199 and the detailed how perfectly. It really nailed 00:05:58.199 --> 00:06:00.759 the concept and the proposed execution. Yeah. 00:06:00.879 --> 00:06:03.579 So perfectly matching the core problem and the 00:06:03.579 --> 00:06:06.199 detailed plan of what a human researcher came 00:06:06.199 --> 00:06:08.680 up with and actually got published after the 00:06:08.680 --> 00:06:11.079 cutoff. That's pretty wild. And maybe unsurprisingly, 00:06:11.100 --> 00:06:13.199 given that depth of alignment with what was truly 00:06:13.199 --> 00:06:15.899 novel and published, it also topped the charts 00:06:15.899 --> 00:06:19.350 for novelty overall. So AI scientists seems to 00:06:19.350 --> 00:06:21.850 be the agent you'd go to for generating bold, 00:06:22.069 --> 00:06:24.529 deeply aligned and potentially really novel ideas. 00:06:24.790 --> 00:06:27.029 OK, bold and novel, maybe the big swings, the 00:06:27.029 --> 00:06:28.670 breakthrough stuff. What about the other one 00:06:28.670 --> 00:06:30.470 you mentioned, AI researcher? Did it have a different 00:06:30.470 --> 00:06:33.709 profile? It did. AI researchers scored best on 00:06:33.709 --> 00:06:36.569 feasibility per step. It's hard to give the exact 00:06:36.569 --> 00:06:39.029 number without the study's full context. But 00:06:39.029 --> 00:06:41.350 the source mentions a specific number, 17 times 00:06:41.350 --> 00:06:43.810 10 to the minus three, I think. But the point 00:06:43.810 --> 00:06:45.829 was that its ideas looked more practical, more. 00:06:46.509 --> 00:06:48.509 buildable right now compared to the others in 00:06:48.509 --> 00:06:54.149 the test ah so maybe less blue sky more grounded 00:06:54.149 --> 00:06:57.889 one for big new concepts and one for ideas that 00:06:57.889 --> 00:07:00.449 seem more ready to actually implement or you 00:07:00.449 --> 00:07:02.310 know build upon today that seems to be the clear 00:07:02.310 --> 00:07:05.209 distinction emerging from the scoring yeah The 00:07:05.209 --> 00:07:07.170 citation -weighted feasibility score indicated 00:07:07.170 --> 00:07:09.449 that AI research was surfacing plans or methods 00:07:09.449 --> 00:07:11.430 that align more with techniques already gaining 00:07:11.430 --> 00:07:14.089 significant adoption or maybe even commercial 00:07:14.089 --> 00:07:17.389 traction in the field. Okay, this study is super 00:07:17.389 --> 00:07:19.029 interesting from a technical standpoint, but 00:07:19.029 --> 00:07:20.490 let's translate this. What does this actually 00:07:20.490 --> 00:07:22.850 mean for, you know, me, the person trying to 00:07:22.850 --> 00:07:25.250 use AI for brainstorming or come up with new 00:07:25.250 --> 00:07:27.709 projects? What are the practical takeaways from 00:07:27.709 --> 00:07:29.290 this whole thing? All right, this is where we 00:07:29.290 --> 00:07:31.470 really get to the so what's for you. The first 00:07:31.470 --> 00:07:33.310 big takeaway is about quality having layers. 00:07:33.870 --> 00:07:37.750 You know, just because an AI idea sounds relevant 00:07:37.750 --> 00:07:40.509 or even aligns perfectly with a problem like 00:07:40.509 --> 00:07:43.129 AI scientists did, doesn't automatically mean 00:07:43.129 --> 00:07:45.769 the details are there to actually implement it 00:07:45.769 --> 00:07:48.810 easily. There's a clear difference between nailing 00:07:48.810 --> 00:07:51.589 the core concept and providing a truly practical, 00:07:51.829 --> 00:07:54.449 buildable plan. Yeah, like the AI scientists 00:07:54.449 --> 00:07:57.439 got the what and the why down cold. Maybe perfectly, 00:07:57.620 --> 00:07:59.759 but maybe AI researcher was better at the how, 00:07:59.839 --> 00:08:02.199 the actual practical steps. Exactly. Think of 00:08:02.199 --> 00:08:05.199 it like maybe an architect providing a beautiful 00:08:05.199 --> 00:08:08.480 visionary drawing versus the detailed blueprints 00:08:08.480 --> 00:08:11.100 and engineering plans needed to actually construct 00:08:11.100 --> 00:08:13.360 the building. Both are necessary, right? But 00:08:13.360 --> 00:08:15.759 they represent different layers of quality or 00:08:15.759 --> 00:08:17.459 usefulness depending on what you need right now. 00:08:17.680 --> 00:08:19.699 I like that analogy. Vision versus blueprint. 00:08:19.959 --> 00:08:22.139 Okay, that makes sense. What else should we take 00:08:22.139 --> 00:08:24.860 away? Well, that citation -aware scoring approach 00:08:24.860 --> 00:08:27.639 they used. That concept itself is a valuable 00:08:27.639 --> 00:08:30.360 lens you can apply, even outside of this specific 00:08:30.360 --> 00:08:33.580 study. When an AI gives you an idea, thinking 00:08:33.580 --> 00:08:35.679 about how much related work or how many already 00:08:35.679 --> 00:08:38.419 adopted methodologies exist for it, you know, 00:08:38.419 --> 00:08:40.759 checking the citation pulse of the underlying 00:08:40.759 --> 00:08:43.799 techniques is like getting a market or technical 00:08:43.799 --> 00:08:47.460 feasibility signal. It's a quick check. If an 00:08:47.460 --> 00:08:50.440 AI suggests a method for, I don't know, analyzing 00:08:50.440 --> 00:08:53.019 customer data, I could kind of ask myself, based 00:08:53.019 --> 00:08:55.179 on what I know, are people actually using this 00:08:55.179 --> 00:08:57.639 method successfully? Is it proven? Is it gaining 00:08:57.639 --> 00:09:00.620 traction? Right. It's a way to gauge how speculative 00:09:00.620 --> 00:09:03.440 or how grounded the idea is in current practice. 00:09:03.820 --> 00:09:05.960 And then there's the critical point about testing 00:09:05.960 --> 00:09:08.799 against truly unseen data. If you're relying 00:09:08.799 --> 00:09:11.779 on AI for genuinely fresh brainstorming, for 00:09:11.779 --> 00:09:13.700 coming up with ideas that are new to you and 00:09:13.700 --> 00:09:15.720 hopefully new to the world, you really need to 00:09:15.720 --> 00:09:18.659 be wary of that training set echo. Training set 00:09:18.659 --> 00:09:20.559 echo, yeah. I like that term. It feels like the 00:09:20.559 --> 00:09:22.679 AI is just humming a tune it heard before, maybe 00:09:22.679 --> 00:09:24.840 slightly remixed, not composing something really 00:09:24.840 --> 00:09:27.840 new. Precisely. If the models you're using for 00:09:27.840 --> 00:09:30.600 brainstorming haven't been rigorously tested 00:09:30.600 --> 00:09:33.019 against information they absolutely could not 00:09:33.019 --> 00:09:36.100 have seen during training, like... In this benchmark, 00:09:36.320 --> 00:09:38.480 you can't be sure if they're generating genuinely 00:09:38.480 --> 00:09:41.000 novel ideas or just variations of things that 00:09:41.000 --> 00:09:43.299 already exist. That kind of testing is vital 00:09:43.299 --> 00:09:46.120 if novelty is really your goal. That makes total 00:09:46.120 --> 00:09:47.980 sense. You want to know if it's actually inventing 00:09:47.980 --> 00:09:49.940 something or just giving you a slightly different 00:09:49.940 --> 00:09:51.559 version of something you already know is out 00:09:51.559 --> 00:09:53.279 there. Maybe even something I already know is 00:09:53.279 --> 00:09:55.620 out there. And this really leads directly to 00:09:55.620 --> 00:09:58.919 the idea of matching the tool to the task. Don't 00:09:58.919 --> 00:10:01.120 fall into the trap of looking for one single 00:10:01.120 --> 00:10:04.940 best AI agent for all your ideation needs. Based 00:10:04.940 --> 00:10:07.279 on this study, AI scientists seems like the one 00:10:07.279 --> 00:10:10.220 you'd use for inspiring those bold, maybe slightly 00:10:10.220 --> 00:10:13.059 more theoretical, highly novel ideas that could 00:10:13.059 --> 00:10:15.259 really push boundaries. The big breakthrough 00:10:15.259 --> 00:10:18.779 concept ideas, the real moonshots, maybe. Yeah, 00:10:18.820 --> 00:10:21.860 the big swings. Well, AI researcher. with its 00:10:21.860 --> 00:10:24.440 higher feasibility score, seems better suited 00:10:24.440 --> 00:10:26.720 for surfacing plans or ideas you could actually 00:10:26.720 --> 00:10:29.860 ship or build upon right now using methods that 00:10:29.860 --> 00:10:31.779 are already proving their worth. Right. You need 00:10:31.779 --> 00:10:33.659 to pick the agent or maybe even use different 00:10:33.659 --> 00:10:36.259 agents in different stages that fits your specific 00:10:36.259 --> 00:10:39.570 goal for that brainstorming session. What are 00:10:39.570 --> 00:10:41.909 you trying to do today? So it's less about finding 00:10:41.909 --> 00:10:44.870 the perfect AI unicorn and more about understanding 00:10:44.870 --> 00:10:47.490 the strengths of different tools and using the 00:10:47.490 --> 00:10:49.870 right one for the right job you need done today, 00:10:50.070 --> 00:10:51.870 like having different tools in your toolbox. 00:10:52.190 --> 00:10:54.210 Exactly. And, you know, this applies whether 00:10:54.210 --> 00:10:56.789 you're in academic research or, as the source 00:10:56.789 --> 00:10:59.090 mentions, on a marketing team needing campaign 00:10:59.090 --> 00:11:01.990 ideas, a product team brainstorming new features, 00:11:02.090 --> 00:11:04.690 or even an investment team evaluating potential 00:11:04.690 --> 00:11:08.159 market gaps. When you turn to generative AI for 00:11:08.159 --> 00:11:11.480 ideation, explicitly apply this two -step filter 00:11:11.480 --> 00:11:14.379 derived from this study. First, is the idea truly 00:11:14.379 --> 00:11:16.399 on target for the problem you're trying to solve? 00:11:16.539 --> 00:11:19.139 Okay, step one, relevance. And second, can it 00:11:19.139 --> 00:11:21.039 actually be built or implemented with reasonable 00:11:21.039 --> 00:11:25.740 effort? Step two, feasibility. Is it relevant 00:11:25.740 --> 00:11:30.820 and is it doable? Right. Using that filter consistently 00:11:30.820 --> 00:11:33.940 helps you cut through the noise of maybe brilliant 00:11:33.940 --> 00:11:37.039 sounding but impractical ideas faster and helps 00:11:37.039 --> 00:11:39.519 you spot potentially winning actionable ideas 00:11:39.519 --> 00:11:42.159 more efficiently. It gives you a framework. That's 00:11:42.159 --> 00:11:43.879 super practical. It gives you like a concrete 00:11:43.879 --> 00:11:46.519 framework to evaluate what AI spits out instead 00:11:46.519 --> 00:11:49.360 of just feeling overwhelmed or unsure if the 00:11:49.360 --> 00:11:51.720 idea is any good or just noise. It does. It moves 00:11:51.720 --> 00:11:55.419 you from passively receiving AI output to. actively 00:11:55.419 --> 00:11:57.539 and critically evaluating it like a good investor 00:11:57.539 --> 00:11:59.720 or good editor really okay so that study gives 00:11:59.720 --> 00:12:02.679 us a deep look at ai's ideation capability which 00:12:02.679 --> 00:12:05.480 is you know a huge step but the source also touches 00:12:05.480 --> 00:12:07.899 on how ai is impacting other areas right now 00:12:07.899 --> 00:12:10.019 showing the breadth of development let's just 00:12:10.019 --> 00:12:11.860 take a quick look around the landscape it sketches 00:12:11.860 --> 00:12:13.980 out maybe touch on a few highlights yeah it's 00:12:13.980 --> 00:12:15.919 good to see these specific examples to ground 00:12:15.919 --> 00:12:18.580 the broader trends we discuss gives it context 00:12:19.129 --> 00:12:21.470 Totally. Like Amazon's apparently building this 00:12:21.470 --> 00:12:26.750 big AI brain for its warehouse robots called 00:12:26.750 --> 00:12:29.350 Proteus, I think. It sounds like it'll let the 00:12:29.350 --> 00:12:31.929 robots follow plain English orders from humans, 00:12:32.090 --> 00:12:34.409 which is kind of wild to think about. Plus, new 00:12:34.409 --> 00:12:36.509 models like Wellspring for optimizing delivery 00:12:36.509 --> 00:12:39.330 routes and an upgraded Sculpt model for stocking 00:12:39.330 --> 00:12:41.830 shelves smarter. Yeah, that's AI moving directly 00:12:41.830 --> 00:12:44.129 into complex physical operations and logistics, 00:12:44.470 --> 00:12:47.789 aiming for massive efficiency gains. It's not 00:12:47.789 --> 00:12:49.919 just soft. anymore. It's interacting with the 00:12:49.919 --> 00:12:53.419 physical world much more. And AlphaFold3. Isomorphic 00:12:53.419 --> 00:12:55.539 Labs, which is part of DeepMind, you know, says 00:12:55.539 --> 00:12:57.940 it can now predict the structure of protein interactions 00:12:57.940 --> 00:13:00.159 with other molecules, not just single proteins. 00:13:00.360 --> 00:13:02.659 This could potentially unlock way faster drug 00:13:02.659 --> 00:13:04.580 design, even for diseases that were previously 00:13:04.580 --> 00:13:07.200 considered Undruggable. That's pretty huge for 00:13:07.200 --> 00:13:09.419 medicine and biotech. Oh, absolutely. Predicting 00:13:09.419 --> 00:13:11.600 molecular interactions beyond just protein folding 00:13:11.600 --> 00:13:14.379 has been a major barrier for a long time. AlphaFold's 00:13:14.379 --> 00:13:17.259 progress here is genuinely a significant potential 00:13:17.259 --> 00:13:20.120 accelerator for pharmaceutical discovery if it 00:13:20.120 --> 00:13:23.840 holds up. Big if, but huge potential. And OpenAI 00:13:23.840 --> 00:13:25.799 still growing like crazy, apparently hitting 00:13:25.799 --> 00:13:28.639 3 million. paying business users now. That's 00:13:28.639 --> 00:13:31.399 a lot. And they're adding features aimed at making 00:13:31.399 --> 00:13:34.500 ChatGPT more useful in daily work, like connectors, 00:13:34.639 --> 00:13:36.399 which sounds like it integrates with things like 00:13:36.399 --> 00:13:39.860 Google Drive, maybe, and this record mode in 00:13:39.860 --> 00:13:41.820 the app for summarizing meetings you record. 00:13:42.480 --> 00:13:46.399 Those features really signal a move towards embedding 00:13:46.399 --> 00:13:50.600 AI deeper into core business workflows and knowledge 00:13:50.600 --> 00:13:52.799 management. They want it to be an indispensable 00:13:52.799 --> 00:13:55.200 tool for daily productivity, not just something 00:13:55.200 --> 00:13:57.299 you chat with occasionally. There's even a policy 00:13:57.299 --> 00:14:00.899 angle mentioned. Anthropic CEO Dario Amodei apparently 00:14:00.899 --> 00:14:04.000 called for a national AI transparency law, not 00:14:04.000 --> 00:14:06.080 freeze on development, importantly, but saying 00:14:06.080 --> 00:14:08.080 we need transparency about these models. That 00:14:08.080 --> 00:14:10.259 raises an important ongoing discussion about 00:14:10.259 --> 00:14:12.659 how societies and governments will oversee and 00:14:12.659 --> 00:14:15.700 regulate powerful AI models as they become more 00:14:15.700 --> 00:14:17.500 integrated into our infrastructure and decision 00:14:17.500 --> 00:14:20.480 making. Transparency is often seen as a key part 00:14:20.480 --> 00:14:23.279 of responsible development, building trust. And 00:14:23.279 --> 00:14:26.700 a cool safety application. Volvo's upcoming EX60 00:14:26.700 --> 00:14:30.320 electric vehicle will have this AI -driven multi 00:14:30.320 --> 00:14:33.460 -adaptive seatbelt system. It uses sensors and 00:14:33.460 --> 00:14:36.639 AI to adjust based on something like 11 different 00:14:36.639 --> 00:14:38.820 crash profiles in real time to try and protect 00:14:38.820 --> 00:14:42.080 passengers better. Using AI for real -time predictive 00:14:42.080 --> 00:14:44.320 safety adjustments in a physical product like 00:14:44.320 --> 00:14:47.070 a car seatbelt. That's a pretty concrete and 00:14:47.070 --> 00:14:49.809 potentially impactful application. Very tangible 00:14:49.809 --> 00:14:52.460 benefit. And on the business side, this new company, 00:14:52.620 --> 00:14:55.240 Shield Technology Partners, just got $100 million 00:14:55.240 --> 00:14:58.519 in funding to launch an AI -enabled managed IT 00:14:58.519 --> 00:15:02.059 services platform. Their strategy is to use shared 00:15:02.059 --> 00:15:04.600 AI agents to automate a lot of the routine IT 00:15:04.600 --> 00:15:07.779 support tasks that bog down human teams. Leveraging 00:15:07.779 --> 00:15:10.100 AI for service delivery and automation, especially 00:15:10.100 --> 00:15:12.539 in areas like IT support, where tasks are often 00:15:12.539 --> 00:15:14.899 repetitive but require specific knowledge, makes 00:15:14.899 --> 00:15:17.000 a lot of sense for scaling expertise and improving 00:15:17.000 --> 00:15:18.759 efficiency. You'll probably see a lot more of 00:15:18.759 --> 00:15:21.230 that. And just too quick. Cool tools that got 00:15:21.230 --> 00:15:23.789 to mention. Eleven Labs dropped version three 00:15:23.789 --> 00:15:26.730 of their platform for even more expressive text 00:15:26.730 --> 00:15:29.269 to speech. That includes emotion tags, which 00:15:29.269 --> 00:15:32.269 is pretty wild for TTS realism, and chat for 00:15:32.269 --> 00:15:34.429 data, which apparently lets you scrape web pages 00:15:34.429 --> 00:15:36.669 just using plain language prompts, which sounds 00:15:36.669 --> 00:15:39.669 way easier than coding it yourself. Yeah, tools 00:15:39.669 --> 00:15:41.730 like those continue to lower the barrier to entry 00:15:41.730 --> 00:15:45.649 for using AI for specific complex tasks, making 00:15:45.649 --> 00:15:48.370 things like creating expressive audio or extracting 00:15:48.370 --> 00:15:50.620 web data accessible. to a much wider audience, 00:15:50.779 --> 00:15:53.740 democratizing the tank in a way. So, wow, okay, 00:15:53.799 --> 00:15:56.000 putting it all together, a lot happening. Okay, 00:15:56.080 --> 00:15:57.899 so let's maybe try and unpack the bigger picture 00:15:57.899 --> 00:16:00.179 from this deep dive. The main takeaway is pretty 00:16:00.179 --> 00:16:02.940 stark, right? AI has genuinely moved way beyond 00:16:02.940 --> 00:16:04.820 just answering your search queries or writing 00:16:04.820 --> 00:16:06.879 simple emails. It's now actively stepping into 00:16:06.879 --> 00:16:09.500 the creative space, generating ideas, even complex 00:16:09.500 --> 00:16:11.779 ones like potential research proposals that, 00:16:11.879 --> 00:16:14.960 based on this study, can stack up against novel 00:16:14.960 --> 00:16:17.830 human -written papers. Yeah, it's not just summarizing 00:16:17.830 --> 00:16:20.049 the past. It's helping sketch out the future 00:16:20.049 --> 00:16:22.950 in a way. And, you know, just like investors 00:16:22.950 --> 00:16:25.889 vet startups or like this AI idea bench study 00:16:25.889 --> 00:16:28.450 scored these agents, we really need robust ways, 00:16:28.549 --> 00:16:30.210 maybe that two -step filter we talked about, 00:16:30.309 --> 00:16:33.230 to evaluate the quality of these AI -generated 00:16:33.230 --> 00:16:35.590 ideas ourselves. Looking at things like genuine 00:16:35.590 --> 00:16:38.169 novelty and practical feasibility is absolutely 00:16:38.169 --> 00:16:40.289 crucial to cut through the noise. Absolutely. 00:16:40.450 --> 00:16:44.230 An idea needs to not just sound good or be statistically 00:16:44.230 --> 00:16:47.799 novel. It needs to have legs. needs to be potentially 00:16:47.799 --> 00:16:50.360 achievable. And that means applying those filters 00:16:50.360 --> 00:16:53.330 is on target for the problem. And can it actually 00:16:53.330 --> 00:16:55.549 be built or implemented with reasonable resources? 00:16:55.909 --> 00:16:57.970 So here's something, I guess, to leave you thinking 00:16:57.970 --> 00:16:59.750 about, something to chew on after hearing all 00:16:59.750 --> 00:17:02.769 this. Given that AI can now generate ideas that 00:17:02.769 --> 00:17:05.130 actually align with and stack up against human 00:17:05.130 --> 00:17:07.150 written research papers published in top venues, 00:17:07.410 --> 00:17:09.750 how should we even begin to rethink the whole 00:17:09.750 --> 00:17:12.109 process of creative ideation itself? Does it 00:17:12.109 --> 00:17:15.009 fundamentally change things? This raises an important 00:17:15.009 --> 00:17:17.210 question for all of us, doesn't it? In a world 00:17:17.210 --> 00:17:19.690 where AI is becoming a constant potential co 00:17:19.690 --> 00:17:22.009 -creator, how do we distinguish between... ideas 00:17:22.009 --> 00:17:25.109 that are merely novel, maybe novel just based 00:17:25.109 --> 00:17:27.190 on recombining its training data in a clever 00:17:27.190 --> 00:17:30.309 way versus those that are truly impactful, truly 00:17:30.309 --> 00:17:32.490 achievable and meaningful in the real world? 00:17:33.009 --> 00:17:35.670 Where's the real insight versus just clever pattern 00:17:35.670 --> 00:17:38.809 mashing? Right. Like what kind of human oversight, 00:17:38.990 --> 00:17:41.509 what kind of human evaluation and curation becomes 00:17:41.509 --> 00:17:43.549 the most crucial part of the process when AI 00:17:43.549 --> 00:17:46.569 is doing so much of the initial generation and 00:17:46.569 --> 00:17:49.190 heavy lifting on concepts? What's our role now? 00:17:49.430 --> 00:17:52.230 It fundamentally shifts our role. perhaps, from 00:17:52.230 --> 00:17:54.970 being the sole generators of ideas to becoming 00:17:54.970 --> 00:17:57.269 expert curators, evaluators, maybe strategic 00:17:57.269 --> 00:18:00.150 prompters and refiners of AI generated concepts. 00:18:00.809 --> 00:18:03.690 Our value might move up the chain, so to speak. 00:18:04.210 --> 00:18:05.789 Definitely something to think about the next 00:18:05.789 --> 00:18:07.529 time you sit down to brainstorm, whether it's 00:18:07.529 --> 00:18:09.809 with AI helping out or just you and a whiteboard.