WEBVTT 00:00:00.000 --> 00:00:03.020 Okay, let's unpack this. Imagine something that 00:00:03.020 --> 00:00:06.559 scores like way off the charts, genius level, 00:00:06.719 --> 00:00:11.220 on an actual Mensa IQ test. It's brilliant with 00:00:11.220 --> 00:00:15.580 text, with logic, you know, but then you ask 00:00:15.580 --> 00:00:17.879 it to look at something, reason about what it 00:00:17.879 --> 00:00:20.539 sees, and suddenly it's, well, it's struggling, 00:00:20.760 --> 00:00:24.260 scoring way below average. Yeah. Kind of weird, 00:00:24.420 --> 00:00:26.679 right? Like, how does that even work? It's a 00:00:26.679 --> 00:00:29.280 really fascinating paradox, actually. And we've 00:00:29.280 --> 00:00:32.100 got this stack of sources today, articles, some 00:00:32.100 --> 00:00:34.539 research papers covering quite a range of things 00:00:34.539 --> 00:00:36.479 happening with AI right now. Right. From those 00:00:36.479 --> 00:00:39.000 cognitive quirks you mentioned to, you know, 00:00:39.020 --> 00:00:40.880 really practical tools people are building and 00:00:40.880 --> 00:00:43.679 even some pretty serious risks we need to talk 00:00:43.679 --> 00:00:45.939 about. Totally. So our mission for this deep 00:00:45.939 --> 00:00:47.619 dive is really just to sift through all this 00:00:47.619 --> 00:00:49.340 and pull out the important stuff, the nuggets 00:00:49.340 --> 00:00:51.399 that help you get a clearer picture of what's 00:00:51.399 --> 00:00:53.719 actually going on with AI. know, beyond the hype. 00:00:53.840 --> 00:00:55.560 Get you informed, maybe raise a few eyebrows. 00:00:55.780 --> 00:00:57.740 Exactly. OK, let's just dive right into that 00:00:57.740 --> 00:01:00.380 first big thing the sources point out. this ai 00:01:00.380 --> 00:01:03.179 and iq score situation we're seeing reports that 00:01:03.179 --> 00:01:06.500 some ai models the text only ones uh -huh just 00:01:06.500 --> 00:01:08.959 language no pictures exactly they're nailing 00:01:08.959 --> 00:01:12.500 these standard iq tests like open ai's oh three 00:01:12.500 --> 00:01:14.920 right oh three that's a big yeah it scored a 00:01:14.920 --> 00:01:18.840 like a staggering 135 on a mensa test i mean 00:01:18.840 --> 00:01:21.120 that's firmly in the genius category way above 00:01:21.120 --> 00:01:24.359 the average human which is like 90 to 110. And 00:01:24.359 --> 00:01:26.140 what's really striking looking at the sources 00:01:26.140 --> 00:01:28.640 is it's not just O3. You've got Clawed Force 00:01:28.640 --> 00:01:31.920 on it hitting 127, Gemini 2 .0, Flash thinking 00:01:31.920 --> 00:01:36.200 right there at 126, Gemini 2 .5 Pro at 124. Wow. 00:01:36.560 --> 00:01:39.319 These are consistently high scores. And the really 00:01:39.319 --> 00:01:41.920 interesting bit, the sources say the top 10 performers 00:01:41.920 --> 00:01:45.159 on these tests, all text -only models, every 00:01:45.159 --> 00:01:47.650 single one. Right. That feels a bit backwards, 00:01:47.810 --> 00:01:49.989 doesn't it? Because all the buzz is about multimodal 00:01:49.989 --> 00:01:53.209 AI, models that can see and hear and process 00:01:53.209 --> 00:01:55.370 text. Yeah, the ones that seem more human -like 00:01:55.370 --> 00:01:57.329 in their inputs. But here's where it gets really 00:01:57.329 --> 00:01:59.629 interesting and maybe a bit confusing based on 00:01:59.629 --> 00:02:01.969 what we're reading. The models that can see the 00:02:01.969 --> 00:02:04.969 multimodals. Like GBT -40 with vision. Precisely. 00:02:05.209 --> 00:02:07.629 When those models were given tasks needing structured 00:02:07.629 --> 00:02:10.129 reasoning, especially involving visual stuff, 00:02:10.389 --> 00:02:12.189 you know, things you'd expect them to be good 00:02:12.189 --> 00:02:14.810 at, they actually performed worse than average 00:02:14.810 --> 00:02:18.900 humans. Significantly worse sometimes. So this 00:02:18.900 --> 00:02:22.159 raises a big question. What is going on? Why 00:02:22.159 --> 00:02:25.219 is GBT 4 .0 vision scoring, what was it, 63? 00:02:25.659 --> 00:02:29.300 And Grok 3 think vision scoring 60. Yeah, those 00:02:29.300 --> 00:02:31.419 numbers are kind of shocking. I mean, in human 00:02:31.419 --> 00:02:33.479 terms, those scores are down in the borderline 00:02:33.479 --> 00:02:35.419 intellectual disability range. So you have this. 00:02:36.060 --> 00:02:39.280 massive kind of perplexing gap opening up. Genius 00:02:39.280 --> 00:02:42.639 text models here and struggling vision models 00:02:42.639 --> 00:02:44.319 over there, at least on these specific reasoning 00:02:44.319 --> 00:02:47.120 tasks. So what's the takeaway here? The sources 00:02:47.120 --> 00:02:49.620 seem to be pointing out this mismatch. Despite 00:02:49.620 --> 00:02:52.039 all the excitement and marketing around multimodal 00:02:52.039 --> 00:02:53.919 AI, they're just not there yet when it comes 00:02:53.919 --> 00:02:56.539 to complex logic combined with visual understanding. 00:02:56.840 --> 00:02:59.840 Right. It kind of widens that gap between the 00:02:59.840 --> 00:03:02.439 hype and today's reality. Maybe, just maybe, 00:03:02.539 --> 00:03:05.180 if your problem is pure logic or abstract reasoning, 00:03:05.400 --> 00:03:08.099 a simpler text model is actually, you know, smarter 00:03:08.099 --> 00:03:11.020 for now. Even Meta's Llama for Maverick, which 00:03:11.020 --> 00:03:13.819 does handle vision, the source mentioned it scored 00:03:13.819 --> 00:03:18.960 105, which is... OK, above average, but not in 00:03:18.960 --> 00:03:21.699 that elite tier with the text only champs. Yeah. 00:03:21.819 --> 00:03:23.939 Interesting distinction. And we should remember 00:03:23.939 --> 00:03:26.939 these IQ scores, while they tell us something 00:03:26.939 --> 00:03:29.199 about a certain kind of intelligence, they're 00:03:29.199 --> 00:03:31.879 definitely not the whole story. Right. One source 00:03:31.879 --> 00:03:34.379 highlights this Apple study looking at large 00:03:34.379 --> 00:03:36.780 reasoning models. Yeah. Yeah. The big names again, 00:03:36.939 --> 00:03:40.840 O3, Claude 3 .7 Sonnet, DeepSeek R1, Gemini, 00:03:40.860 --> 00:03:44.439 and found something kind of critical. They can 00:03:44.439 --> 00:03:46.710 actually just collapse. Completely. When you 00:03:46.710 --> 00:03:49.949 throw really complex problems at them. Wow. Collapse. 00:03:49.969 --> 00:03:52.289 You mean just stop working. Not even a bad answer. 00:03:52.330 --> 00:03:54.750 Just nothing. Pretty much. That's pretty eye 00:03:54.750 --> 00:03:56.530 opening. So it's not just about the peak score 00:03:56.530 --> 00:03:58.930 they can hit. It's about like robustness. Yeah. 00:03:58.949 --> 00:04:01.090 How they handle the hard stuff. Exactly. Brittleness 00:04:01.090 --> 00:04:03.389 is the word that comes to mind. And another source 00:04:03.389 --> 00:04:06.370 they ran a bunch of coding tests on, like 14 00:04:06.370 --> 00:04:09.629 major LLMs. Oh, yeah. How'd that go? Well, they 00:04:09.629 --> 00:04:13.030 found five clear winners. Models that consistently 00:04:13.030 --> 00:04:16.040 produced good code. But they also identified 00:04:16.040 --> 00:04:19.980 models you'd probably want to avoid if you're 00:04:19.980 --> 00:04:22.319 relying on them for coding. And the interesting 00:04:22.319 --> 00:04:25.879 point was that just having pro in the name doesn't 00:04:25.879 --> 00:04:28.720 guarantee better code. Performance varies wildly 00:04:28.720 --> 00:04:30.980 depending on the specific task you give it. Right. 00:04:31.079 --> 00:04:34.319 Task specific performance. Makes sense. And speaking 00:04:34.319 --> 00:04:36.920 of specific tasks, there was that mention of 00:04:36.920 --> 00:04:40.759 a YC backed startup. They built a research agent. 00:04:40.839 --> 00:04:42.579 Yeah, a frontier research agent. And it scored, 00:04:42.660 --> 00:04:46.540 what, 94 .9 % on OpenAI's SimpleQA benchmark? 00:04:47.199 --> 00:04:49.160 Which is all about answering questions based 00:04:49.160 --> 00:04:52.000 on provided text. High score. So it seems like 00:04:52.000 --> 00:04:54.720 when you focus an AI on one specific thing, it 00:04:54.720 --> 00:04:56.759 can get incredibly good at that thing, even if 00:04:56.759 --> 00:04:58.600 the general purpose models are still figuring 00:04:58.600 --> 00:05:00.879 things out. Yeah, specialized versus generalized 00:05:00.879 --> 00:05:03.180 intelligence. That's a key theme, I think. Okay, 00:05:03.240 --> 00:05:05.779 let's shift gears maybe. From performance and 00:05:05.779 --> 00:05:09.160 limits to more practical stuff. What's cool is 00:05:09.160 --> 00:05:11.579 seeing AI features actually starting to show 00:05:11.579 --> 00:05:13.939 up in everyday tools. Or just entirely new tools 00:05:13.939 --> 00:05:17.259 based on AI. Right, like... Google Gemini is 00:05:17.259 --> 00:05:19.839 apparently testing temporary chats. That lets 00:05:19.839 --> 00:05:22.040 you talk to it without the data being used for 00:05:22.040 --> 00:05:25.319 training. Kind of like ChatGPT's incognito mode. 00:05:25.519 --> 00:05:28.339 Yeah, that feels like a response to privacy concerns, 00:05:28.540 --> 00:05:32.000 giving users more control, which is good. And 00:05:32.000 --> 00:05:34.480 Gemini is also adding recurring tasks, you know, 00:05:34.480 --> 00:05:37.079 scheduling things. Oh, like ChatGPT already has. 00:05:37.240 --> 00:05:39.220 Exactly. Catching up on features. They might 00:05:39.220 --> 00:05:42.579 seem small, but for everyday use, setting reminders, 00:05:42.899 --> 00:05:45.129 automating little things. That's actually pretty 00:05:45.129 --> 00:05:47.189 useful. Totally. And the source has listed a 00:05:47.189 --> 00:05:49.269 whole bunch of specific tools, too. Did any jump 00:05:49.269 --> 00:05:51.250 out at you? Well, Glimpse turning photos into 00:05:51.250 --> 00:05:52.970 videos in the browser sounds kind of neat. Yeah. 00:05:53.009 --> 00:05:55.949 And Kling AI 2 .1 promising faster, cheaper, 00:05:56.089 --> 00:05:58.589 better video rendering. That could be a big deal 00:05:58.589 --> 00:06:01.110 for creators. Yeah, definitely. And Moonlit for 00:06:01.110 --> 00:06:04.569 building content workflows. Fusebase AI agents 00:06:04.569 --> 00:06:07.910 for teamwork, like a smarter notion maybe. And 00:06:07.910 --> 00:06:10.689 Agora, an AI search engine just for e -commerce. 00:06:10.829 --> 00:06:12.730 Lots of niche applications popping up. Put those 00:06:12.730 --> 00:06:14.689 quick hits. Apple doing live translation messages 00:06:14.689 --> 00:06:16.449 in FaceTime. That's pretty cool right on your 00:06:16.449 --> 00:06:18.649 phone. Microsoft putting out a free AI video 00:06:18.649 --> 00:06:22.170 creator. Oh, the wildly easy to use one. That's 00:06:22.170 --> 00:06:24.790 the one. And then on the flip side, Antropic. 00:06:25.259 --> 00:06:28.160 quietly killing its Claude Explains blog after 00:06:28.160 --> 00:06:31.100 just a month. Oh, really? I miss that. Yeah, 00:06:31.180 --> 00:06:32.920 it just shows how fast things change. Not every 00:06:32.920 --> 00:06:35.620 idea sticks. Yeah. Even for the big players. 00:06:35.759 --> 00:06:38.680 True. The pace is just relentless. And there 00:06:38.680 --> 00:06:41.500 were mentions of resources, too, like tutorials 00:06:41.500 --> 00:06:44.300 for prompt engineering, guides for building your 00:06:44.300 --> 00:06:46.819 own AI research assistant. Yeah, it gives you 00:06:46.819 --> 00:06:48.420 a feel for what people are actively trying to 00:06:48.420 --> 00:06:50.319 do with this tech beyond just chatting with it. 00:06:50.360 --> 00:06:52.680 Right, building things. But, you know, it's not 00:06:52.680 --> 00:06:55.639 all just... Cool tools and high scores. The sources 00:06:55.639 --> 00:07:00.720 also included a really tough story. Okay. A 16 00:07:00.720 --> 00:07:02.920 -year -old boy who tragically died by suicide. 00:07:03.439 --> 00:07:06.399 Apparently, criminals used fake AI -generated 00:07:06.399 --> 00:07:10.519 nude photos to blackmail him for $3 ,000. Oh, 00:07:10.519 --> 00:07:14.360 man. That's horrific. AI -generated fakes. Yeah. 00:07:14.620 --> 00:07:16.620 And the FBI is warning these kinds of scams are 00:07:16.620 --> 00:07:18.939 targeting more teens. It's just a really stark, 00:07:19.120 --> 00:07:21.360 heartbreaking reminder of the potential for misuse. 00:07:21.500 --> 00:07:23.579 These tools can cause devastating real world 00:07:23.579 --> 00:07:26.819 harm. Oh, God, that's just awful. It really just 00:07:26.819 --> 00:07:30.160 it slams home the need for more awareness, right? 00:07:30.240 --> 00:07:32.980 More protection, especially for kids. Education 00:07:32.980 --> 00:07:35.199 about this stuff and frankly, consequences for 00:07:35.199 --> 00:07:37.160 the people doing it. It's not abstract anymore. 00:07:37.339 --> 00:07:39.899 Absolutely. The dark side is incredibly real 00:07:39.899 --> 00:07:44.939 and dangerous. Shifting gears completely. To 00:07:44.939 --> 00:07:47.480 the business world. OK. The source has also touched 00:07:47.480 --> 00:07:50.139 on the money side. This company, AnySphere, they 00:07:50.139 --> 00:07:54.040 just raised $900 million. Wow. $900 million. 00:07:54.319 --> 00:07:56.759 Yeah. Huge funding round. Right. Puts their valuation 00:07:56.759 --> 00:08:00.540 at nearly $10 billion. And apparently their AI 00:08:00.540 --> 00:08:02.920 tool is already bringing in $200 million a year. 00:08:03.040 --> 00:08:05.480 That's serious cash. Shows the level of belief 00:08:05.480 --> 00:08:08.160 and maybe progress in certain corners of the 00:08:08.160 --> 00:08:10.240 AI business world. Definitely. And I thought 00:08:10.240 --> 00:08:12.670 it was interesting. The Zapier. CEO share their 00:08:12.670 --> 00:08:15.129 internal chart for measuring AI fluency. Oh, 00:08:15.189 --> 00:08:18.310 yeah. What was that like? To scale from unacceptable 00:08:18.310 --> 00:08:21.550 use of AI all the way up to transformative. Kind 00:08:21.550 --> 00:08:23.550 of makes you think, doesn't it? Where do you 00:08:23.550 --> 00:08:25.529 or your company fall on that spectrum right now? 00:08:25.670 --> 00:08:27.610 Yeah, that's a good self -assessment tool. How 00:08:27.610 --> 00:08:29.509 well are you actually using this stuff? Right. 00:08:29.649 --> 00:08:32.250 Practical perspective. OK, let's wrap up with 00:08:32.250 --> 00:08:34.529 something really forward looking, kind of mind 00:08:34.529 --> 00:08:37.690 bending, actually. One source dives into this 00:08:37.690 --> 00:08:42.639 new biomolecular AI. from MIT and a company called 00:08:42.639 --> 00:08:45.299 Recursion. It's called Bolts2. Yeah, this was 00:08:45.299 --> 00:08:47.700 pretty wild. What's really fascinating is how 00:08:47.700 --> 00:08:50.080 it's aiming to speed up drug research dramatically. 00:08:50.960 --> 00:08:53.840 Okay, so Bolts2 predicts something called binding 00:08:53.840 --> 00:08:57.179 affinity, which is basically how strongly a potential 00:08:57.179 --> 00:08:59.299 drug molecule will stick to its target in the 00:08:59.299 --> 00:09:01.720 body, like a protein involved in a disease. Right, 00:09:01.779 --> 00:09:04.000 that's crucial for a drug to work. Exactly. Getting 00:09:04.000 --> 00:09:06.360 that prediction right is key. Bolts2 does it 00:09:06.360 --> 00:09:10.059 with, they say, physics -grade accuracy. But 00:09:10.059 --> 00:09:13.320 here's the kicker. It does it a thousand times 00:09:13.320 --> 00:09:15.940 faster than the old school computer simulations. 00:09:16.299 --> 00:09:18.399 Wait, say that again? A thousand times faster? 00:09:18.419 --> 00:09:21.419 A thousand times, yeah. Wow. That's not incremental. 00:09:21.519 --> 00:09:23.720 That's transformative speed. What exactly does 00:09:23.720 --> 00:09:26.059 it predict? Just that sticking power? No, it's 00:09:26.059 --> 00:09:28.120 more comprehensive. It's a foundation model, 00:09:28.299 --> 00:09:31.740 kind of like an LLM, but for biology. It predicts 00:09:31.740 --> 00:09:34.139 both the 3D shapes of molecules and how they 00:09:34.139 --> 00:09:36.600 bind together. It builds on their earlier model, 00:09:36.720 --> 00:09:39.470 Bolts 1. which was already seen as an open source 00:09:39.470 --> 00:09:42.029 competitor to AlphaFold 3 for structure prediction. 00:09:42.289 --> 00:09:44.950 OK, AlphaFold was huge for protein shapes. Right. 00:09:45.070 --> 00:09:48.690 But Boltz 2 is apparently the first AI to model 00:09:48.690 --> 00:09:51.049 both the structure and the binding affinity together 00:09:51.049 --> 00:09:54.590 jointly in one go. And that combined approach 00:09:54.590 --> 00:09:57.649 seems to be the key. So it's accurate, too, not 00:09:57.649 --> 00:10:00.129 just fast. That's what the source claims. They 00:10:00.129 --> 00:10:02.389 say it matches the accuracy of really complex, 00:10:02.490 --> 00:10:05.669 slow physics -based methods and beats other methods 00:10:05.669 --> 00:10:09.269 on standard tests like OpenFE and CASP16. Those 00:10:09.269 --> 00:10:11.289 are like the Olympics for these kinds of predictions. 00:10:11.470 --> 00:10:13.570 And importantly, they apparently tested some 00:10:13.570 --> 00:10:15.610 of its predictions in the real world, prospectively, 00:10:15.710 --> 00:10:18.210 and confirmed they were strong binders. So it's 00:10:18.210 --> 00:10:21.289 validated. OK, so it's fast, accurate, validated, 00:10:21.350 --> 00:10:24.820 and built for practical use. Optimized for GPUs 00:10:24.820 --> 00:10:27.100 and stuff. Yeah, exactly. Designed for large 00:10:27.100 --> 00:10:29.440 -scale deployment. And their goal, according 00:10:29.440 --> 00:10:32.100 to the source, is pretty ambitious. They want 00:10:32.100 --> 00:10:35.519 Bolt 2 to be the go -to open platform for structure 00:10:35.519 --> 00:10:37.840 and affinity. Kind of like what AlphaFold became 00:10:37.840 --> 00:10:40.779 just for structure. Wow. If they pull that off, 00:10:40.919 --> 00:10:43.139 it could genuinely change the pace of discovering 00:10:43.139 --> 00:10:45.919 new medicines. Totally reshape it. Okay, that 00:10:45.919 --> 00:10:49.019 was quite a journey. We went from AI being a 00:10:49.019 --> 00:10:51.759 text genius, but maybe visually challenged. Uh 00:10:51.759 --> 00:10:54.620 -huh, the paradox. To spotting limitations, checking 00:10:54.620 --> 00:10:57.279 out all those new tools, facing the really serious 00:10:57.279 --> 00:11:00.299 risks. Yeah, the blackmail story was heavy. Seeing 00:11:00.299 --> 00:11:03.179 the huge money involved and then blasting off 00:11:03.179 --> 00:11:05.659 into predicting molecules a thousand times faster 00:11:05.659 --> 00:11:08.799 for drug discovery. It's a lot. It really shows 00:11:08.799 --> 00:11:11.120 the sheer breadth of AI today, doesn't it? From 00:11:11.120 --> 00:11:14.000 grappling with logic problems to simulating physics 00:11:14.000 --> 00:11:16.500 at incredible speeds. It really makes you wonder 00:11:16.500 --> 00:11:18.620 about the future. Like, how do we bridge that 00:11:18.620 --> 00:11:20.980 gap between tech smarts and visual reasoning 00:11:20.980 --> 00:11:23.820 and the big one? How do we make sure these incredibly 00:11:23.820 --> 00:11:27.120 powerful tools are used for good, like finding 00:11:27.120 --> 00:11:29.580 cures and not for horrific things like those 00:11:29.580 --> 00:11:32.059 scams? And maybe this raises an even deeper question 00:11:32.059 --> 00:11:35.389 for you to think about. Given how AI is soaring 00:11:35.389 --> 00:11:37.750 in these super specialized complex areas like 00:11:37.750 --> 00:11:40.429 drug research, while still fumbling with things 00:11:40.429 --> 00:11:43.509 humans find easy, like some visual tasks, what 00:11:43.509 --> 00:11:45.570 does that split tell us about intelligence itself? 00:11:46.009 --> 00:11:47.990 You know, human intelligence versus artificial 00:11:47.990 --> 00:11:50.230 intelligence. What even is it? Definitely something 00:11:50.230 --> 00:11:50.649 that you want.