WEBVTT 00:00:00.000 --> 00:00:03.580 Okay, so welcome back to the deep dive. Today's 00:00:03.580 --> 00:00:06.799 dive is, uh, well. we're really digging into 00:00:06.799 --> 00:00:09.160 some fascinating stuff from the stack of material 00:00:09.160 --> 00:00:11.140 we've been looking at articles some fresh research 00:00:11.140 --> 00:00:13.759 notes you know everything centered around what's 00:00:13.759 --> 00:00:17.260 bubbling up right now in ai our mission today 00:00:17.260 --> 00:00:19.559 for you listening is to just kind of cut through 00:00:19.559 --> 00:00:22.260 the noise right get to the really good insights 00:00:22.260 --> 00:00:25.379 the key developments maybe uh some surprising 00:00:25.379 --> 00:00:27.460 things you might have missed we're gonna unpack 00:00:27.460 --> 00:00:30.160 all this material together yeah absolutely it's 00:00:30.160 --> 00:00:34.170 uh it's easy to feel a bit swamped with how fast 00:00:34.170 --> 00:00:36.909 AI is moving. So this dive is about helping you 00:00:36.909 --> 00:00:39.189 connect the dots, understand the significance 00:00:39.189 --> 00:00:40.929 of these latest developments. You know, it's 00:00:40.929 --> 00:00:43.229 like we're looking at both the really practical, 00:00:43.369 --> 00:00:45.530 immediate applications and some of these big, 00:00:45.630 --> 00:00:47.810 almost foundational shifts happening underneath 00:00:47.810 --> 00:00:52.030 the surface. Exactly. So let's jump in. And I 00:00:52.030 --> 00:00:53.869 want to start with something that when we were 00:00:53.869 --> 00:00:56.509 going through the sources felt incredibly impactful, 00:00:56.570 --> 00:00:58.920 maybe a little. Mind -blowing, honestly. Okay. 00:00:59.000 --> 00:01:01.240 It's this project called Fraggle coming out of 00:01:01.240 --> 00:01:04.260 Singapore. Fraggle, right. The core idea here 00:01:04.260 --> 00:01:08.099 is detecting incredibly tiny traces of cancer 00:01:08.099 --> 00:01:11.780 using just a small blood sample. Like, picture 00:01:11.780 --> 00:01:14.700 this using AI to potentially spot cancer before 00:01:14.700 --> 00:01:16.719 you even have symptoms. I mean, that feels like 00:01:16.719 --> 00:01:18.980 sci -fi, right? It really does. But the material 00:01:18.980 --> 00:01:21.780 suggests it's getting real. What's really striking 00:01:21.780 --> 00:01:24.599 about Fraggle, based on what we read, is how 00:01:24.599 --> 00:01:26.829 it does it. It's looking for something called 00:01:26.829 --> 00:01:30.030 circulating tumor DNA or ctDNA. ctDNA, yeah. 00:01:30.209 --> 00:01:32.129 These are microscopic bits of DNA that break 00:01:32.129 --> 00:01:34.450 off from cancer cells and float around in your 00:01:34.450 --> 00:01:37.409 bloodstream. Okay. Now, the key detail the research 00:01:37.409 --> 00:01:40.049 highlighted is that these ctDNA fragments tend 00:01:40.049 --> 00:01:42.670 to be slightly different sizes than the healthy 00:01:42.670 --> 00:01:44.829 DNA fragments also floating around. Different 00:01:44.829 --> 00:01:47.930 sizes. Yeah. And Fraggle's AI is trained to spot 00:01:47.930 --> 00:01:50.010 that subtle size difference. It's using this. 00:01:50.700 --> 00:01:53.859 This quiet biological signature as its marker. 00:01:54.000 --> 00:01:55.939 Oh, OK. So it's not trying to find a whole tumor. 00:01:56.040 --> 00:01:58.319 It's looking for these molecular crumbs like 00:01:58.319 --> 00:02:00.719 tiny clues. Exactly. These tiny clues in the 00:02:00.719 --> 00:02:02.959 blood. Yeah. And the practical uses they outline 00:02:02.959 --> 00:02:04.739 in the sources, they're pretty significant, right? 00:02:04.819 --> 00:02:06.599 Like it could be that early warning system you 00:02:06.599 --> 00:02:09.560 mentioned or. Or catching a cancer relapse way 00:02:09.560 --> 00:02:12.240 sooner. Right. Or even knowing quickly if a specific 00:02:12.240 --> 00:02:14.530 treatment isn't working. That kind of thing. 00:02:14.650 --> 00:02:17.050 Right. And the sources emphasize this isn't just 00:02:17.050 --> 00:02:19.930 theoretical anymore. They've trained the AI on 00:02:19.930 --> 00:02:22.909 data from whole genome sequencing, looking at 00:02:22.909 --> 00:02:26.090 DNA sizes from actual cancer patients versus 00:02:26.090 --> 00:02:28.729 healthy individuals. And it's already entered 00:02:28.729 --> 00:02:31.310 clinical trials with over 100 patients in Singapore 00:02:31.310 --> 00:02:33.909 who are currently undergoing cancer treatment. 00:02:34.229 --> 00:02:36.889 So it's, you know, moving through those necessary 00:02:36.889 --> 00:02:39.509 steps towards real world use. OK, but here's 00:02:39.509 --> 00:02:41.789 where my jaw kind of dropped. Digging into the 00:02:41.789 --> 00:02:45.240 cost part. We looked at what a typical CT DNA 00:02:45.240 --> 00:02:48.500 test costs today, and the sources put it anywhere 00:02:48.500 --> 00:02:52.500 up to, what, $780? Yeah, around that. Expensive. 00:02:52.500 --> 00:02:54.580 But Fraggle's method, it's coming in at just 00:02:54.580 --> 00:02:58.560 $39. $39! I mean, that's a wild difference. It 00:02:58.560 --> 00:03:00.900 really is. $39. And the implications of that 00:03:00.900 --> 00:03:03.259 massive cost reduction are profound, according 00:03:03.259 --> 00:03:05.580 to the material. Like what? Well, it suddenly 00:03:05.580 --> 00:03:07.810 makes frequent monitoring feasible. The sources 00:03:07.810 --> 00:03:10.050 even used the phrase like a live feed for tracking. 00:03:10.229 --> 00:03:13.689 Wow, a live feed. And this is huge. It could 00:03:13.689 --> 00:03:16.110 potentially make high quality cancer monitoring 00:03:16.110 --> 00:03:19.169 accessible globally, not just something limited 00:03:19.169 --> 00:03:21.949 to wealthy nations or specialized private hospitals. 00:03:22.050 --> 00:03:24.550 That's huge. And they designed it smartly, too, 00:03:24.629 --> 00:03:26.750 didn't they? The sources pointed out it was built 00:03:26.750 --> 00:03:29.530 with adoption in mind. It integrates seamlessly 00:03:29.530 --> 00:03:32.389 with standard DNA profiling labs already out 00:03:32.389 --> 00:03:34.569 there. You don't need new expensive machines 00:03:34.569 --> 00:03:36.990 or require massive retraining for technicians. 00:03:37.150 --> 00:03:39.789 Right. It's designed to work alongside existing 00:03:39.789 --> 00:03:43.789 tools, which makes rolling it out. Well, much 00:03:43.789 --> 00:03:46.210 more realistic. Okay. Yeah, it's that focus on 00:03:46.210 --> 00:03:48.250 practical integration that turns it from an interesting 00:03:48.250 --> 00:03:51.110 lab project into something with real potential 00:03:51.110 --> 00:03:53.409 to change clinical practice. So what does it 00:03:53.409 --> 00:03:55.550 all mean then? What this all means is it points 00:03:55.550 --> 00:03:59.490 towards a future where maybe, just maybe, cancer 00:03:59.490 --> 00:04:01.949 detection could become as relatively routine 00:04:01.949 --> 00:04:04.430 as, say, getting your cholesterol checked during 00:04:04.430 --> 00:04:06.969 an annual physical. That would fundamentally 00:04:06.969 --> 00:04:09.610 shift the paradigm for how we manage the disease. 00:04:10.139 --> 00:04:13.000 Wow. Okay. So that's a super tangible, impactful 00:04:13.000 --> 00:04:16.120 application, obviously. Now let's pivot from 00:04:16.120 --> 00:04:18.240 something concrete in healthcare to something 00:04:18.240 --> 00:04:22.300 a bit more abstract, maybe, but potentially even 00:04:22.300 --> 00:04:24.899 more foundational for AI itself. We're going 00:04:24.899 --> 00:04:29.199 to talk about MIT's SEAL project. Ah, SEAL. And 00:04:29.199 --> 00:04:33.480 the hook here is, what if AI could learn... By 00:04:33.480 --> 00:04:36.779 literally teaching itself, like self -improvement 00:04:36.779 --> 00:04:39.980 on steroids. This is fascinating stuff from the 00:04:39.980 --> 00:04:42.439 research papers we reviewed. SEAL stands for 00:04:42.439 --> 00:04:45.699 self -adapting LLMs, large language models. Self 00:04:45.699 --> 00:04:47.920 -adapting LLMs, right. It's a framework where 00:04:47.920 --> 00:04:50.699 an LLM doesn't just process data fed to it, it 00:04:50.699 --> 00:04:53.300 trains itself. How does that even work? Like, 00:04:53.339 --> 00:04:55.379 what's the mechanism? Based on the descriptions, 00:04:55.600 --> 00:04:57.899 it involves the AI creating its own synthetic 00:04:57.899 --> 00:05:00.720 data to learn from. Its own data. Yeah. It updates 00:05:00.720 --> 00:05:02.660 its own internal instructions on how to learn 00:05:02.660 --> 00:05:05.120 more effectively. And it even makes its own adjustments, 00:05:05.180 --> 00:05:07.519 like tweaking its weights, the parameters within 00:05:07.519 --> 00:05:09.699 the model that determine how it processes information. 00:05:10.160 --> 00:05:12.100 So it's like... It's essentially generating its 00:05:12.100 --> 00:05:14.399 own self -edits or revision notes, as you could 00:05:14.399 --> 00:05:17.740 think of it. Okay, that's kind of weirdly human 00:05:17.740 --> 00:05:20.079 -like, like writing your own study guide before 00:05:20.079 --> 00:05:23.060 a test. Exactly. And the results cited in the 00:05:23.060 --> 00:05:25.720 source material are pretty wild. They report 00:05:25.720 --> 00:05:28.839 it's already outperforming GPT -4 .1 on some 00:05:28.839 --> 00:05:31.759 specific tasks. Already. Wow. And this is key. 00:05:31.879 --> 00:05:34.259 It learned more effectively from the data it 00:05:34.259 --> 00:05:36.519 generated itself, its own notes, than it did 00:05:36.519 --> 00:05:39.360 from material generated by GPT -4 .1 for it to 00:05:39.360 --> 00:05:42.550 learn from. Whoa. Hold on. So the AI's own teaching 00:05:42.550 --> 00:05:45.670 method, its own notes, are better for it than 00:05:45.670 --> 00:05:48.209 being taught by another top AI model. That's 00:05:48.209 --> 00:05:50.189 what the results suggest. Yeah. And it raises 00:05:50.189 --> 00:05:51.970 this really interesting question. What does it 00:05:51.970 --> 00:05:54.269 mean if an AI system can figure out the best 00:05:54.269 --> 00:05:57.029 way for it to learn, tailoring its process to 00:05:57.029 --> 00:05:59.230 its own internal architecture or style? Like 00:05:59.230 --> 00:06:01.209 finding its own learning style. Exactly. Much 00:06:01.209 --> 00:06:02.850 like how some humans find their own revision 00:06:02.850 --> 00:06:04.889 notes more effective than just rereading a textbook. 00:06:05.089 --> 00:06:07.290 Yeah. Yeah. And they mentioned a dramatic improvement 00:06:07.290 --> 00:06:09.149 in something specific. Right. Like puzzle solving. 00:06:09.230 --> 00:06:11.000 I remember reading that. They did. The source 00:06:11.000 --> 00:06:13.300 highlighted a jump in certain puzzle -solving 00:06:13.300 --> 00:06:16.420 tasks. Using standard methods, the AI had a 0 00:06:16.420 --> 00:06:20.079 % success rate. Zero. Zero percent. Okay. But 00:06:20.079 --> 00:06:22.040 after going through the SEAL self -training process, 00:06:22.420 --> 00:06:28.079 it jumped to 72 .5 % success. 72 .5. From zero. 00:06:28.180 --> 00:06:31.600 From zero. That's a huge, huge leap. A huge leap. 00:06:31.639 --> 00:06:33.779 Totally. And if we connect this to the bigger 00:06:33.779 --> 00:06:37.009 picture. What SEAL and similar frameworks they 00:06:37.009 --> 00:06:38.709 mentioned, like Sakana's work on dynamic graph 00:06:38.709 --> 00:06:43.750 models, DGM. What they're exploring is LLMs that 00:06:43.750 --> 00:06:46.709 can potentially evolve continuously without needing 00:06:46.709 --> 00:06:49.569 to be retrained from scratch by humans every 00:06:49.569 --> 00:06:52.589 time. Continuously evolving. Yeah. This mechanism. 00:06:53.079 --> 00:06:54.860 This idea of self -improvement and continuous 00:06:54.860 --> 00:06:57.800 adaptation. It's central to a lot of the theoretical 00:06:57.800 --> 00:07:00.779 discussions around things like AGI or artificial 00:07:00.779 --> 00:07:03.560 general intelligence. And even speculation about 00:07:03.560 --> 00:07:05.480 super intelligence. It's not just about AI getting 00:07:05.480 --> 00:07:08.199 slightly smarter. It's about changing a fundamental 00:07:08.199 --> 00:07:10.560 mechanism by which it gets smarter, potentially 00:07:10.560 --> 00:07:13.259 in real time. OK, so circling back to you listening, 00:07:13.439 --> 00:07:16.430 what does all this mean? It means the AI landscape 00:07:16.430 --> 00:07:18.470 we're looking at isn't just about incremental 00:07:18.470 --> 00:07:21.949 updates to the tools we use. It's about AI potentially 00:07:21.949 --> 00:07:25.050 changing in really fundamental ways, involving 00:07:25.050 --> 00:07:27.370 its own capabilities and learning processes, 00:07:27.709 --> 00:07:29.949 maybe right before our eyes. It's a totally different 00:07:29.949 --> 00:07:32.069 dynamic. It's a different ballgame, really. Okay. 00:07:32.170 --> 00:07:35.370 So we've looked at a concrete application in 00:07:35.370 --> 00:07:38.569 healthcare, a foundational shift in how AI learns. 00:07:39.029 --> 00:07:41.029 Now let's kind of sweep up some of the other 00:07:41.029 --> 00:07:43.470 really interesting... nuggets from the material 00:07:43.470 --> 00:07:45.870 we reviewed. It's a mix of practical stuff, some 00:07:45.870 --> 00:07:49.509 industry gossip, and definitely some quirky things. 00:07:49.829 --> 00:07:52.269 Sounds good. Starting on the educational side, 00:07:52.470 --> 00:07:55.250 something we noted was Anthropic releasing a 00:07:55.250 --> 00:07:58.750 free 12 -lesson course they call AI Fluency. 00:07:58.949 --> 00:08:00.970 AI Fluency, okay. The source points out it goes 00:08:00.970 --> 00:08:02.990 beyond just prompting tips, which is what a lot 00:08:02.990 --> 00:08:05.329 of courses focus on. This one actually involves 00:08:05.329 --> 00:08:08.430 planning and executing a real AI project yourself. 00:08:08.850 --> 00:08:11.050 Oh, hands -on. Yeah, hands -on, and you get a 00:08:11.050 --> 00:08:13.319 certificate. could be a useful resource for anyone 00:08:13.319 --> 00:08:15.240 wanting to dive deeper themselves. Oh, that sounds 00:08:15.240 --> 00:08:16.620 cool. And then there's the stuff that just makes 00:08:16.620 --> 00:08:19.839 you laugh, like VZero's CAPTCHA contest. Right. 00:08:19.920 --> 00:08:21.819 VZero is running a contest for the most ridiculous 00:08:21.819 --> 00:08:24.360 CAPTCHA. Ridiculous how? The example the source 00:08:24.360 --> 00:08:28.019 gave was literally, are you human or are you 00:08:28.019 --> 00:08:30.660 dancer? Chuckles. Like, is that from a song? 00:08:30.819 --> 00:08:33.820 It is, yeah. The Killers, I think. But as a CAPTCHA, 00:08:33.919 --> 00:08:35.960 it's pretty out there. Totally unhinged. The 00:08:35.960 --> 00:08:39.299 winner gets $1 ,000 in credits, which feels about 00:08:39.299 --> 00:08:41.659 right. for inducing that level of existential 00:08:41.659 --> 00:08:44.539 confusion. Juckles, definitely. On the technical 00:08:44.539 --> 00:08:47.940 side, we saw a mention of a builder combining 00:08:47.940 --> 00:08:50.639 different cutting -edge models, Cloud Code, plus 00:08:50.639 --> 00:08:54.299 OpenAI's O3 model, plus Gemini 2 .5. Okay, stacking 00:08:54.299 --> 00:08:56.340 models. Yeah, all working together through something 00:08:56.340 --> 00:08:59.860 called MCP. Think of it like getting different 00:08:59.860 --> 00:09:02.460 AI superpowers to team up. Like the Avengers. 00:09:02.840 --> 00:09:04.740 The source described it exactly like the Avengers 00:09:04.740 --> 00:09:09.110 of AI models. Okay, MCP. What's MCP stand for? 00:09:09.190 --> 00:09:11.909 Like, is that a specific framework or? You know, 00:09:11.929 --> 00:09:13.769 the source didn't spell out the acronym, unfortunately, 00:09:13.970 --> 00:09:15.809 but the description implied it's a method or 00:09:15.809 --> 00:09:17.870 platform for getting diverse models to collaborate 00:09:17.870 --> 00:09:20.549 on tasks, sharing strengths. Got it. It just 00:09:20.549 --> 00:09:22.529 highlights this trend of people trying to orchestrate 00:09:22.529 --> 00:09:25.570 multiple powerful AIs rather than relying on 00:09:25.570 --> 00:09:29.309 just one. OK, got it. Like a meta layer for AIs. 00:09:30.129 --> 00:09:33.289 OK, now here's where things get a little less 00:09:33.289 --> 00:09:36.149 fun, maybe. Based on some analysis we read. There's 00:09:36.149 --> 00:09:38.750 this idea gaining traction that chat GPT and 00:09:38.750 --> 00:09:41.149 other early generative models might have polluted 00:09:41.149 --> 00:09:44.070 the Internet so badly that it's actually hindering 00:09:44.070 --> 00:09:46.970 future AI development. Yeah, this is a significant 00:09:46.970 --> 00:09:49.330 potential downside highlighted in the material. 00:09:49.549 --> 00:09:51.830 The concept is sometimes referred to as model 00:09:51.830 --> 00:09:54.389 collapse. Model collapse. Right. I've heard that 00:09:54.389 --> 00:09:57.720 term. The theory is that as AI models are increasingly 00:09:57.720 --> 00:10:00.779 trained on data from the Internet and the Internet 00:10:00.779 --> 00:10:02.980 is increasingly filled with text and images generated 00:10:02.980 --> 00:10:06.639 by earlier AI models, we start training new models 00:10:06.639 --> 00:10:09.980 on the output of old models rather than on truly 00:10:09.980 --> 00:10:13.820 human generated data. It's like feeding photocopies 00:10:13.820 --> 00:10:15.659 to a copier over and over and eventually the 00:10:15.659 --> 00:10:18.039 copies just degrade. Exactly. That's a great 00:10:18.039 --> 00:10:20.659 analogy. You lose the richness, the nuance, the 00:10:20.659 --> 00:10:24.059 sheer originality of genuinely. human expression 00:10:24.059 --> 00:10:26.960 and data. The sources noted that this AI generated 00:10:26.960 --> 00:10:30.259 spam has potentially tainted a lot of modern 00:10:30.259 --> 00:10:34.000 Internet data. What's striking is they said explicitly 00:10:34.000 --> 00:10:37.299 that data generated before 2022 is now considered 00:10:37.299 --> 00:10:40.820 gold. Before 2022 is gold. Wow. Yeah, because 00:10:40.820 --> 00:10:43.539 it's much less likely to be this kind of AI generated 00:10:43.539 --> 00:10:46.539 synthetic stuff. It's a consequence that, you 00:10:46.539 --> 00:10:48.759 know, maybe wasn't fully anticipated when these 00:10:48.759 --> 00:10:51.429 models first came out. That is a wild, unintended 00:10:51.429 --> 00:10:54.169 consequence. Like, the Internet just got less 00:10:54.169 --> 00:10:56.350 useful for training the very things it helped 00:10:56.350 --> 00:10:58.730 create. Potentially, yeah. And it raises big 00:10:58.730 --> 00:11:00.529 questions about where future training data will 00:11:00.529 --> 00:11:02.990 come from, right? Yes. Okay, speaking of models 00:11:02.990 --> 00:11:05.289 interacting, we saw some mentions of a little 00:11:05.289 --> 00:11:08.830 AI beef, didn't we? Oh, yes. Claude Fore apparently 00:11:08.830 --> 00:11:11.490 co -authored what one source described as a spicy 00:11:11.490 --> 00:11:14.029 takedown of a recent research paper put out by 00:11:14.029 --> 00:11:16.649 Apple. A takedown? Yeah, saying it, and I quote, 00:11:16.850 --> 00:11:20.320 kind of sucks. Chuckles. Seriously. ai models 00:11:20.320 --> 00:11:22.320 reviewing and criticizing each other's academic 00:11:22.320 --> 00:11:24.500 work now that's a whole new level of meta it 00:11:24.500 --> 00:11:26.820 really is it shows this emerging dynamic within 00:11:26.820 --> 00:11:30.419 the ai research community where The models themselves 00:11:30.419 --> 00:11:32.919 are starting to participate in the academic discourse, 00:11:33.159 --> 00:11:36.259 even if prompted by humans, of course. It's fascinating 00:11:36.259 --> 00:11:38.399 to watch that unfold. Totally fascinating. And 00:11:38.399 --> 00:11:40.879 just briefly touching on the industry and business 00:11:40.879 --> 00:11:44.419 side from the sources, we read that Google apparently 00:11:44.419 --> 00:11:47.360 had to invent this role, a chief AI architect, 00:11:47.679 --> 00:11:50.200 because despite having incredible research models, 00:11:50.440 --> 00:11:53.039 they still seem to struggle turning that cutting 00:11:53.039 --> 00:11:56.159 edge stuff into usable, consumer ready products. 00:11:56.519 --> 00:11:58.840 Yeah, it's just a disconnect between their theoretical. 00:11:58.919 --> 00:12:01.340 medical breakthroughs, and their practical application 00:12:01.340 --> 00:12:03.820 pipeline, which is, you know, a common challenge. 00:12:04.000 --> 00:12:06.500 True. Meanwhile, big money is still flowing into 00:12:06.500 --> 00:12:09.019 the space. One venture studio was noted for raising 00:12:09.019 --> 00:12:12.539 $190 million, specifically targeting AI applications 00:12:12.539 --> 00:12:14.820 in healthcare and finance. Yeah, those remain 00:12:14.820 --> 00:12:17.200 huge areas for AI development and investment. 00:12:17.850 --> 00:12:19.850 Makes sense. And there were even mentions kind 00:12:19.850 --> 00:12:23.269 of whispers of potential cracks in the OpenAI 00:12:23.269 --> 00:12:25.669 Microsoft relationship. Yeah. Saw that, too. 00:12:25.730 --> 00:12:27.950 Always interesting dynamics between the big players. 00:12:28.090 --> 00:12:32.269 While OpenAI also reportedly bagged a $200 million 00:12:32.269 --> 00:12:35.509 deal with the U .S. military for war gaming. 00:12:35.750 --> 00:12:37.830 Right. So that just shows the diverse and sometimes 00:12:37.830 --> 00:12:40.850 conflicting arenas where AI is having an impact 00:12:40.850 --> 00:12:44.129 from defense to health care to corporate strategy. 00:12:44.470 --> 00:12:47.529 It's everywhere. It really is. And just quickly 00:12:47.529 --> 00:12:49.909 to wrap up this segment, the source has listed 00:12:49.909 --> 00:12:51.769 a whole bunch of these new, often specialized 00:12:51.769 --> 00:12:54.029 AI tools coming out, giving you a sense of the 00:12:54.029 --> 00:12:56.070 breadth of what's being built. Right, like C 00:12:56.070 --> 00:12:58.830 -Dance Pro for generating videos from text. Instance 00:12:58.830 --> 00:13:01.190 for turning ideas into apps or games without 00:13:01.190 --> 00:13:03.570 coding. Fluidworks for personalized guidance. 00:13:03.970 --> 00:13:06.330 Wondrish for building no -code webpages easily. 00:13:06.649 --> 00:13:09.009 Scribe for automatically creating how -to guides 00:13:09.009 --> 00:13:11.840 from screen recordings. That sounds useful. Yeah, 00:13:11.879 --> 00:13:14.139 definitely. And even updates to existing tools, 00:13:14.279 --> 00:13:16.860 like ChatGPT Canvas now letting you export to 00:13:16.860 --> 00:13:20.019 PDF, or Google's Video Model VO already having 00:13:20.019 --> 00:13:22.179 a short film premiere at Tribeca Festival. Yeah, 00:13:22.220 --> 00:13:25.019 stuff is just happening everywhere at all levels. 00:13:25.139 --> 00:13:27.080 It's hard to keep up. It really is. And just 00:13:27.080 --> 00:13:29.159 one last quick hit from the material, a little 00:13:29.159 --> 00:13:31.659 reminder that individual effort still matters. 00:13:32.279 --> 00:13:35.320 A story about a woman who won $10K in a machine 00:13:35.320 --> 00:13:38.539 learning competition in just one week. Nice. 00:13:38.759 --> 00:13:42.340 So it's not all huge companies and labs. Individuals 00:13:42.340 --> 00:13:44.639 are still making waves, too. Absolutely. Good 00:13:44.639 --> 00:13:47.419 reminder. Okay, wow. So we've covered a lot of 00:13:47.419 --> 00:13:49.779 ground in this deep dive. We started with that 00:13:49.779 --> 00:13:52.440 incredible, potentially revolutionary application 00:13:52.440 --> 00:13:55.740 of AI in healthcare with Fragle. Yeah, the $39 00:13:55.740 --> 00:13:58.809 test. Right. moved the cutting edge of how AI 00:13:58.809 --> 00:14:02.990 learns and evolves with MIT's SEAL project. The 00:14:02.990 --> 00:14:06.009 self -teaching AI. Exactly. And then just got 00:14:06.009 --> 00:14:09.090 a rapid fire snapshot of the incredibly diverse. 00:14:09.710 --> 00:14:12.070 sometimes weird and rapidly changing landscape 00:14:12.070 --> 00:14:14.570 with all those other points from model collapse 00:14:14.570 --> 00:14:17.809 and AI beef to new tools and industry shifts. 00:14:17.990 --> 00:14:19.690 Hopefully going through these sources together 00:14:19.690 --> 00:14:21.509 helps you see some of the underlying patterns, 00:14:21.710 --> 00:14:24.149 the key developments, and ultimately the potential 00:14:24.149 --> 00:14:26.450 impact on you. It's by getting that knowledge 00:14:26.450 --> 00:14:28.850 quickly, but still getting a real sense of the 00:14:28.850 --> 00:14:30.720 depth and breadth of what's happening. Right. 00:14:30.820 --> 00:14:33.700 This is just a snapshot, of course, but hopefully 00:14:33.700 --> 00:14:37.559 it gives you a much clearer picture of some of 00:14:37.559 --> 00:14:39.480 the most interesting and significant things happening 00:14:39.480 --> 00:14:42.440 in AI right now. And if we leave you with one 00:14:42.440 --> 00:14:44.860 final thought, based on everything we've discussed 00:14:44.860 --> 00:14:47.960 today, if AI systems are becoming increasingly 00:14:47.960 --> 00:14:51.340 capable of teaching themselves, adapting and 00:14:51.340 --> 00:14:53.980 evolving continuously without constant human 00:14:53.980 --> 00:14:56.590 intervention. What does that fundamentally change 00:14:56.590 --> 00:14:58.850 about our timeline for technological progress 00:14:58.850 --> 00:15:01.129 or even our definition of intelligence itself 00:15:01.129 --> 00:15:03.389 in the years to come? Yeah. Something to really 00:15:03.389 --> 00:15:05.990 mull over. Absolutely. Keep exploring. Keep learning.