WEBVTT 00:00:00.000 --> 00:00:02.720 Imagine a future where AI spots a neurological 00:00:02.720 --> 00:00:05.900 disease maybe weeks before symptoms even show 00:00:05.900 --> 00:00:07.960 up. And then it actually guides a therapy using 00:00:07.960 --> 00:00:10.939 light. Yeah, a light therapy to literally help 00:00:10.939 --> 00:00:14.519 save brain cells. Well, that future. We're going 00:00:14.519 --> 00:00:16.199 to dive into some research today that suggests 00:00:16.199 --> 00:00:18.960 it might be closer than you think. Welcome to 00:00:18.960 --> 00:00:22.140 the Deep Dive. We're here again to unpack your 00:00:22.140 --> 00:00:24.480 latest stack of sources. Yeah, pulling out those 00:00:24.480 --> 00:00:27.140 key insights, the important facts, maybe some 00:00:27.140 --> 00:00:29.589 surprising stuff too. Today, we've got a really 00:00:29.589 --> 00:00:32.549 fascinating breakthrough. It combines AI and 00:00:32.549 --> 00:00:35.990 light therapy for Parkinson's research. And then 00:00:35.990 --> 00:00:38.210 we'll swing wide, look at AI out in the wild. 00:00:38.429 --> 00:00:41.270 Some useful things, some, well, some controversies 00:00:41.270 --> 00:00:43.450 too. Right. And finally, we'll dig into what 00:00:43.450 --> 00:00:46.369 makes AI seem creative, what's actually going 00:00:46.369 --> 00:00:48.609 on there. Our goal, as always, is to give you 00:00:48.609 --> 00:00:50.950 that clear path, help you grasp these complex 00:00:50.950 --> 00:00:54.350 topics. And get those aha moments without feeling 00:00:54.350 --> 00:00:57.659 overloaded. So let's just jump right in. So this 00:00:57.659 --> 00:01:00.039 first deep dive, it takes us into something I 00:01:00.039 --> 00:01:02.280 think is truly groundbreaking. It's a recent 00:01:02.280 --> 00:01:06.079 study pairing AI with optogenetics. Right. Optogenetics. 00:01:06.099 --> 00:01:08.299 That's using light to control cells, basically. 00:01:08.439 --> 00:01:11.379 Exactly. And it's not just about spotting Parkinson's 00:01:11.379 --> 00:01:14.159 early in these mouse models. It's about a targeted 00:01:14.159 --> 00:01:16.180 treatment. Yeah. The potential there is huge. 00:01:16.359 --> 00:01:18.480 What's really cool is the strategy they used. 00:01:18.519 --> 00:01:20.819 Kind of two parts. Okay. First, the AI part. 00:01:21.260 --> 00:01:24.060 They built this incredibly detailed 3D pose estimation 00:01:24.060 --> 00:01:28.480 system. So like a motion tracker, but super detailed. 00:01:28.700 --> 00:01:31.480 Exactly. Super, super detailed. For mice, it 00:01:31.480 --> 00:01:33.980 tracks over 340 different movement features. 00:01:34.260 --> 00:01:37.920 Wow, 340. That's a lot of detail for tracking 00:01:37.920 --> 00:01:40.900 tiny mouse movement. It is. And that huge amount 00:01:40.900 --> 00:01:44.400 of data lets the AI learn subtle patterns. Which 00:01:44.400 --> 00:01:47.519 leads to... The accuracy, I guess. Precisely. 00:01:47.540 --> 00:01:50.760 The system hits 90 % accuracy for spotting early 00:01:50.760 --> 00:01:53.980 stage Parkinson's. 90%. And this is the key part. 00:01:54.060 --> 00:01:57.870 It does it. weeks before the standard tests could 00:01:57.870 --> 00:02:00.230 even detect anything. Weeks earlier, that early 00:02:00.230 --> 00:02:02.989 detection window, that changes everything, potentially. 00:02:02.989 --> 00:02:05.469 Especially with neurodegenerative diseases, time 00:02:05.469 --> 00:02:07.650 is critical. So that's the detection side. But 00:02:07.650 --> 00:02:09.650 then there's the therapy part. That's where the 00:02:09.650 --> 00:02:12.129 optogenetics comes in. Right. So the researchers 00:02:12.129 --> 00:02:16.509 use these special wireless LED cages. Wireless 00:02:16.509 --> 00:02:19.250 cages. Yeah, to deliver pulses of blue light 00:02:19.250 --> 00:02:21.990 directly into specific brain regions in the mice. 00:02:22.379 --> 00:02:25.759 Okay. And they targeted mice with mild Parkinson's 00:02:25.759 --> 00:02:28.580 symptoms? Mm -hmm. They applied this light therapy 00:02:28.580 --> 00:02:32.039 on an alternate day schedule. And the results 00:02:32.039 --> 00:02:34.400 for those mice? Genuinely profound is the word 00:02:34.400 --> 00:02:36.659 the source used. It actually prevented the disease 00:02:36.659 --> 00:02:39.080 from getting much worse. Prevented severe progression. 00:02:39.139 --> 00:02:42.219 That's significant. It's huge. They saw about 00:02:42.219 --> 00:02:45.340 90 % of the dopamine neurons preserved in the 00:02:45.340 --> 00:02:48.280 mice that got the treatment. 90%. Compared to 00:02:48.280 --> 00:02:51.280 what in the untreated ones? Only on 38%. So a 00:02:51.280 --> 00:02:53.500 massive difference. Wow. And it wasn't just the 00:02:53.500 --> 00:02:55.759 neurons. Their gait went back to normal. Coordination 00:02:55.759 --> 00:02:58.800 got better. Tremors reduced. Real functional 00:02:58.800 --> 00:03:00.939 improvements. That really paints a picture of 00:03:00.939 --> 00:03:02.919 what intervention could look like. What about 00:03:02.919 --> 00:03:05.699 mice with more severe PD? The effects were more 00:03:05.699 --> 00:03:08.340 limited there, understandably. But it still showed 00:03:08.340 --> 00:03:10.580 some positive impact, suggesting potential even 00:03:10.580 --> 00:03:13.439 later on. So this whole setup, AI detecting, 00:03:13.759 --> 00:03:16.539 light treating, feels like a blueprint. Exactly. 00:03:16.639 --> 00:03:18.979 They're calling it closed -loop neuromedicine. 00:03:19.060 --> 00:03:22.120 Meaning? Meaning the AI diagnoses, the optogenetics 00:03:22.120 --> 00:03:24.819 delivers targeted therapy, and then crucially, 00:03:24.900 --> 00:03:28.280 both parts adapt over time. The system learns. 00:03:28.479 --> 00:03:32.819 It gets smarter. Like stacking Lego blocks of 00:03:32.819 --> 00:03:35.060 data and light to fight the disease. That's a 00:03:35.060 --> 00:03:37.319 great analogy. Yeah. Stacking data and light. 00:03:37.419 --> 00:03:40.080 It's intelligent. It's responsive. Yeah. Now, 00:03:40.199 --> 00:03:42.319 obviously, this is early days. It's mouse models, 00:03:42.400 --> 00:03:44.300 not humans. We have to stress that. Of course. 00:03:44.319 --> 00:03:47.020 Big jump from mice to people. But it really does 00:03:47.020 --> 00:03:49.219 highlight the potential, doesn't it? Combining 00:03:49.219 --> 00:03:51.659 AI and bioengineering for these really tough 00:03:51.659 --> 00:03:54.020 diseases. It absolutely does. So stepping back 00:03:54.020 --> 00:03:56.539 a bit, what's the biggest takeaway for you seeing 00:03:56.539 --> 00:03:59.759 AI and this light therapy working together so 00:03:59.759 --> 00:04:02.139 precisely? Precisely, so adaptively. For me, 00:04:02.159 --> 00:04:05.900 it signals a shift towards truly responsive medicine, 00:04:06.159 --> 00:04:09.120 AI guiding adaptive, highly targeted treatments. 00:04:09.479 --> 00:04:12.159 Okay, so from the microscopic world inside the 00:04:12.159 --> 00:04:14.199 brain, let's zoom out. Let's look at AI's impact 00:04:14.199 --> 00:04:16.819 kind of everywhere else in our daily lives. Yeah, 00:04:16.879 --> 00:04:18.819 the sources cover a real mix this time. Some 00:04:18.819 --> 00:04:21.420 really useful tools, some pretty big controversies, 00:04:21.420 --> 00:04:24.600 and some, well, ambitious visions for the future. 00:04:24.779 --> 00:04:28.019 It really is touching everything. One source 00:04:28.019 --> 00:04:31.300 mentioned a Redditor using AI. Oh, yeah, that 00:04:31.300 --> 00:04:33.259 was pretty cool. They built this tool with AI 00:04:33.259 --> 00:04:36.620 to filter LinkedIn job postings much more effectively. 00:04:36.819 --> 00:04:39.279 How many jobs did it scrape? Over 4 .1 million 00:04:39.279 --> 00:04:42.160 directly. And they made the tool free, trying 00:04:42.160 --> 00:04:44.600 to tackle that whole ghost jobs problem, help 00:04:44.600 --> 00:04:47.300 people find real active roles. That's a fantastic 00:04:47.300 --> 00:04:50.259 example of AI being used for something practical, 00:04:50.439 --> 00:04:52.939 kind of democratizing access. Definitely. And 00:04:52.939 --> 00:04:56.160 speaking of practical value, Andrew Ng. you know, 00:04:56.160 --> 00:04:59.000 from Google Brain. Right. He laid out five big 00:04:59.000 --> 00:05:01.639 opportunities where he sees potential for, well, 00:05:01.699 --> 00:05:04.480 creating significant wealth using AI, a clear 00:05:04.480 --> 00:05:07.399 sign of the economic shift happening. It's definitely 00:05:07.399 --> 00:05:10.079 changing the landscape. On the maybe stranger 00:05:10.079 --> 00:05:12.560 side of things. Light laugh. Yeah. Yeah, the 00:05:12.560 --> 00:05:15.459 AI ASMR food challenges are apparently back and 00:05:15.459 --> 00:05:17.800 going viral. Wait, AI eating? Yeah, generating 00:05:17.800 --> 00:05:21.139 clips of, like, digital food crunching and sizzling. 00:05:21.259 --> 00:05:24.600 Tools like BaseLab's AI making these millions 00:05:24.600 --> 00:05:27.000 of views. The Internet's a weird place sometimes. 00:05:27.139 --> 00:05:29.660 It certainly is. Okay, but back to useful resources. 00:05:30.100 --> 00:05:33.480 Anthropic released a guide. Yeah, a really detailed 00:05:33.480 --> 00:05:36.779 step -by -step guide on prompt engineering. Which 00:05:36.779 --> 00:05:39.790 is basically... How to talk to an AI effectively. 00:05:40.149 --> 00:05:42.550 Exactly. How to craft good instructions to get 00:05:42.550 --> 00:05:44.930 the results you want. Super valuable for anyone 00:05:44.930 --> 00:05:47.050 really using these tools. Definitely sounds useful, 00:05:47.170 --> 00:05:49.790 but it's not all positive applications, is it? 00:05:49.790 --> 00:05:52.129 There are challenges. No, definitely not. We're 00:05:52.129 --> 00:05:54.870 seeing real issues emerge, like these AI crawlers, 00:05:54.889 --> 00:05:57.470 the bots from Meta, OpenAI. What are they doing? 00:05:57.790 --> 00:05:59.290 They're just hammering websites with traffic. 00:05:59.509 --> 00:06:01.730 One source mentioned a bot hitting a site 39 00:06:01.730 --> 00:06:05.290 ,000 times a minute. 39 ,000 per minute. Yeah. 00:06:05.579 --> 00:06:07.579 It's overwhelming smaller sites, making it hard 00:06:07.579 --> 00:06:10.019 for them to even stay online. Huge strain. That's 00:06:10.019 --> 00:06:12.079 a serious unintended consequence. And speaking 00:06:12.079 --> 00:06:14.100 of things happening without much transparency, 00:06:14.500 --> 00:06:18.480 YouTube. Ah, yeah. YouTube apparently used AI 00:06:18.480 --> 00:06:21.660 to silently edit something like 20 billion shorts 00:06:21.660 --> 00:06:25.360 videos. Silently edited. Without creators knowing. 00:06:25.680 --> 00:06:28.600 Seems that way. Subtle edits, maybe stabilization 00:06:28.600 --> 00:06:31.120 or something, but done without explicit consent. 00:06:31.660 --> 00:06:33.899 And once you notice it, it feels kind of off. 00:06:34.040 --> 00:06:36.220 Yeah. That raises big questions about control 00:06:36.220 --> 00:06:38.639 and transparency, doesn't it? Absolutely. Where 00:06:38.639 --> 00:06:41.660 is that line? How much transparency do creators 00:06:41.660 --> 00:06:45.160 deserve when platforms use AI on their work? 00:06:45.279 --> 00:06:47.279 It's a critical question. It really is. On a 00:06:47.279 --> 00:06:48.839 totally different scale, though, the investment 00:06:48.839 --> 00:06:52.240 is just pouring in. Field AI. Right. Raised $405 00:06:52.240 --> 00:06:56.160 million. Yeah. Backed by huge names, Bezos, Intel. 00:06:56.379 --> 00:06:59.670 And their goal is... Ambitious. Very. Building 00:06:59.670 --> 00:07:03.329 a universal AI robot brain. One AI that can work 00:07:03.329 --> 00:07:05.910 across basically any type of robot. Incorporating 00:07:05.910 --> 00:07:08.509 physics too for safety. Exactly. Physics -based 00:07:08.509 --> 00:07:10.810 understanding for smarter, safer decisions in 00:07:10.810 --> 00:07:13.430 the real world. Whoa. Just imagine scaling that. 00:07:13.529 --> 00:07:17.329 One AI brain adaptable to like... a billion different 00:07:17.329 --> 00:07:20.750 robot tasks manufacturing logistics maybe even 00:07:20.750 --> 00:07:23.029 disaster response it's a moment where you really 00:07:23.029 --> 00:07:25.569 see the potential scale yeah mind -boggling implications 00:07:25.569 --> 00:07:28.089 it is things are moving so fast we had some quick 00:07:28.089 --> 00:07:31.730 hits too right yeah rapid fire grok 2 .5 is open 00:07:31.730 --> 00:07:34.730 source now grok 3 maybe in six months okay meta's 00:07:34.730 --> 00:07:36.649 partnering with mid journey on image and video 00:07:36.649 --> 00:07:40.620 ai Big combo there. OpenAI warned people about 00:07:40.620 --> 00:07:43.420 dodgy investments like unauthorized SPVs. Good 00:07:43.420 --> 00:07:46.199 to know. AI art is getting scarily good tests. 00:07:46.339 --> 00:07:49.939 Show aesthetic AI pieces are basically indistinguishable 00:07:49.939 --> 00:07:53.699 from human work now. Wow. And some really cool 00:07:53.699 --> 00:07:56.439 research looking at tiny movements in bee brains. 00:07:56.819 --> 00:08:00.079 Might hold clues for smarter AI design. Bee brains. 00:08:00.160 --> 00:08:02.199 OK, that's a lot happening all at once. It really 00:08:02.199 --> 00:08:05.000 shows the speed. So with all these amazing tools, 00:08:05.120 --> 00:08:07.980 but also these potential downsides. How do we 00:08:07.980 --> 00:08:10.120 strike that balance? How do we manage responsible 00:08:10.120 --> 00:08:13.540 use? It demands constant vigilance, really clear 00:08:13.540 --> 00:08:16.060 ethical guidelines that can adapt as fast as 00:08:16.060 --> 00:08:17.959 the tech does. Okay, let's shift gears one last 00:08:17.959 --> 00:08:20.399 time to something, oh, maybe a bit philosophical, 00:08:20.540 --> 00:08:24.639 the idea of AI creativity. Ah, yes. Models like 00:08:24.639 --> 00:08:27.620 Daolati, stable diffusion, they feel creative, 00:08:27.839 --> 00:08:30.040 don't they? They really do. They improvise, they 00:08:30.040 --> 00:08:32.299 blend patterns, they make things that seem genuinely 00:08:32.299 --> 00:08:35.879 new. But the sources suggest it's mine. Well, 00:08:35.940 --> 00:08:38.710 it's more like an illusion. A very, very clever 00:08:38.710 --> 00:08:41.929 illusion based on math. Not magic, just math. 00:08:42.309 --> 00:08:44.789 Kind of takes the mystery out of it, maybe. Slight 00:08:44.789 --> 00:08:47.690 laugh, a little bit. But the math itself is fascinating. 00:08:47.870 --> 00:08:49.769 These are diffusion models, right? Explain that 00:08:49.769 --> 00:08:52.009 again simply. They basically start with random 00:08:52.009 --> 00:08:55.649 noise, like static on our old TV. Okay. And then 00:08:55.649 --> 00:08:58.470 they gradually remove the noise step by step, 00:08:58.529 --> 00:09:01.429 following patterns they learned, until a clear 00:09:01.429 --> 00:09:05.679 image emerges, denoising. Got it. Start with 00:09:05.679 --> 00:09:08.179 chaos and with order, but how does that lead 00:09:08.179 --> 00:09:11.659 to creativity? It comes down to two key quirks 00:09:11.659 --> 00:09:13.620 in how they're designed. The first one is called 00:09:13.620 --> 00:09:16.860 locality. Locality. Yeah, it means the AI doesn't 00:09:16.860 --> 00:09:19.059 see the whole picture at once. It focuses on 00:09:19.059 --> 00:09:21.480 generating just one small patch of pixels at 00:09:21.480 --> 00:09:24.220 a time. Like building a mosaic tile by tile without 00:09:24.220 --> 00:09:26.460 seeing the full design initially. Exactly like 00:09:26.460 --> 00:09:29.019 that. It forces the model to synthesize the bigger 00:09:29.019 --> 00:09:31.899 picture from lots of small local decisions. Okay, 00:09:31.940 --> 00:09:34.220 that's one quirk. What's the second? Equivariance. 00:09:34.259 --> 00:09:35.940 This basically means if you shift the input, 00:09:36.039 --> 00:09:38.480 the output shifts the same way. Like if I ask 00:09:38.480 --> 00:09:40.580 for a cat on the left and then ask for a cat 00:09:40.580 --> 00:09:42.960 on the right, the catness moves. Pretty much. 00:09:43.100 --> 00:09:45.840 The generated features move predictably with 00:09:45.840 --> 00:09:48.019 changes in the input position or orientation. 00:09:48.440 --> 00:09:52.559 So locality and equivariance. Yeah. How do those 00:09:52.559 --> 00:09:55.440 two things make it creative? Because together 00:09:55.440 --> 00:09:57.799 they actually stop the model from just perfectly 00:09:57.799 --> 00:10:00.740 copying images from its training data. Ah, they 00:10:00.740 --> 00:10:02.700 constrain it. Right. It can't just reproduce. 00:10:02.960 --> 00:10:05.019 It's forced to assemble those little patches 00:10:05.019 --> 00:10:07.539 based on the patterns it knows, but in ways that 00:10:07.539 --> 00:10:10.379 fit the prompt and the neighboring patches. That 00:10:10.379 --> 00:10:12.620 assembly process leads to novel combinations. 00:10:13.200 --> 00:10:16.799 So the novelty, the apparent improvisation, comes 00:10:16.799 --> 00:10:18.759 from the constraints of building it piece by 00:10:18.759 --> 00:10:21.690 piece. That seems to be the idea. Originality 00:10:21.690 --> 00:10:23.769 emerges as a sort of byproduct of the architecture 00:10:23.769 --> 00:10:26.529 itself. That's actually quite profound. And researchers 00:10:26.529 --> 00:10:28.750 tested this. They did. They built a simplified 00:10:28.750 --> 00:10:31.710 mathematical model called an ELS machine to mimic 00:10:31.710 --> 00:10:34.730 just these core principles. And its output was 00:10:34.730 --> 00:10:38.129 90 % identical to the actual complex trained 00:10:38.129 --> 00:10:41.450 AI models. 90 % identical, just from the core 00:10:41.450 --> 00:10:44.590 mathematical principles. Yeah. It strongly suggests 00:10:44.590 --> 00:10:48.110 that this creativity we see isn't some emergent 00:10:48.110 --> 00:10:52.330 consciousness, but a determined... So it's following 00:10:52.330 --> 00:10:55.169 rules, very complex rules, that happen to produce 00:10:55.169 --> 00:10:58.470 outputs we perceive as novel and creative. Precisely. 00:10:58.750 --> 00:11:00.870 But then the source takes it a step further, 00:11:01.049 --> 00:11:03.649 drawing a parallel to human creativity. How so? 00:11:03.990 --> 00:11:06.090 It suggests maybe our own creativity isn't so 00:11:06.090 --> 00:11:08.830 different. Maybe we also largely recombine things 00:11:08.830 --> 00:11:11.070 we've seen and learned in new ways. That's a 00:11:11.070 --> 00:11:13.870 big thought, that we get originality for free 00:11:13.870 --> 00:11:16.409 just by how our own minds piece things together. 00:11:16.919 --> 00:11:19.100 You know, I still wrestle with prompt drift myself 00:11:19.100 --> 00:11:21.399 sometimes trying to get these models to produce 00:11:21.399 --> 00:11:23.340 something truly unexpected, something beyond 00:11:23.340 --> 00:11:25.759 recombination. It's harder than it looks. It 00:11:25.759 --> 00:11:28.080 really is. Which leads to the big question here. 00:11:28.159 --> 00:11:31.639 Yeah. If AI creativity is essentially this deterministic 00:11:31.639 --> 00:11:34.759 process of recombination based on its structure, 00:11:34.899 --> 00:11:38.220 does that change how we think about human creativity 00:11:38.220 --> 00:11:41.179 at its core? It definitely suggests human and 00:11:41.179 --> 00:11:44.059 AI creativity might share more fundamental mechanisms 00:11:44.059 --> 00:11:46.879 than we previously thought. A lot to chew on 00:11:46.879 --> 00:11:49.059 there. So wrapping things up for today, we've 00:11:49.059 --> 00:11:51.940 gone from neurons and light therapy. All the 00:11:51.940 --> 00:11:54.200 way to the mechanics of digital art creation. 00:11:54.519 --> 00:11:57.580 It's really spanned the breadth of AI's impact 00:11:57.580 --> 00:12:00.779 right now. Yeah, we've seen AI as this incredibly 00:12:00.779 --> 00:12:05.320 precise potential guide for medicine. And as 00:12:05.320 --> 00:12:09.460 a powerful force online. useful tools, but also 00:12:09.460 --> 00:12:12.679 some real problems and controversies. And then 00:12:12.679 --> 00:12:15.100 this dive into AI as a surprisingly creative 00:12:15.100 --> 00:12:18.539 artist, making us question originality itself. 00:12:18.860 --> 00:12:20.919 You know, the thread connecting all of this for 00:12:20.919 --> 00:12:23.860 me is AI's power to process information and make 00:12:23.860 --> 00:12:26.639 connections. Yeah, often in ways that still genuinely 00:12:26.639 --> 00:12:28.799 surprise us. It keeps opening doors we didn't 00:12:28.799 --> 00:12:31.169 even know were there. This deep dive really shows 00:12:31.169 --> 00:12:33.509 just the incredible scope of what AI is doing. 00:12:33.590 --> 00:12:35.789 It's transformative, truly. Changing science, 00:12:35.909 --> 00:12:38.149 definitely changing how we work, and maybe even 00:12:38.149 --> 00:12:40.909 changing how we understand art, creativity, originality. 00:12:41.049 --> 00:12:43.049 So thinking back on everything we covered, what's 00:12:43.049 --> 00:12:45.070 the one thing that really stands out to you today? 00:12:45.590 --> 00:12:49.409 Maybe it's this. If AI's creativity stems from 00:12:49.409 --> 00:12:51.870 combining existing pieces in new ways because 00:12:51.870 --> 00:12:54.330 of its structure, what does that suggest about 00:12:54.330 --> 00:12:57.299 the ultimate source of innovation? Even human 00:12:57.299 --> 00:13:00.240 innovation. Is it all just clever recombination 00:13:00.240 --> 00:13:02.659 at some level? That is definitely a great question 00:13:02.659 --> 00:13:04.639 to ponder. Thank you for joining us on the Deep 00:13:04.639 --> 00:13:06.879 Dive. We'll be back soon, ready to explore another 00:13:06.879 --> 00:13:09.740 stack of sources. Until next time, keep digging 00:13:09.740 --> 00:13:11.639 deeper. Out to your own music.