WEBVTT 00:00:00.000 --> 00:00:01.820 Hey, have you ever used an AI tool and afterwards 00:00:01.820 --> 00:00:05.219 felt a little fuzzy, like your own brain kind 00:00:05.219 --> 00:00:06.719 of took a backseat for a bit? Oh, definitely. 00:00:06.960 --> 00:00:08.800 It's a weird sensation, right? It really is. 00:00:09.080 --> 00:00:11.660 Well, today we're going to take a deep dive into 00:00:11.660 --> 00:00:16.059 that very feeling. We're exploring this groundbreaking 00:00:16.059 --> 00:00:21.160 MIT study that really unpacks what AI might be 00:00:21.160 --> 00:00:23.239 doing to our thinking. But we're also balancing 00:00:23.239 --> 00:00:26.519 that. Exactly. Balancing that with some truly 00:00:26.519 --> 00:00:29.300 mind -blowing advancements happening in AI itself. 00:00:29.519 --> 00:00:32.340 We've pulled together a stack of sources for 00:00:32.340 --> 00:00:35.020 you from that key MID research to all sorts of 00:00:35.020 --> 00:00:37.420 updates and breakthroughs. It's going to be a 00:00:37.420 --> 00:00:40.159 lot to explore. Yeah. What's so fascinating here 00:00:40.159 --> 00:00:42.920 is just the sheer speed of change in AI right 00:00:42.920 --> 00:00:45.359 now and what that ultimately means for you, the 00:00:45.359 --> 00:00:47.859 listener. We're talking about, sure, profound 00:00:47.859 --> 00:00:50.119 opportunities, absolutely, but also some really 00:00:50.119 --> 00:00:53.640 nuanced challenges, especially to our... cognition 00:00:53.640 --> 00:00:56.100 right this deep dive isn't just about what AI 00:00:56.100 --> 00:00:57.899 can do though believe me we'll get into some 00:00:57.899 --> 00:01:00.359 incredible new tools it's also about what it 00:01:00.359 --> 00:01:03.399 does to us and how we adapt or how we should 00:01:03.399 --> 00:01:07.079 adapt maybe okay let's dig into this let's call 00:01:07.079 --> 00:01:10.599 it concerning finding from MIT It's all about 00:01:10.599 --> 00:01:13.519 what chat GPT or, you know, LLMs in general might 00:01:13.519 --> 00:01:15.640 be doing to our brains when we use them for stuff 00:01:15.640 --> 00:01:17.760 like writing essays. And the core finding, it's 00:01:17.760 --> 00:01:19.900 kind of counterintuitive, like you said. Yeah. 00:01:20.260 --> 00:01:23.200 AI actually made people better at the task, at 00:01:23.200 --> 00:01:26.959 writing, but simultaneously maybe worse at thinking. 00:01:27.579 --> 00:01:30.000 Sounds a little wild, doesn't it? It does. So 00:01:30.000 --> 00:01:31.579 how did they even figure this out? What was the 00:01:31.579 --> 00:01:34.659 setup? Well, the study was pretty rigorously 00:01:34.659 --> 00:01:37.079 designed, actually, to isolate these effects. 00:01:37.540 --> 00:01:40.859 MIT brought in 54 participants over four months, 00:01:40.980 --> 00:01:43.319 all focused on essay writing. Okay. They split 00:01:43.319 --> 00:01:45.659 them into three groups. One group used ChatGPT. 00:01:46.019 --> 00:01:48.159 Another used Google, kind of like, you know, 00:01:48.180 --> 00:01:50.359 how we might have done research back in the day, 00:01:50.379 --> 00:01:52.719 just basic search. Right, 90s kid style. Exactly. 00:01:52.939 --> 00:01:55.719 And then the third group, totally analog. Just 00:01:55.719 --> 00:01:58.400 their brain, pen and paper, no digital help. 00:01:58.540 --> 00:02:00.700 Interesting. And the results? They were really 00:02:00.700 --> 00:02:03.500 striking. The ChatGPT group showed a significant 00:02:03.500 --> 00:02:06.780 drop in brain connectivity, measured by EEG, 00:02:06.920 --> 00:02:09.340 you know, mapping brain activity. I mean, it 00:02:09.340 --> 00:02:11.599 literally flatlined in some areas during the 00:02:11.599 --> 00:02:14.039 writing test. Flatlined? Seriously? Yeah. And 00:02:14.039 --> 00:02:15.979 what was even more telling was the follow -up. 00:02:16.020 --> 00:02:20.419 When these same ChatGPT users were asked to write 00:02:20.419 --> 00:02:24.509 without AI afterwards, They performed like novices, 00:02:24.550 --> 00:02:26.590 like people who hadn't been practicing writing 00:02:26.590 --> 00:02:29.669 for months, which really highlights the dependency, 00:02:30.069 --> 00:02:32.550 you know? Wow. So the brain's literally like 00:02:32.550 --> 00:02:35.509 disengaging. That's pretty intense. It really 00:02:35.509 --> 00:02:37.550 is. And get this. Here's a truly astonishing 00:02:37.550 --> 00:02:42.629 statistic from the study. Okay. 83%. 83 % of 00:02:42.629 --> 00:02:45.229 the chat GPT users couldn't quote a single line 00:02:45.229 --> 00:02:48.710 they had just written using the AI. 83%. Compared 00:02:48.710 --> 00:02:51.590 to what? Compared to only 11 % of the brain -only 00:02:51.590 --> 00:02:54.569 group. Oh, man. It's almost like if you let AI 00:02:54.569 --> 00:02:56.849 do the heavy lifting, your brain just goes, cool, 00:02:56.949 --> 00:03:00.030 I'm clocking out now. That's kind of terrifying, 00:03:00.229 --> 00:03:01.909 actually, because here's where it gets really 00:03:01.909 --> 00:03:04.770 interesting, you know, the maybe creepiest part. 00:03:04.889 --> 00:03:07.409 The users themselves felt fine. They didn't report 00:03:07.409 --> 00:03:10.789 any mental fog, no perceived decline, nothing. 00:03:10.969 --> 00:03:13.810 Right. The subjective experience was okay. Exactly. 00:03:13.870 --> 00:03:16.250 The only way this drop was visible was on those 00:03:16.250 --> 00:03:19.129 EEGs. Yeah. It's like this invisible trade -off, 00:03:19.129 --> 00:03:21.349 isn't it? You get efficiency, maybe better output 00:03:21.349 --> 00:03:24.250 sometimes, but at a hidden cognitive cost. Yeah. 00:03:24.430 --> 00:03:26.990 And MIT even coined a term for this, cognitive 00:03:26.990 --> 00:03:29.250 debt. Right. Think of it like the Google Maps 00:03:29.250 --> 00:03:32.759 effect. Yeah. The more you rely on GPS. The worse 00:03:32.759 --> 00:03:35.219 your own sense of direction gets. Exactly. You're 00:03:35.219 --> 00:03:38.120 essentially borrowing against your future cognitive 00:03:38.120 --> 00:03:41.199 capacity for that short -term convenience. That's 00:03:41.199 --> 00:03:44.379 a heavy concept. It is. But, and this is crucial, 00:03:44.539 --> 00:03:47.419 the study is absolutely not anti -AI. That's 00:03:47.419 --> 00:03:49.860 a key takeaway. Okay, good. Because they actually 00:03:49.860 --> 00:03:52.840 discovered a brain -friendly AI workflow within 00:03:52.840 --> 00:03:55.240 the study. So it raises this important question 00:03:55.240 --> 00:03:57.939 for all of us. How do we engage productively? 00:03:58.039 --> 00:04:00.400 How do we use these tools without racking up 00:04:00.400 --> 00:04:02.539 that debt? But what did they find? It was a brain 00:04:02.539 --> 00:04:04.460 -friendly way. The optimal workflow they identified 00:04:04.460 --> 00:04:07.879 was first think, really process the ideas yourself, 00:04:08.120 --> 00:04:11.639 then write your own draft, and then only then 00:04:11.639 --> 00:04:14.819 use chat GPT to maybe polish, refine, or get 00:04:14.819 --> 00:04:17.730 feedback. Ah, so the AI comes in later. Exactly. 00:04:17.990 --> 00:04:20.430 And what's truly fascinating here is that people 00:04:20.430 --> 00:04:23.610 who used this method, they actually showed stronger 00:04:23.610 --> 00:04:26.230 brain connectivity than anyone else in the study. 00:04:26.370 --> 00:04:28.490 Stronger than even the brain -only group. Stronger 00:04:28.490 --> 00:04:30.370 than the brain -only group. That's a huge insight, 00:04:30.490 --> 00:04:33.649 right? It suggests thoughtful AI integration 00:04:33.649 --> 00:04:37.209 can actually enhance cognition. Wow. But it really 00:04:37.209 --> 00:04:39.990 highlights that if you just rely on AI from the 00:04:39.990 --> 00:04:42.170 start, letting it do the initial heavy lifting, 00:04:42.410 --> 00:04:46.699 well, your cognitive core can, you know... get 00:04:46.699 --> 00:04:49.139 a bit mushy. Right. So if you're still in that 00:04:49.139 --> 00:04:51.699 phase of, oh, AI can just write it for me, maybe 00:04:51.699 --> 00:04:54.199 rethink that. Let your brain take the wheel. 00:04:54.240 --> 00:04:56.220 At least sometimes your hippocampus will thank 00:04:56.220 --> 00:04:58.180 you later. That's such an important distinction. 00:04:58.420 --> 00:05:00.519 It's not about avoiding the tool. It's about 00:05:00.519 --> 00:05:03.620 how you use it thoughtfully. Precisely. But hey, 00:05:03.720 --> 00:05:05.800 it's not all about potential brain drain, right? 00:05:07.319 --> 00:05:09.160 There's some incredible stuff happening out there, 00:05:09.199 --> 00:05:11.509 too. Let's kind of zoom out a bit. Talk about 00:05:11.509 --> 00:05:13.930 the broader AI landscape. What's been buzzing? 00:05:14.129 --> 00:05:16.490 Yeah, there's a lot going on. For sure. Take 00:05:16.490 --> 00:05:19.670 AI video models. Just recently, we saw ByteDance's 00:05:19.670 --> 00:05:22.889 Seedance 1 .0 and then this kind of sneaky new 00:05:22.889 --> 00:05:25.250 contender, Hailu O2. Oh, yeah. They jumped up 00:05:25.250 --> 00:05:27.550 the leaderboards fast. Right. Leapfrogging models 00:05:27.550 --> 00:05:30.790 like Google's VO3 almost overnight just shows 00:05:30.790 --> 00:05:33.350 how dynamic this space is. Innovation cycles 00:05:33.350 --> 00:05:36.290 are shrinking like crazy. Wild. And speaking 00:05:36.290 --> 00:05:39.480 of VO3, apparently clever TikTok users. are racking 00:05:39.480 --> 00:05:41.779 up millions of followers creating ASMR videos 00:05:41.779 --> 00:05:45.399 using its new native audio features. Interesting 00:05:45.399 --> 00:05:49.680 use case. Yeah. VO3 seems to be like a goldmine 00:05:49.680 --> 00:05:52.740 for creators right now. Oh, and get this. You 00:05:52.740 --> 00:05:55.800 can even generate images with ChatGPT just by 00:05:55.800 --> 00:05:58.319 texting a specific 1 -800 number on WhatsApp. 00:05:58.579 --> 00:06:01.209 On WhatsApp. Really? Yeah. Don't even need an 00:06:01.209 --> 00:06:03.290 account, apparently. You get one free image daily. 00:06:03.470 --> 00:06:05.610 Have you tried that yourself? I haven't tried 00:06:05.610 --> 00:06:07.470 the WhatsApp thing yet, no. But I've seen some 00:06:07.470 --> 00:06:09.610 impressive outputs people are sharing. It just 00:06:09.610 --> 00:06:11.769 speaks to how accessible these tools are becoming, 00:06:11.930 --> 00:06:13.930 you know? Totally. But if we connect this all 00:06:13.930 --> 00:06:16.550 back to the bigger picture, it's not just exciting 00:06:16.550 --> 00:06:19.350 new toys. There are also some really significant 00:06:19.350 --> 00:06:21.930 considerations, things we need to be aware of. 00:06:22.149 --> 00:06:24.689 Like what? Well, for instance, Jeffrey Hinton, 00:06:24.889 --> 00:06:27.730 you know, often called the godfather of AI. Yeah. 00:06:28.060 --> 00:06:30.319 Heard of him. He's openly stated there's maybe 00:06:30.319 --> 00:06:34.680 a 10 % to 20 % chance that AI will displace humans 00:06:34.680 --> 00:06:37.540 completely in some sectors. He even mentioned 00:06:37.540 --> 00:06:40.060 specific fields he thinks face a heavier risk, 00:06:40.199 --> 00:06:44.079 saying they should be terrified. Wow. That's 00:06:44.079 --> 00:06:47.079 sobering. It is. It's not necessarily a prediction, 00:06:47.160 --> 00:06:49.579 more like a stark warning about the potential 00:06:49.579 --> 00:06:52.740 scale of disruption. And then you has Reid Hoffman, 00:06:52.860 --> 00:06:55.699 LinkedIn's co -founder. What's his take? What's 00:06:55.699 --> 00:06:57.680 fascinating about his advice is he's saying things 00:06:57.680 --> 00:07:01.220 like prompt hacks and what he calls vibe coding, 00:07:01.439 --> 00:07:03.079 you know, just guessing prompts without real 00:07:03.079 --> 00:07:05.060 understanding. He says that's not going to be 00:07:05.060 --> 00:07:07.360 enough to really thrive. He argues it's going 00:07:07.360 --> 00:07:09.579 to come down to actual judgment, critical thinking, 00:07:09.740 --> 00:07:12.560 applying these tools strategically. That's what 00:07:12.560 --> 00:07:15.019 will differentiate people. That makes sense beyond 00:07:15.019 --> 00:07:18.100 just tinkering. Right. And on a really practical 00:07:18.100 --> 00:07:19.959 note, something you definitely need to keep in 00:07:19.959 --> 00:07:24.300 mind is data privacy. Ah, yes. Important. Our 00:07:24.300 --> 00:07:26.560 sources indicate that a pretty significant chunk, 00:07:26.699 --> 00:07:31.060 like 63 % of ChatGPT user data, actually contains 00:07:31.060 --> 00:07:35.100 personally identifiable information, PII. 63%, 00:07:35.100 --> 00:07:39.209 that's high. It is. Yet only about 22 % of users 00:07:39.209 --> 00:07:41.709 are even aware they can opt out of data collection. 00:07:42.129 --> 00:07:45.149 And crucially, your inputs and outputs, they're 00:07:45.149 --> 00:07:47.410 being saved indefinitely. Indefinitely. Okay, 00:07:47.470 --> 00:07:49.329 that's something everyone should know. Definitely 00:07:49.329 --> 00:07:51.490 something to be aware of, you know, as you interact 00:07:51.490 --> 00:07:53.089 with these platforms. Absolutely. That privacy 00:07:53.089 --> 00:07:55.750 point is huge, knowing what's happening with 00:07:55.750 --> 00:07:58.949 your data. Yeah. But, okay. On a more positive 00:07:58.949 --> 00:08:01.529 note for specific industries, we are seeing some 00:08:01.529 --> 00:08:03.709 amazing efficiency gains, right? Oh, for sure. 00:08:03.930 --> 00:08:07.550 Like Nobla AI. This startup just got $70 million 00:08:07.550 --> 00:08:10.009 in Series C funding. Right, for the clinical 00:08:10.009 --> 00:08:12.490 documentation tools. Exactly. Used by like 85 00:08:12.490 --> 00:08:14.910 ,000 clinicians, and it cuts their documentation 00:08:14.910 --> 00:08:17.870 time by over 50%. That's pretty huge for healthcare 00:08:17.870 --> 00:08:20.470 efficiency, isn't it? Massive. Reducing burnout, 00:08:20.709 --> 00:08:23.790 freeing up time for patients. Big potential there. 00:08:23.970 --> 00:08:25.829 And then you look at the big players. Amazon 00:08:25.829 --> 00:08:28.509 is pouring a massive $100 billion into automation. 00:08:28.730 --> 00:08:32.049 Yeah. Over 1 ,000 AI projects underway. $100 00:08:32.049 --> 00:08:35.149 billion. Wow. Yeah. But interestingly, they've 00:08:35.149 --> 00:08:37.570 also announced some AI hiring freezes in certain 00:08:37.570 --> 00:08:40.370 areas. It just shows how rapidly this stuff is 00:08:40.370 --> 00:08:43.370 transforming operations across the board. Big 00:08:43.370 --> 00:08:45.149 shifts happening. Okay. But here's where it gets 00:08:45.149 --> 00:08:47.429 really interesting again. Let's talk about a 00:08:47.429 --> 00:08:52.980 major breakthrough. Chinese AI startup, Minimax. 00:08:53.179 --> 00:08:56.759 Ah, yes. Minimax and M1. They've kind of changed 00:08:56.759 --> 00:08:59.659 the game with their M1 model. We're talking major 00:08:59.659 --> 00:09:03.059 leap forward in AI capabilities here. Definitely 00:09:03.059 --> 00:09:04.620 pushing boundaries. It's the most impressive 00:09:04.620 --> 00:09:08.399 feature. Get ready for this. A one million token 00:09:08.399 --> 00:09:12.320 contest window. One million. That's huge. It's 00:09:12.320 --> 00:09:14.539 the biggest we've ever seen. Like, wow. Just 00:09:14.539 --> 00:09:17.240 massive. The amount of information it can hold 00:09:17.240 --> 00:09:19.840 in its working memory. At once. Game -changing 00:09:19.840 --> 00:09:23.419 potential. And here's the kicker. Get this. This 00:09:23.419 --> 00:09:26.240 whole model. Trained in just three weeks. Three 00:09:26.240 --> 00:09:28.840 weeks. That's incredibly fast. For under $540 00:09:28.840 --> 00:09:31.799 ,000. Under half a million? Yeah. Like not even 00:09:31.799 --> 00:09:33.200 close to a million dollars. Yeah. It's pretty 00:09:33.200 --> 00:09:34.980 incredible, right? Makes you really wonder about 00:09:34.980 --> 00:09:37.100 development costs and accessibility going forward. 00:09:37.299 --> 00:09:40.320 That efficiency is almost as impressive as the 00:09:40.320 --> 00:09:43.149 context window itself. So connecting this to 00:09:43.149 --> 00:09:46.490 the bigger picture, the significance of a 1 million 00:09:46.490 --> 00:09:50.629 token context window is hard to overstate. Explain 00:09:50.629 --> 00:09:53.370 that. What does it really mean? It means instead 00:09:53.370 --> 00:09:55.889 of having that short term memory loss you see 00:09:55.889 --> 00:09:57.590 in a lot of current models, you know, where they 00:09:57.590 --> 00:09:59.129 forget the start of the conversation. Right. 00:09:59.190 --> 00:10:01.250 Very frustrating. This thing can hold an entire 00:10:01.250 --> 00:10:04.049 project or a whole book or like a massive code 00:10:04.049 --> 00:10:06.230 base in its head and still reason through it 00:10:06.230 --> 00:10:09.470 coherently. It basically gives the AI a truly 00:10:09.470 --> 00:10:12.870 expansive working memory. OK, so much more capable 00:10:12.870 --> 00:10:16.169 for complex tasks. Exactly. Our sources show 00:10:16.169 --> 00:10:18.350 it performs exceptionally well across the board, 00:10:18.490 --> 00:10:21.169 but it really shines in areas like software engineering. 00:10:21.740 --> 00:10:24.879 using tools as an agent and obviously that long 00:10:24.879 --> 00:10:26.840 context reasoning. And how did they do it so 00:10:26.840 --> 00:10:29.360 fast and cheap? Well, what's fascinating here 00:10:29.360 --> 00:10:32.500 is how they achieved it. Minimax used a new reinforcement 00:10:32.500 --> 00:10:34.799 learning method they developed called CISPO. 00:10:34.919 --> 00:10:37.960 Yeah. Essentially, it's just a smarter, more 00:10:37.960 --> 00:10:40.059 efficient way for the AI to learn from trial 00:10:40.059 --> 00:10:43.539 and error. This led to a 2x faster training speed 00:10:43.539 --> 00:10:46.500 without needing, you know, those colossal GPU 00:10:46.500 --> 00:10:49.259 farms or mega budgets. So cleverness over brute 00:10:49.259 --> 00:10:51.899 force. Kind of proved that clever architecture 00:10:51.899 --> 00:10:54.539 and optimization can actually beat sheer compute 00:10:54.539 --> 00:10:57.629 power sometimes, you know. Which is encouraging. 00:10:57.809 --> 00:11:00.009 Definitely. And the implications. Well, this 00:11:00.009 --> 00:11:02.870 kind of long context capability is what serious 00:11:02.870 --> 00:11:05.950 AI agents really need to become practical. Think 00:11:05.950 --> 00:11:07.730 about agents that don't forget what you said 00:11:07.730 --> 00:11:11.029 five prompts ago or agents that can scan hundreds 00:11:11.029 --> 00:11:13.250 of pages of legal documents or medical records 00:11:13.250 --> 00:11:15.750 before accurately answering your questions. Right. 00:11:15.809 --> 00:11:18.750 Handling human scale data. Exactly. It truly 00:11:18.750 --> 00:11:21.309 opens the door to like personalized memory systems, 00:11:21.389 --> 00:11:24.789 much smarter assistants, reliable legal or medical 00:11:24.789 --> 00:11:27.799 AI. This is a. pretty transformative for practical 00:11:27.799 --> 00:11:30.940 applications down the line. So, wow. We've really 00:11:30.940 --> 00:11:32.899 gone on a journey here from the subtle kind of 00:11:32.899 --> 00:11:36.240 unseen ways AI might be like changing our brains. 00:11:36.360 --> 00:11:39.259 The cognitive dead idea. Yeah. To what's buzzing 00:11:39.259 --> 00:11:41.740 in AI tools right now, the good and the concerning, 00:11:41.860 --> 00:11:45.220 all the way to this massive breakthrough in long 00:11:45.220 --> 00:11:48.379 context models with mini Macs. It's a lot. It 00:11:48.379 --> 00:11:51.759 is. It's clear AI is advancing incredibly rapidly 00:11:51.759 --> 00:11:54.519 on so many fronts and understanding its impact, 00:11:54.639 --> 00:11:57.970 both the good and the. potentially challenging 00:11:57.970 --> 00:12:00.549 is just crucial for all of us navigating this 00:12:00.549 --> 00:12:03.009 absolutely ultimately I think this deep dive 00:12:03.009 --> 00:12:05.929 really shows us AI's immense potential but also 00:12:05.929 --> 00:12:10.629 that vital need for intentional, thoughtful engagement. 00:12:11.230 --> 00:12:13.750 It's about how you choose to integrate it into 00:12:13.750 --> 00:12:17.210 your workflow, into your life, not just letting 00:12:17.210 --> 00:12:19.629 it passively take over or accepting its influence 00:12:19.629 --> 00:12:21.909 without question. Right. Being deliberate. Yeah. 00:12:21.990 --> 00:12:24.350 Which raises an important question maybe for 00:12:24.350 --> 00:12:26.809 you listening. How does this information shape 00:12:26.809 --> 00:12:28.789 your approach? How will you think about learning 00:12:28.789 --> 00:12:30.830 and working with AI going forward? Something 00:12:30.830 --> 00:12:33.549 to definitely consider. Exactly. It's not about 00:12:33.549 --> 00:12:35.409 being anti -AI, not at all. It's about being 00:12:35.409 --> 00:12:38.429 smart about how we use these incredibly powerful... 00:12:38.440 --> 00:12:40.700 Powerful tools. Well said. So maybe the next 00:12:40.700 --> 00:12:43.000 time you reach for an AI tool, just pause for 00:12:43.000 --> 00:12:45.639 a second. Ask yourself, is this enhancing my 00:12:45.639 --> 00:12:48.779 thinking or is my brain kind of clocking out 00:12:48.779 --> 00:12:50.600 here? Think about it. We'll catch you on the 00:12:50.600 --> 00:12:51.120 next deep dive.