WEBVTT 00:00:00.000 --> 00:00:02.120 Have you ever wondered if that free AI you're 00:00:02.120 --> 00:00:04.559 hearing about is maybe actually a hidden money 00:00:04.559 --> 00:00:08.259 pit for businesses? Or if a machine could genuinely 00:00:08.259 --> 00:00:11.740 outperform a doctor in training on really complex 00:00:11.740 --> 00:00:14.769 medical stuff? Today, we're taking a deep dive 00:00:14.769 --> 00:00:17.350 into some pretty surprising AI realities. Welcome 00:00:17.350 --> 00:00:20.329 to the deep dive. Yeah, we're here today to unpack 00:00:20.329 --> 00:00:22.969 a fascinating stack of sources. They really challenge 00:00:22.969 --> 00:00:25.809 some of our core assumptions about artificial 00:00:25.809 --> 00:00:27.969 intelligence. That's right. Our mission today 00:00:27.969 --> 00:00:29.929 really is to give you a shortcut to understanding 00:00:29.929 --> 00:00:34.030 AI's rapidly changing landscape. We'll start 00:00:34.030 --> 00:00:36.649 by busting the myth of cheap open source AI. 00:00:36.829 --> 00:00:39.109 It's not always what it seems. Definitely not. 00:00:39.659 --> 00:00:41.939 Then we'll sort of race through a series of rapid 00:00:41.939 --> 00:00:44.200 -fire AI, highlights everything from, believe 00:00:44.200 --> 00:00:47.219 it or not, viral cat videos to some big corporate 00:00:47.219 --> 00:00:49.899 strategy shifts. And finally, we'll grapple with 00:00:49.899 --> 00:00:53.280 a truly groundbreaking study, an AI that, well, 00:00:53.439 --> 00:00:56.560 it just outperformed human medical interns. It's 00:00:56.560 --> 00:00:59.200 quite the journey into AI's surprising new landscape. 00:01:00.520 --> 00:01:03.500 Okay, so let's unpack this first big idea. Many 00:01:03.500 --> 00:01:05.480 people assume open -source AI is always the most 00:01:05.480 --> 00:01:07.500 affordable option. It just feels intuitive, right? 00:01:07.540 --> 00:01:09.790 Free code. It does feel that way but a recent 00:01:09.790 --> 00:01:13.030 study tells a very different story about the 00:01:13.030 --> 00:01:17.079 actual cost. Yeah. Tell us about that. It really 00:01:17.079 --> 00:01:19.159 does. A new NASA research study just dropped, 00:01:19.260 --> 00:01:21.159 and their findings are pretty eye -opening, especially 00:01:21.159 --> 00:01:24.079 for anyone actually running AI systems. They 00:01:24.079 --> 00:01:26.459 found that these so -called free open -weight 00:01:26.459 --> 00:01:30.280 models can actually cost you more in the long 00:01:30.280 --> 00:01:32.480 run than maybe using something like OpenAI's 00:01:32.480 --> 00:01:35.000 APIs. Oh, so? Well, the research shows these 00:01:35.000 --> 00:01:38.019 open -source models can burn through 1 .5 to 00:01:38.019 --> 00:01:41.379 4 times more tokens than their closed counterparts. 00:01:41.900 --> 00:01:43.959 Okay, wait. When you say tokens, what exactly 00:01:43.959 --> 00:01:46.099 are we talking about here? For folks maybe not 00:01:46.099 --> 00:01:48.799 deep in the weeds on this. Good question. So 00:01:48.799 --> 00:01:51.920 tokens are like the small pieces of words or 00:01:51.920 --> 00:01:54.140 data and AI processes. You can think of them 00:01:54.140 --> 00:01:56.459 as tiny building blocks of information the AI 00:01:56.459 --> 00:01:59.579 works with. So more tokens just means more work 00:01:59.579 --> 00:02:01.719 for the AI, more processing power needed. And 00:02:01.719 --> 00:02:04.219 ultimately more cost to you, the user. Right, 00:02:04.239 --> 00:02:07.140 exactly. So if a model uses more tokens for the 00:02:07.140 --> 00:02:09.960 very same task, it's essentially less efficient. 00:02:10.539 --> 00:02:13.080 And the study highlighted this inefficiency even 00:02:13.080 --> 00:02:15.080 further, didn't it? It got pretty extreme in 00:02:15.080 --> 00:02:18.479 some cases. Oh, absolutely. For really simple 00:02:18.479 --> 00:02:21.139 Q &A tasks, some of these open source models 00:02:21.139 --> 00:02:25.379 used a shocking 10 times more tokens. Imagine 00:02:25.379 --> 00:02:27.819 asking, you know, what's the capital of Australia? 00:02:28.139 --> 00:02:31.120 And the AI basically writes a short novel to 00:02:31.120 --> 00:02:34.439 give you Canberra. It's overkill. Wow. That's 00:02:34.439 --> 00:02:37.849 significant. Yeah. Meanwhile, closed models like 00:02:37.849 --> 00:02:41.050 OpenAI's 04 Mini, they demonstrated superior 00:02:41.050 --> 00:02:44.930 efficiency, particularly with complex tasks like 00:02:44.930 --> 00:02:47.810 math, where their internal reasoning seems much 00:02:47.810 --> 00:02:49.770 more compressed. That's a huge difference in 00:02:49.770 --> 00:02:51.150 efficiency. And you mentioned something like, 00:02:51.189 --> 00:02:53.310 what, 51 % of companies are already running AI 00:02:53.310 --> 00:02:56.849 in production. This kind of inefficiency quickly 00:02:56.849 --> 00:02:59.629 translates into runaway compute costs. It really 00:02:59.629 --> 00:03:02.110 highlights how that per token pricing we see 00:03:02.110 --> 00:03:04.729 advertised can be, well, pretty deceiving if 00:03:04.729 --> 00:03:06.689 the model just eats way more tokens to get the 00:03:06.689 --> 00:03:09.849 job done. Precisely. Your seemingly cheaper open 00:03:09.849 --> 00:03:13.030 source option could quietly, you know, devastate 00:03:13.030 --> 00:03:14.810 your compute budget if you're not paying close 00:03:14.810 --> 00:03:17.009 attention to its actual token consumption. So 00:03:17.009 --> 00:03:19.189 why is this happening? What's the fundamental 00:03:19.189 --> 00:03:21.469 difference in how these models reason that leads 00:03:21.469 --> 00:03:23.960 to such a big gap? Well, it seems to come down 00:03:23.960 --> 00:03:26.060 to their architecture, maybe their training philosophy. 00:03:26.580 --> 00:03:30.659 Closed source providers are internally compressing 00:03:30.659 --> 00:03:33.280 their reasoning pathways. They've really optimized 00:03:33.280 --> 00:03:36.539 their models to perform complex tasks with fewer 00:03:36.539 --> 00:03:39.500 internal thinking steps, basically. That shrinks 00:03:39.500 --> 00:03:42.120 the token count significantly. Open source developers, 00:03:42.319 --> 00:03:44.580 on the other hand, often extend these reasoning 00:03:44.580 --> 00:03:47.460 chains. They might add more explicit step -by 00:03:47.460 --> 00:03:51.319 -step thinking, maybe for accuracy or robustness, 00:03:51.319 --> 00:03:53.800 perhaps. to cover more edge cases. Well, that 00:03:53.800 --> 00:03:55.919 means more tokens. But that means more tokens 00:03:55.919 --> 00:03:58.439 and inevitably more cost. It's a clear trade 00:03:58.439 --> 00:04:00.240 -off. So it's not just about getting the right 00:04:00.240 --> 00:04:02.699 answer anymore. It's becoming about how efficiently 00:04:02.699 --> 00:04:05.460 the AI gets to that answer. Exactly. Token discipline. 00:04:05.639 --> 00:04:07.860 It isn't just some technical jargon anymore. 00:04:07.979 --> 00:04:10.360 It's actually becoming crucial for managing your 00:04:10.360 --> 00:04:13.539 budget and, frankly, making AI deployment sustainable 00:04:13.539 --> 00:04:16.389 long -term. So for businesses out there trying 00:04:16.389 --> 00:04:19.949 to save money with AI, what's the real key takeaway 00:04:19.949 --> 00:04:22.790 here, the core message? The key isn't just the 00:04:22.790 --> 00:04:25.170 per token price. It's the total token efficiency 00:04:25.170 --> 00:04:31.100 that matters. Okay, let's shift gears now. Let's 00:04:31.100 --> 00:04:34.120 talk about some of the most talked about AI happenings 00:04:34.120 --> 00:04:36.620 right now. Kind of a wild mix, really showcasing 00:04:36.620 --> 00:04:39.300 how AI is just popping up everywhere. It really 00:04:39.300 --> 00:04:42.759 is. We recently saw an AI builder spark this 00:04:42.759 --> 00:04:45.439 online challenge, right? Yeah. Inviting all AI 00:04:45.439 --> 00:04:47.620 creators to share their coolest AI -made art, 00:04:47.720 --> 00:04:50.579 videos, tools. Yeah, that was neat. And the result 00:04:50.579 --> 00:04:52.860 was this incredible sort of crowdsourced stream 00:04:52.860 --> 00:04:56.019 of creativity. It really shows how AI is maybe... 00:04:56.769 --> 00:04:58.990 democratizing artistic creation in a way. And 00:04:58.990 --> 00:05:00.569 speaking of access, if you're looking to get 00:05:00.569 --> 00:05:02.730 into that, Harvard University is offering 12 00:05:02.730 --> 00:05:05.769 free online AI courses in 2025. That definitely 00:05:05.769 --> 00:05:08.350 widens access to some pretty critical knowledge. 00:05:08.550 --> 00:05:11.670 Okay, now here's where it gets truly interesting. 00:05:11.810 --> 00:05:14.269 Yeah. Maybe a little weird. AI CAD videos have 00:05:14.269 --> 00:05:16.569 gone super viral. Oh, I saw some of these. And 00:05:16.569 --> 00:05:18.910 they're surprisingly bizarre. We're talking like... 00:05:19.019 --> 00:05:22.180 buff cats on revenge missions or uh billy eilish 00:05:22.180 --> 00:05:24.899 meows dubbed into 30 seconds soap operas millions 00:05:24.899 --> 00:05:27.600 are genuinely addicted it's like a whole new 00:05:27.600 --> 00:05:31.459 genre of internet culture just spawned beat whoa 00:05:31.459 --> 00:05:33.839 the sophistication in these little narratives 00:05:33.839 --> 00:05:36.240 is kind of wild makes you realize how far ai 00:05:36.240 --> 00:05:38.810 generation has come even if You know, many of 00:05:38.810 --> 00:05:40.870 us still wrestle with prompt drift, just trying 00:05:40.870 --> 00:05:43.170 to get simpler outputs. Yeah, it's a truly unique 00:05:43.170 --> 00:05:46.069 corner of the Internet, that's for sure. On a 00:05:46.069 --> 00:05:48.170 more practical note, Genspark AI just launched 00:05:48.170 --> 00:05:50.069 something called AI Developer, which is like 00:05:50.069 --> 00:05:53.009 a vibe coding tool. Apparently, one user built 00:05:53.009 --> 00:05:56.430 a working Mario game in just five prompts. Five 00:05:56.430 --> 00:05:59.310 prompts. Wow. Talk about rapid prototyping. That 00:05:59.310 --> 00:06:01.310 really lowers the barrier to entry for coding, 00:06:01.430 --> 00:06:03.930 doesn't it? Absolutely. We also saw some more 00:06:03.930 --> 00:06:06.329 controversial news recently, though. Ignite Tech 00:06:06.329 --> 00:06:09.279 CEO. He controversially cut 80 percent of staff 00:06:09.279 --> 00:06:11.720 who resisted AI adoption. His quote was something 00:06:11.720 --> 00:06:15.100 like belief was harder than skills. Oof. They 00:06:15.100 --> 00:06:18.819 even reportedly instituted AI only Mondays. It 00:06:18.819 --> 00:06:20.540 kind of highlights the intense pressure some 00:06:20.540 --> 00:06:23.240 companies feel to integrate AI really fast, sometimes 00:06:23.240 --> 00:06:26.540 at the cost of human jobs. But then on the flip 00:06:26.540 --> 00:06:29.899 side, Duolingo's CEO clarified that their AI 00:06:29.899 --> 00:06:34.019 first memo was widely misunderstood. He stated 00:06:34.019 --> 00:06:37.319 no full -time staff lost jobs due to AI, and 00:06:37.319 --> 00:06:40.100 their AI Fridays are now just a weekly internal 00:06:40.100 --> 00:06:42.980 thing for exploration. Okay, so maybe more of 00:06:42.980 --> 00:06:45.560 a PR misstep there than a policy shift. Sounds 00:06:45.560 --> 00:06:47.779 like it. An important clarification about their 00:06:47.779 --> 00:06:50.100 approach, though. And finally, on the investment 00:06:50.100 --> 00:06:52.540 front, the U .S. government, alongside NVIDIA, 00:06:53.100 --> 00:06:55.899 is investing a pretty significant amount, $152 00:06:55.899 --> 00:06:58.879 million, into building open -source AI models 00:06:58.879 --> 00:07:01.899 specifically for science. That's good to see. 00:07:02.079 --> 00:07:03.680 Yeah, the idea is to help universities catch 00:07:03.680 --> 00:07:06.800 up, especially as the costs for private, cutting 00:07:06.800 --> 00:07:09.819 -edge AI models continue to soar. So thinking 00:07:09.819 --> 00:07:11.680 about all these incredibly diverse headlines, 00:07:12.060 --> 00:07:14.439 what's the biggest takeaway here? What connects 00:07:14.439 --> 00:07:17.120 them? I think it's that AI isn't just a tech 00:07:17.120 --> 00:07:19.860 trend anymore. It's fundamentally reshaping culture, 00:07:20.000 --> 00:07:22.949 business, and education. OK, now let's zip through 00:07:22.949 --> 00:07:25.250 some quick, intriguing insights from the AI world. 00:07:25.430 --> 00:07:27.689 These really reveal how we're starting to interact 00:07:27.689 --> 00:07:29.449 with it, maybe without even noticing sometimes 00:07:29.449 --> 00:07:32.910 day to day. All right. Rapid fire. PwC University 00:07:32.910 --> 00:07:35.410 apparently offers five distinct strategies to 00:07:35.410 --> 00:07:37.829 help avoid what they're calling AI paralysis. 00:07:38.709 --> 00:07:42.670 Essentially, how not to get. killed by AI, in 00:07:42.670 --> 00:07:45.610 their words. Character AI is betting big on these 00:07:45.610 --> 00:07:47.670 persona -based AIs. You know, the idea is to 00:07:47.670 --> 00:07:50.209 give you a personalized bestie for conversation, 00:07:50.550 --> 00:07:53.649 companionship, maybe more. Interesting. On a 00:07:53.649 --> 00:07:56.490 more scientific note, AI is now actually designing 00:07:56.490 --> 00:07:59.149 bizarre new physics experiments that... get this, 00:07:59.290 --> 00:08:03.410 actually work. Really? That's wild. Yeah. Google's 00:08:03.410 --> 00:08:05.610 former AI lead offered an interesting perspective, 00:08:05.670 --> 00:08:08.009 too, saying it's basically too late now to get 00:08:08.009 --> 00:08:11.209 a PhD specifically for the AI boom timing. Right. 00:08:11.269 --> 00:08:13.889 Yeah. And finally, Gemini Canvas now lets anyone 00:08:13.889 --> 00:08:15.850 tweak app designs using just simple descriptive 00:08:15.850 --> 00:08:18.889 words like make the button blue. So putting these 00:08:18.889 --> 00:08:20.790 snapshots together, what do they reveal about 00:08:20.790 --> 00:08:22.949 how AI is kind of weaving itself into our daily 00:08:22.949 --> 00:08:25.689 lives and workflows? I'd say AI is fast becoming 00:08:25.689 --> 00:08:28.310 a creative partner and just a seamless day. tool 00:08:28.310 --> 00:08:30.970 for many people. All right. This next deep dive 00:08:30.970 --> 00:08:35.350 is, well, it's truly remarkable. Maybe bordering 00:08:35.350 --> 00:08:39.289 on astounding, frankly. GPT -5 has demonstrably 00:08:39.289 --> 00:08:42.289 outperformed human medical interns in specific 00:08:42.289 --> 00:08:45.610 diagnostic tasks. This isn't just like a small 00:08:45.610 --> 00:08:47.870 step forward. It feels like a significant leap. 00:08:48.110 --> 00:08:50.769 Yeah, this comes from a new study by Emory University's 00:08:50.769 --> 00:08:53.409 radiation oncology team. And they really put 00:08:53.409 --> 00:08:56.049 GPT -5 through its paces. They pitted it against 00:08:56.049 --> 00:08:59.990 earlier AI models like GPT -4 -0 and actual human 00:08:59.990 --> 00:09:03.210 medical interns. And crucially, GPT -5 wasn't 00:09:03.210 --> 00:09:05.769 fed exam answers or pre -digested information. 00:09:06.070 --> 00:09:09.009 It was a pure test of its analytical and reasoning 00:09:09.009 --> 00:09:11.870 capabilities in a real diagnostic context or 00:09:11.870 --> 00:09:14.690 simulated anyway. And how exactly did they test 00:09:14.690 --> 00:09:17.480 its reasoning? What kind of prompts? were involved. 00:09:17.620 --> 00:09:19.440 That seems key. They use something called zero 00:09:19.440 --> 00:09:21.639 -shot chain of thought. Which is a bit technical, 00:09:21.740 --> 00:09:23.980 but basically means the AI thinks step by step 00:09:23.980 --> 00:09:26.659 to reach an answer without needing specific examples 00:09:26.659 --> 00:09:29.139 or training on those exact problems first. It's 00:09:29.139 --> 00:09:30.620 kind of like it figures out the logical path 00:09:30.620 --> 00:09:33.059 on its own, showing its internal thinking process. 00:09:33.659 --> 00:09:36.000 Okay, so it's reasoning from first principles, 00:09:36.100 --> 00:09:37.759 essentially. And they gave it some incredibly 00:09:37.759 --> 00:09:40.179 complex scenarios too, right? Demanding more 00:09:40.179 --> 00:09:42.399 than just crunching text. Oh, absolutely. They 00:09:42.399 --> 00:09:45.080 used multimodal prompts. Now that means they 00:09:45.080 --> 00:09:48.240 combined patient history, so the text, with actual 00:09:48.240 --> 00:09:50.720 medical images. Things like... like CT scans, 00:09:51.080 --> 00:09:54.899 MRIs, or x -rays. GPT -5 had to understand both 00:09:54.899 --> 00:09:57.620 the visual data and the textual context, then 00:09:57.620 --> 00:09:59.720 connect the dots to make a diagnosis. That's 00:09:59.720 --> 00:10:02.360 a very human -like, high -level task. And the 00:10:02.360 --> 00:10:05.590 results were, well, quite definitive. Probably 00:10:05.590 --> 00:10:07.649 surprising for many in the medical field, I'd 00:10:07.649 --> 00:10:10.429 imagine. Yeah, GPT -5 crushed it. On these multimodal 00:10:10.429 --> 00:10:13.149 reasoning tasks, it showed a nearly 30 % gain 00:10:13.149 --> 00:10:15.809 in logic and over 36 % gain in understanding 00:10:15.809 --> 00:10:18.769 compared to GPT -4. That's a huge jump between 00:10:18.769 --> 00:10:22.049 versions. And where GPT -4 actually lagged behind 00:10:22.049 --> 00:10:25.129 the interns by about 5 % to 15 % on these specific 00:10:25.129 --> 00:10:28.970 tasks, GPT -5 surged ahead by over 24%. It's 00:10:28.970 --> 00:10:31.429 a significant quantifiable leap in its diagnostic 00:10:31.429 --> 00:10:33.350 capabilities in this controlled setting. But 00:10:33.350 --> 00:10:34.850 there's always a catch, isn't there? A but. But 00:10:34.850 --> 00:10:37.190 this was all in ideal lab settings. Real hospitals 00:10:37.190 --> 00:10:40.129 are just incredibly messy. Very different environment. 00:10:40.450 --> 00:10:42.610 Yeah. You've got incomplete records, the complexities 00:10:42.610 --> 00:10:45.850 of human emotion, huge ethical concerns, legal 00:10:45.850 --> 00:10:48.730 constraints everywhere. And AI might ace an exam 00:10:48.730 --> 00:10:52.289 like this, but building bedside trust with patients, 00:10:52.509 --> 00:10:54.230 that's a completely different challenge. Two 00:10:54.230 --> 00:10:56.509 sec silence. Honestly, I still wrestle with the 00:10:56.509 --> 00:10:59.529 concept of fully trusting AI in critical life 00:10:59.529 --> 00:11:01.909 and death scenarios myself. It's a huge leap 00:11:01.909 --> 00:11:04.210 from the lab to the actual clinic floor. Right. 00:11:04.309 --> 00:11:07.529 And if GPT -4 was maybe like a helpful but still 00:11:07.529 --> 00:11:10.830 learning med student in these tests, GPT -5 is 00:11:10.830 --> 00:11:12.970 certainly performing at the level of an attending 00:11:12.970 --> 00:11:15.210 physician, maybe even beyond, on these specific 00:11:15.210 --> 00:11:18.230 reasoning tasks. The capabilities showcased here 00:11:18.230 --> 00:11:20.669 are truly astounding. They really push the boundaries 00:11:20.669 --> 00:11:22.990 of what we thought AI could do autonomously in 00:11:22.990 --> 00:11:25.610 diagnostics. Yeah, this isn't just about technical 00:11:25.610 --> 00:11:27.970 accuracy. It brings up huge questions about regulations, 00:11:28.350 --> 00:11:30.409 liability, how existing medical workflows would 00:11:30.409 --> 00:11:32.730 even adapt. So the implication seems clear, at 00:11:32.730 --> 00:11:35.309 least from this study. In specific, complex tasks 00:11:35.309 --> 00:11:37.950 involving both text and image data from patients, 00:11:38.330 --> 00:11:42.230 GPT -5 is demonstrably beyond doctor level, outperforming 00:11:42.230 --> 00:11:44.649 humans in multimodal reasoning within that test 00:11:44.649 --> 00:11:47.279 environment. It's pretty staggering. So what 00:11:47.279 --> 00:11:50.059 do you think is the biggest hurdle for AI moving 00:11:50.059 --> 00:11:53.120 from these amazing lab results to actually being 00:11:53.120 --> 00:11:56.299 used widely in real world hospitals? Oh, it's 00:11:56.299 --> 00:11:58.659 got to be trust, dealing with human complexity, 00:11:58.799 --> 00:12:02.059 and just adapting to all that real world chaos. 00:12:02.399 --> 00:12:05.179 So wrapping this all up, what does this all mean 00:12:05.179 --> 00:12:08.399 for us? We've seen that AI's promises, and maybe 00:12:08.399 --> 00:12:11.179 its pitfalls, are often far more complex than 00:12:11.179 --> 00:12:13.399 they might initially appear. From those hidden 00:12:13.399 --> 00:12:15.559 costs that can really impact a company's bottom 00:12:15.559 --> 00:12:18.299 line to these truly groundbreaking capabilities 00:12:18.299 --> 00:12:20.960 emerging in fields like medicine. Yeah, it seems 00:12:20.960 --> 00:12:23.399 like efficiency, ethical application, and just 00:12:23.399 --> 00:12:26.759 having a nuanced understanding of AI's true strengths 00:12:26.759 --> 00:12:29.320 and weaknesses, those are becoming absolutely 00:12:29.320 --> 00:12:31.639 paramount for navigating this whole evolving 00:12:31.639 --> 00:12:34.059 landscape. This deep dive really just scratched 00:12:34.059 --> 00:12:36.220 the surface of our sources today. It really did. 00:12:36.360 --> 00:12:38.259 We definitely encourage you to keep exploring 00:12:38.259 --> 00:12:40.100 these fascinating shifts. You know, ask yourself, 00:12:40.279 --> 00:12:43.120 how might AI's efficiency gains affect your work? 00:12:43.220 --> 00:12:45.980 And where will we see AI's beyond human abilities 00:12:45.980 --> 00:12:48.659 maybe emerge next outside of the lab? And here's 00:12:48.659 --> 00:12:52.639 something to maybe ponder. If AI can now demonstrably 00:12:52.639 --> 00:12:56.139 out -diagnose human interns in specific, complex 00:12:56.139 --> 00:12:59.700 tasks. What truly fundamental human skills will 00:12:59.700 --> 00:13:02.159 remain irreplaceable in an increasingly AI -driven 00:13:02.159 --> 00:13:04.000 world? That's the big question, isn't it? Thank 00:13:04.000 --> 00:13:05.360 you for joining us on this deep dive.