WEBVTT 00:00:00.000 --> 00:00:04.440 We often think that AI progress, it just means 00:00:04.440 --> 00:00:06.799 going bigger, right? More scale. It's kind of 00:00:06.799 --> 00:00:09.640 the default assumption. Bigger models, massive 00:00:09.640 --> 00:00:12.240 compute, huge data centers. That's what gets 00:00:12.240 --> 00:00:14.699 all the attention, the brute force method. But 00:00:14.699 --> 00:00:16.980 what if the really big step forward, the next 00:00:16.980 --> 00:00:20.179 one, comes from something, well, tiny? architecturally 00:00:20.179 --> 00:00:23.280 smart we've got some data showing a small recursive 00:00:23.280 --> 00:00:26.179 ai that's actually out reasoning the giants the 00:00:26.179 --> 00:00:28.859 trillion parameter models on really tough benchmarks 00:00:28.859 --> 00:00:31.100 welcome to the deep dive yeah yeah you sent over 00:00:31.100 --> 00:00:32.820 some truly fascinating sources this week it's 00:00:32.820 --> 00:00:35.000 like ai is splitting going down two very different 00:00:35.000 --> 00:00:36.979 roads at the same time oh it's our job today 00:00:36.979 --> 00:00:40.079 clarity let's try and separate that uh massive 00:00:40.079 --> 00:00:43.000 computation hype from the real tangible algorithmic 00:00:43.000 --> 00:00:45.320 intelligence that's actually popping up it seems 00:00:45.320 --> 00:00:47.320 smart design might just be starting to beat sheer 00:00:47.320 --> 00:00:49.710 muscle Okay, so here's the roadmap for this deep 00:00:49.710 --> 00:00:51.770 dive, just for you. First, we're going to unpack 00:00:51.770 --> 00:00:54.750 this quiet revolution of these tiny efficient 00:00:54.750 --> 00:00:57.880 models, TRM they're calling it. why they're winning 00:00:57.880 --> 00:00:59.960 on abstract reasoning then second we'll look 00:00:59.960 --> 00:01:03.039 at ai agents the age of agents seems to be upon 00:01:03.039 --> 00:01:05.079 us we'll dive into how they're being integrated 00:01:05.079 --> 00:01:08.340 the sort of market wars and also the societal 00:01:08.340 --> 00:01:11.760 friction lawsuits jobs that kind of thing and 00:01:11.760 --> 00:01:14.700 finally the story of dr cabot it's pretty shocking 00:01:14.700 --> 00:01:17.040 actually the first ai getting a diagnosis published 00:01:17.040 --> 00:01:19.180 in a top medical journal a huge professional 00:01:19.180 --> 00:01:21.939 step so let's get into this divergence in the 00:01:21.939 --> 00:01:24.200 ai landscape right so for the last what five 00:01:24.200 --> 00:01:27.450 years maybe the mantra in ai It's just been scale. 00:01:27.549 --> 00:01:29.430 Scale solves everything. Your LLM isn't good 00:01:29.430 --> 00:01:31.510 enough. Fine. Throw more data, more parameters, 00:01:31.609 --> 00:01:34.269 more GPUs at it. Just make it bigger. But now 00:01:34.269 --> 00:01:36.430 Samsung seems to be really challenging that idea 00:01:36.430 --> 00:01:40.049 with this tiny TRM model. And when we say tiny. 00:01:40.810 --> 00:01:42.730 We mean compared to those trillion parameter 00:01:42.730 --> 00:01:45.269 monsters. You know, those things are like enormous 00:01:45.269 --> 00:01:48.069 computer brains needing billions of dollars and 00:01:48.069 --> 00:01:51.370 just mountains of data to build. Right. But this 00:01:51.370 --> 00:01:54.310 TRM, it's going head to head with some pretty 00:01:54.310 --> 00:01:57.170 serious models, like even the older Open AIO 00:01:57.170 --> 00:02:00.069 3 Mini, which is still a solid baseline, and 00:02:00.069 --> 00:02:03.310 Google's newest Gemini 2 .5 Pro on complex reasoning 00:02:03.310 --> 00:02:07.030 tasks. And the kicker, it's winning. Yeah. And 00:02:07.030 --> 00:02:09.550 the concept underneath it all is it's really 00:02:09.550 --> 00:02:11.669 compelling because it shifts focus away from 00:02:11.669 --> 00:02:14.330 just raw data crunching towards like structural 00:02:14.330 --> 00:02:17.270 intelligence. Instead of one giant pass through 00:02:17.270 --> 00:02:19.689 the data to get an answer, TRM loops. It goes 00:02:19.689 --> 00:02:21.449 over its own initial answer again and again, 00:02:21.550 --> 00:02:23.030 recursively. You can think of it like a really 00:02:23.030 --> 00:02:24.770 careful student doing a hard math problem. They 00:02:24.770 --> 00:02:26.770 get an answer. Right. But instead of just stopping, 00:02:26.889 --> 00:02:29.210 they spend maybe five minutes just rechecking 00:02:29.210 --> 00:02:31.469 every step, refining that first output until 00:02:31.469 --> 00:02:33.449 they're really, really sure about the logic. 00:02:33.789 --> 00:02:36.270 Exactly. It's self -refinement built in. People 00:02:36.270 --> 00:02:37.949 have tried getting LLMs to do this externally, 00:02:38.090 --> 00:02:39.770 right? Like with chain of thought prompting, 00:02:39.770 --> 00:02:42.689 making it right out its steps. But TRM builds 00:02:42.689 --> 00:02:45.689 that recursive checking right into its core design. 00:02:46.110 --> 00:02:49.370 It doesn't need some massive, expensive prompt 00:02:49.370 --> 00:02:52.550 to get that deep reasoning ability. And the proof 00:02:52.550 --> 00:02:54.710 is pretty striking, especially on abstract tasks, 00:02:54.849 --> 00:02:57.689 the kind designed to really push LLMs past just 00:02:57.689 --> 00:03:00.849 remembering facts. Take Sudoku Extreme. Okay. 00:03:00.969 --> 00:03:03.810 That's famously hard. It needs serious logic, 00:03:03.990 --> 00:03:07.830 constraint satisfaction. TRM got 87 .4 % accuracy. 00:03:08.050 --> 00:03:11.259 The bigger models, around 55%. Wow. Big difference. 00:03:11.419 --> 00:03:14.120 Huge. And on maze -hard problems, testing navigation 00:03:14.120 --> 00:03:18.060 logic, TRM hit 85%. Again, just beating the larger 00:03:18.060 --> 00:03:20.099 models pretty easily. So that's the trade -off 00:03:20.099 --> 00:03:22.180 then, isn't it? TRM isn't going to, like, summarize 00:03:22.180 --> 00:03:24.219 your emails or chat about the news. It won't 00:03:24.219 --> 00:03:26.539 replace a generalist like GPT -4. But its thing 00:03:26.539 --> 00:03:29.520 is deep, abstract, computational reasoning. Yeah, 00:03:29.520 --> 00:03:32.860 it sacrifices that broad, general knowledge for 00:03:32.860 --> 00:03:36.639 being incredibly efficient at one type of problem 00:03:36.639 --> 00:03:39.419 solving. And it proves there's another way forward. 00:03:40.110 --> 00:03:42.150 that doesn't automatically mean you need a billion 00:03:42.150 --> 00:03:44.870 dollar GPU farm to get breakthrough reasoning. 00:03:45.250 --> 00:03:48.629 Whoa. I mean, imagine scaling that. That kind 00:03:48.629 --> 00:03:51.389 of precise self -correcting logic for specialized 00:03:51.389 --> 00:03:54.110 industrial stuff without needing, you know, a 00:03:54.110 --> 00:03:56.409 billion queries or a giant carbon footprint. 00:03:56.689 --> 00:03:59.090 That efficiency. That's actually kind of amazing. 00:03:59.289 --> 00:04:02.069 Points to the future, maybe. Okay, but hang on. 00:04:02.189 --> 00:04:04.569 If this TRM is so much better on these complex 00:04:04.569 --> 00:04:07.289 reasoning benchmarks, why isn't it, like, everywhere? 00:04:07.430 --> 00:04:09.400 What's the practical catch right now? Well, the 00:04:09.400 --> 00:04:11.319 catch is that specialization. It's laser focused. 00:04:11.400 --> 00:04:13.599 It's not built for chatting or writing emails. 00:04:13.780 --> 00:04:15.819 Think of it as a research win, proving smart 00:04:15.819 --> 00:04:18.800 algorithms can beat raw power. OK, so we've talked 00:04:18.800 --> 00:04:20.759 about size versus smarts in the models themselves. 00:04:21.000 --> 00:04:23.220 Now let's pivot. Let's talk application integration. 00:04:23.399 --> 00:04:25.560 The real world. This is really where the rubber 00:04:25.560 --> 00:04:27.360 meets the road, where AI meets the spreadsheet 00:04:27.360 --> 00:04:30.480 and where the friction really starts. Yeah, the 00:04:30.480 --> 00:04:32.879 agentic future isn't some far off theory anymore. 00:04:33.000 --> 00:04:36.360 It's about immediate utility. We're past just 00:04:36.360 --> 00:04:38.339 single prompts now. We're talking autonomous 00:04:38.339 --> 00:04:41.180 systems doing tasks. I mean, look at OpenAI's 00:04:41.180 --> 00:04:43.639 agent builder. It apparently already has like 00:04:43.639 --> 00:04:46.720 50 real use cases, boosting productivity and 00:04:46.720 --> 00:04:49.459 sales marketing operations almost everywhere. 00:04:49.639 --> 00:04:52.660 And just for fun, showing how mainstream agents 00:04:52.660 --> 00:04:56.660 are becoming. Mark your calendars. October 27th, 00:04:56.660 --> 00:04:58.860 there's going to be this hilarious AI poker showdown. 00:04:59.319 --> 00:05:02.420 ChatGPT versus Claude versus Grok versus DeepSeek. 00:05:03.180 --> 00:05:05.079 Poker playing agents. It's like the new competitive 00:05:05.079 --> 00:05:08.459 sport. OK, noted. But yeah, the integration wars, 00:05:08.579 --> 00:05:10.920 they are absolutely heating up. Google just launched 00:05:10.920 --> 00:05:13.560 Gemini CLI extensions that the command line interface 00:05:13.560 --> 00:05:16.519 literally days after OpenAI showed off their 00:05:16.519 --> 00:05:18.860 new app idea. And the key thing here, what's 00:05:18.860 --> 00:05:20.660 really crucial is these extensions let anyone 00:05:20.660 --> 00:05:22.980 publish tools. You can plug in huge platforms, 00:05:23.180 --> 00:05:26.100 Figma for design, Stripe for payments. The AI 00:05:26.100 --> 00:05:27.740 isn't just helping you anymore. It's becoming 00:05:27.740 --> 00:05:29.779 the hub, kind of like the new operating system 00:05:29.779 --> 00:05:31.779 for all your other digital tools. It's a vision 00:05:31.779 --> 00:05:33.199 people have talked about for ages, actually. 00:05:33.519 --> 00:05:35.759 There's this funny historical echo going around 00:05:35.759 --> 00:05:38.720 social media now. Someone dug up a 1985 video, 00:05:39.019 --> 00:05:42.500 Steve Jobs. And he seems to be predicting tools 00:05:42.500 --> 00:05:44.920 exactly like ChatGPT, something that could automate 00:05:44.920 --> 00:05:47.040 using other tools. Kind of wild to see that finally 00:05:47.040 --> 00:05:49.360 happen. Yeah, it is. But integrating this stuff. 00:05:49.949 --> 00:05:52.730 It comes with real tension, complex tension right 00:05:52.730 --> 00:05:55.829 now. On one side, you've got resistance, major 00:05:55.829 --> 00:05:59.949 lawsuits like 17 authors are suing OpenAI right 00:05:59.949 --> 00:06:02.069 now over using copyrighted books for training. 00:06:02.149 --> 00:06:05.149 And you hear big creators like Mr. Beast saying 00:06:05.149 --> 00:06:08.350 publicly that AI means, quote, scary times for 00:06:08.350 --> 00:06:10.149 YouTubers. They're worried about being drowned 00:06:10.149 --> 00:06:13.089 out or replaced. And at the exact same time, 00:06:13.209 --> 00:06:16.189 adoption is just. Warp speed, creating this huge 00:06:16.189 --> 00:06:18.430 industrial split. IBM teams up with Anthropic, 00:06:18.470 --> 00:06:20.629 putting Claude models into their business stuff. 00:06:20.970 --> 00:06:23.889 Deloitte deployed Claude AI to half a million 00:06:23.889 --> 00:06:26.449 employees globally. Half a million. Wow. Yeah, 00:06:26.529 --> 00:06:29.050 500 ,000 people using it internally. But that 00:06:29.050 --> 00:06:31.689 efficiency, it has consequences we really need 00:06:31.689 --> 00:06:33.990 to look at. We're seeing actual job displacement. 00:06:34.290 --> 00:06:36.389 Equiture, an insurance broker, is cutting 400 00:06:36.389 --> 00:06:39.430 jobs. Why? AI is automating accounting and operations. 00:06:40.089 --> 00:06:42.970 And this integration, even for the people building 00:06:42.970 --> 00:06:45.750 it, it's not always smooth sailing, is it? Taking 00:06:45.750 --> 00:06:48.310 an agent from a cool idea to something stable 00:06:48.310 --> 00:06:51.430 and useful, that's tricky work. Oh, absolutely. 00:06:51.589 --> 00:06:54.350 It's not magic yet. I mean, I still wrestle with 00:06:54.350 --> 00:06:56.209 prompt drift myself when I'm trying to build 00:06:56.209 --> 00:06:58.769 complex agents. Yeah. You set up a sequence, 00:06:58.970 --> 00:07:01.769 right? And like three steps in, the system just 00:07:01.769 --> 00:07:04.689 kind of wanders off or misinterprets the data 00:07:04.689 --> 00:07:07.129 it just made. It needs constant fiddling, constant 00:07:07.129 --> 00:07:10.420 tuning. So given all the... The lawsuits, the 00:07:10.420 --> 00:07:13.699 creator anxiety, the actual job cuts. How do 00:07:13.699 --> 00:07:17.040 we balance this massive tech acceleration with 00:07:17.040 --> 00:07:19.399 these really serious societal risks we're seeing? 00:07:19.600 --> 00:07:21.660 Yeah, that's the core issue. Integration really 00:07:21.660 --> 00:07:23.699 needs careful management, clear policies. We 00:07:23.699 --> 00:07:26.319 need to focus on job transitions and figure out 00:07:26.319 --> 00:07:28.920 fair copyright rules basically as fast as the 00:07:28.920 --> 00:07:30.689 tech itself is changing. All right, let's shift 00:07:30.689 --> 00:07:33.430 to maybe the highest stakes area of all, medicine. 00:07:33.670 --> 00:07:35.829 This source material you sent talks about an 00:07:35.829 --> 00:07:38.589 AI hitting a major professional milestone, something 00:07:38.589 --> 00:07:40.870 that absolutely demands transparency and accountability. 00:07:41.269 --> 00:07:43.569 All right, that's Harvard's Dr. Cabot and the 00:07:43.569 --> 00:07:47.629 milestone. It's the first AI system ever to publish 00:07:47.629 --> 00:07:50.350 a diagnosis in the New England Journal of Medicine, 00:07:50.610 --> 00:07:56.199 the NEJM. Just for context, that journal, it's 00:07:56.199 --> 00:07:59.060 arguably the top medical journal globally. It's 00:07:59.060 --> 00:08:01.399 where the absolute best human doctors debate 00:08:01.399 --> 00:08:04.079 the toughest, most complex medical cases. And 00:08:04.079 --> 00:08:06.639 this AI, Dr. Cabot, it doesn't just, you know, 00:08:06.639 --> 00:08:08.500 spit out an answer. That's always the worry with 00:08:08.500 --> 00:08:11.899 LLMs, right? But this thing, crucially, it simulates 00:08:11.899 --> 00:08:16.060 how a doctor actually thinks. So you feed it 00:08:16.060 --> 00:08:18.680 a complex case, symptoms, patient history, lab 00:08:18.680 --> 00:08:20.920 results, the whole picture, and then the system 00:08:20.920 --> 00:08:23.220 goes to work. Okay. It builds a full slide deck 00:08:23.220 --> 00:08:25.120 and talks you through its reasoning audibly. 00:08:25.259 --> 00:08:27.439 It lays out all the possibilities, the differential 00:08:27.439 --> 00:08:30.000 diagnosis, doctors call it. It methodically rules 00:08:30.000 --> 00:08:32.279 out the red herrings, the misleading clues, and 00:08:32.279 --> 00:08:34.840 it backs up every claim by citing actual clinical 00:08:34.840 --> 00:08:37.320 papers. And it does all of that in about five 00:08:37.320 --> 00:08:39.360 minutes? Five minutes, yeah. Wow. It's apparently 00:08:39.360 --> 00:08:42.740 powered by OpenAI's O3 model. but specialized, 00:08:43.059 --> 00:08:46.820 trained on over 100 years of very specific cases 00:08:46.820 --> 00:08:49.139 from Mass General, their CPC cases, clinical 00:08:49.139 --> 00:08:51.580 pathological conferences. So it has this incredibly 00:08:51.580 --> 00:08:54.139 deep historical knowledge base, probably more 00:08:54.139 --> 00:08:56.220 than any single human could have. And here's 00:08:56.220 --> 00:08:58.500 a detail that's, oh, it's kind of freaky, honestly. 00:08:58.659 --> 00:09:01.460 The source material says it even uses human -like 00:09:01.460 --> 00:09:03.620 filler words when it presents its case, like, 00:09:03.639 --> 00:09:07.029 ugh, and, you know. Hmm. Okay, wait. Is that 00:09:07.029 --> 00:09:09.250 actually useful or is it just window dressing? 00:09:09.450 --> 00:09:12.450 Like, are those A's helping the diagnosis or 00:09:12.450 --> 00:09:14.570 is it just trying to sound more human, maybe 00:09:14.570 --> 00:09:17.090 less like a scary robot to build trust? That's 00:09:17.090 --> 00:09:18.950 a really good question. And maybe the most important 00:09:18.950 --> 00:09:21.769 part is how the NEJM handled it. They published 00:09:21.769 --> 00:09:24.330 the AI's diagnosis right next to the human experts 00:09:24.330 --> 00:09:26.769 reasoning for the same case. And critically, 00:09:27.009 --> 00:09:28.889 they didn't hide its flaws. They pointed them 00:09:28.889 --> 00:09:31.769 out. Total transparency, which you absolutely 00:09:31.769 --> 00:09:34.029 need for medical AI. Right. That transparency 00:09:34.029 --> 00:09:37.320 is everything for actual deployment. So while 00:09:37.320 --> 00:09:40.200 this is super impressive, Dr. Cabot isn't quite 00:09:40.200 --> 00:09:42.960 ready for your local hospital yet. But if you're 00:09:42.960 --> 00:09:45.600 curious, you can actually watch 15 of these AI 00:09:45.600 --> 00:09:48.529 case talks online. So you justify its conclusions 00:09:48.529 --> 00:09:51.669 step by step. So how absolutely vital is it then 00:09:51.669 --> 00:09:54.309 that AI cannot just get the answer right, but 00:09:54.309 --> 00:09:56.789 also clearly explained its whole thought process 00:09:56.789 --> 00:09:59.370 to really earn that professional trust? Yeah, 00:09:59.409 --> 00:10:01.629 I think transparency and explaining the why, 00:10:01.789 --> 00:10:03.809 they aren't nice to haves. They're absolutely 00:10:03.809 --> 00:10:06.029 essential. The foundation for using any expert 00:10:06.029 --> 00:10:07.950 system in fields where the stakes are this high, 00:10:08.090 --> 00:10:10.769 hashtag, tag, tag, recap, and outro. So the big 00:10:10.769 --> 00:10:12.509 idea kind of weaving through all the sources 00:10:12.509 --> 00:10:15.649 today seems to be this shift. AI isn't just about 00:10:15.649 --> 00:10:17.919 massive scale. anymore it's moving towards specialized 00:10:17.919 --> 00:10:21.360 efficient quality exactly we saw these tiny recursive 00:10:21.360 --> 00:10:24.100 models like TRM showing that clever design smart 00:10:24.100 --> 00:10:26.639 architecture can actually beat raw computing 00:10:26.639 --> 00:10:29.220 power on tough reasoning tasks and then we saw 00:10:29.220 --> 00:10:31.679 specialized agents like dr. Cabot hitting major 00:10:31.679 --> 00:10:33.879 professional milestones in fields like medicine 00:10:33.879 --> 00:10:36.419 where transparency and explaining yourself are 00:10:36.419 --> 00:10:38.919 non -negotiable right So we definitely encourage 00:10:38.919 --> 00:10:41.820 you to dig deeper into the sources you sent us. 00:10:41.960 --> 00:10:44.519 Maybe read up a bit more on those TRM concepts. 00:10:44.620 --> 00:10:47.080 Or even better, go watch one of those Dr. Cabot 00:10:47.080 --> 00:10:49.860 case talks online. See that high stakes, transparent 00:10:49.860 --> 00:10:52.460 AI reasoning actually happen. It's quite something. 00:10:52.639 --> 00:10:55.320 Yeah, it really is eye opening to watch. So here's 00:10:55.320 --> 00:10:57.620 a final thought to leave you with. If an AI can 00:10:57.620 --> 00:10:59.799 publish a complex, peer -reviewed differential 00:10:59.799 --> 00:11:02.919 diagnosis in just five minutes, what's the next 00:11:02.919 --> 00:11:05.600 traditionally human -exclusive professional achievement 00:11:05.600 --> 00:11:08.279 we should be preparing for as a society? Something 00:11:08.279 --> 00:11:10.100 to think about. Out to your music.