WEBVTT 00:00:00.000 --> 00:00:03.160 So everyone kind of thought AI would replace 00:00:03.160 --> 00:00:05.639 writers first. Right. Or maybe coders. That was 00:00:05.639 --> 00:00:08.339 the next big thing. But what about forecasters? 00:00:08.800 --> 00:00:12.019 Could AI actually predict real world events? 00:00:12.099 --> 00:00:14.919 Yeah, that's a wild thought. Are these models 00:00:14.919 --> 00:00:17.739 maybe already seeing possibilities that, you 00:00:17.739 --> 00:00:20.300 know, we just can't mapping things out? Well. 00:00:20.600 --> 00:00:22.600 Welcome to the Deep Dive. Today we're going to 00:00:22.600 --> 00:00:25.179 dig into some really fascinating source material 00:00:25.179 --> 00:00:29.500 here. Our mission really is to unpack how these 00:00:29.500 --> 00:00:32.460 advanced AI models are doing more than just like 00:00:32.460 --> 00:00:34.479 understanding language. They're actually making 00:00:34.479 --> 00:00:37.259 these nuanced bets on the future, probabilistic 00:00:37.259 --> 00:00:40.159 bets. Okay. So we'll look at models in live prediction 00:00:40.159 --> 00:00:42.219 markets first, then we'll shift over and talk 00:00:42.219 --> 00:00:44.399 about this quiet giant, a new frontier model 00:00:44.399 --> 00:00:46.039 that just sort of appeared. Oh, interesting. 00:00:46.299 --> 00:00:48.460 And finally, we'll touch on some of the AI tools 00:00:48.460 --> 00:00:50.640 that are, you know, already. changing how we 00:00:50.640 --> 00:00:53.479 work every day, get ready for some pretty eye 00:00:53.479 --> 00:00:55.759 -opening stuff, I think. All right, let's jump 00:00:55.759 --> 00:00:58.439 into that first big idea then, AI models stepping 00:00:58.439 --> 00:01:01.159 into prediction markets. There's this new platform 00:01:01.159 --> 00:01:04.620 from the University of Chicago, and it puts top 00:01:04.620 --> 00:01:08.060 AI models right into these live markets. So like 00:01:08.060 --> 00:01:11.959 real money, real events. Seems like it, or at 00:01:11.959 --> 00:01:14.400 least simulated stakes that mirror real markets. 00:01:14.780 --> 00:01:16.840 Think of them as these betting pools for what's 00:01:16.840 --> 00:01:18.540 going to happen next. Okay, like what kind of 00:01:18.540 --> 00:01:20.930 things are they betting on? All sorts. Who will 00:01:20.930 --> 00:01:23.250 win the next election, where crypto prices might 00:01:23.250 --> 00:01:26.650 land, even the outcome of an MLS soccer game. 00:01:26.790 --> 00:01:29.310 Wow. OK, so how does it work? Do they just say 00:01:29.310 --> 00:01:32.599 Team A wins? No, it's more subtle. They make 00:01:32.599 --> 00:01:36.379 probabilistic bets. So instead of just X happens, 00:01:36.540 --> 00:01:39.079 it's more like, okay, I think there's a 70 % 00:01:39.079 --> 00:01:41.159 chance of X happening. Gotcha. So they assign 00:01:41.159 --> 00:01:43.620 odds. Exactly. And what's really cool is they 00:01:43.620 --> 00:01:45.739 apparently also provide explanations for why 00:01:45.739 --> 00:01:47.620 they made that bet. Gives you a peek under the 00:01:47.620 --> 00:01:49.920 hood. Okay. That reasoning part is key. So what 00:01:49.920 --> 00:01:52.019 are the early results? Anything surprising? Oh, 00:01:52.040 --> 00:01:55.299 yeah. Definitely surprising. Take OpenAI's O3 00:01:55.299 --> 00:01:59.209 Mini. apparently turned $1 into $9 betting on 00:01:59.209 --> 00:02:02.349 an MLS game. Whoa, really? Well, the human market, 00:02:02.450 --> 00:02:05.269 you know, the consensus, only gave Toronto FC 00:02:05.269 --> 00:02:07.810 an 11 % chance to win. Okay, pretty low odds. 00:02:07.950 --> 00:02:10.909 Right, but O3 Mini saw it differently. It put 00:02:10.909 --> 00:02:13.810 the odds at 30%. It's a pretty bold bet against 00:02:13.810 --> 00:02:15.969 the crowd. And it paid off, obviously. Nine times 00:02:15.969 --> 00:02:18.069 the money. Okay. Yeah. And you see differences 00:02:18.069 --> 00:02:21.210 between the models, too. Like, Quinn 3 was super 00:02:21.210 --> 00:02:23.610 confident about AI regulation passing, like 75 00:02:23.610 --> 00:02:26.159 % chance. That's bullish. But then Meta's law 00:02:26.159 --> 00:02:28.500 for Maverick was way more cautious on the same 00:02:28.500 --> 00:02:31.539 thing, only 35%. Huh. So they disagree quite 00:02:31.539 --> 00:02:35.060 a bit. They do. And here's another wrinkle. GPT 00:02:35.060 --> 00:02:37.539 -5, often it's the most accurate model overall. 00:02:37.740 --> 00:02:39.680 Right. Gets the most answers correct. Exactly. 00:02:40.060 --> 00:02:42.180 But O3 Mini is the one making the most profit. 00:02:42.360 --> 00:02:45.979 Huh. Okay. So being right and making money aren't 00:02:45.979 --> 00:02:47.639 always the same thing, are they? Apparently not. 00:02:47.949 --> 00:02:50.210 Then you've got Deep Seek R1. Sometimes it just 00:02:50.210 --> 00:02:53.349 went chaos mode is what they call it. Chaos mode. 00:02:53.430 --> 00:02:55.550 What's that mean? Betting 0 % on everything. 00:02:55.689 --> 00:02:57.669 Yeah. Just flat zeros. Okay, that sounds like 00:02:57.669 --> 00:03:01.370 a terrible strategy. You'd think. But somehow, 00:03:01.569 --> 00:03:04.189 it still made money by hitting some major upsets 00:03:04.189 --> 00:03:06.090 that nobody else saw coming. It's kind of wild. 00:03:06.250 --> 00:03:10.030 That is wild. Almost spooky. Any other big calls? 00:03:10.289 --> 00:03:12.770 Yeah, one notable one. Llama 4 Maverick was the 00:03:12.770 --> 00:03:15.780 only model, apparently. To predict the Zoran 00:03:15.780 --> 00:03:17.900 upset. Oh, I remember that the local election 00:03:17.900 --> 00:03:19.860 surprise, no human pollster saw that coming. 00:03:20.000 --> 00:03:23.039 Right. And interestingly, anthropics clawed models. 00:03:23.379 --> 00:03:26.939 They were just absent, not really showing up 00:03:26.939 --> 00:03:29.159 on the leaderboard. Hmm. Strange. Wonder why. 00:03:29.759 --> 00:03:32.960 Good question. The bigger implication here. Hmm. 00:03:33.180 --> 00:03:35.699 What does this all really mean? Yeah. What's 00:03:35.699 --> 00:03:38.180 the takeaway? Well, these models are making bets 00:03:38.180 --> 00:03:40.620 on things like the 2028 election already. Wow, 00:03:40.740 --> 00:03:42.419 okay, looking way ahead. And their predictions 00:03:42.419 --> 00:03:44.580 don't line up with current human polling data. 00:03:44.800 --> 00:03:47.199 Not even close sometimes. So they're seeing something 00:03:47.199 --> 00:03:49.500 different. It really suggests that. Maybe some 00:03:49.500 --> 00:03:51.580 models have an understanding, like a world model, 00:03:51.740 --> 00:03:54.840 that we humans just aren't quite getting. Two 00:03:54.840 --> 00:03:58.300 -sec silence. This could be maybe the cleanest 00:03:58.300 --> 00:04:01.240 test we've seen yet of how well AI can reason 00:04:01.240 --> 00:04:03.250 about the real world. Right, because it's tied 00:04:03.250 --> 00:04:05.370 to actual outcomes, not just language tasks. 00:04:05.710 --> 00:04:09.789 Exactly. And if this holds up, well, Wall Street 00:04:09.789 --> 00:04:11.830 might need to pay attention. Anyone who relies 00:04:11.830 --> 00:04:15.189 on forecasting, really. Yeah, big time. So probing 00:04:15.189 --> 00:04:17.790 question then. What's the biggest implication 00:04:17.790 --> 00:04:21.449 if these models do consistently outperform human 00:04:21.449 --> 00:04:24.629 forecasters? It suggests a profound shift in 00:04:24.629 --> 00:04:26.889 how we understand and use predictive insights. 00:04:27.149 --> 00:04:29.089 A really profound shift. Okay, so let's switch 00:04:29.089 --> 00:04:31.910 gears now. From betting on the future to the 00:04:31.910 --> 00:04:34.629 models that actually do the thinking. Frontier 00:04:34.629 --> 00:04:38.170 models. Right. So get this. A massive new open 00:04:38.170 --> 00:04:41.029 source model just appeared. Like zero hype. Just 00:04:41.029 --> 00:04:43.889 quietly uploaded to Hugging Face. No big announcement. 00:04:44.129 --> 00:04:46.149 Just there. Pretty much. It's called DeepSeek 00:04:46.149 --> 00:04:50.410 V3 .1. And it is a beast. We're talking 685 billion 00:04:50.410 --> 00:04:53.160 parameters. Whoa. Okay, for listeners, parameters 00:04:53.160 --> 00:04:55.040 are kind of like the learned data points, the 00:04:55.040 --> 00:04:57.019 connections inside the AI's brain, right? More 00:04:57.019 --> 00:04:59.699 means more capacity. Exactly. And $685 billion 00:04:59.699 --> 00:05:01.420 puts it right up there with the biggest models 00:05:01.420 --> 00:05:04.240 out there like GPT -5, CLAWD -4. It's a serious 00:05:04.240 --> 00:05:06.379 contender. And it's coming out of China. And 00:05:06.379 --> 00:05:07.819 it's gunning straight for the top dogs, you're 00:05:07.819 --> 00:05:10.199 saying. Open source, though. That's the kicker. 00:05:10.240 --> 00:05:12.899 It's apparently faster than CLAWD, way cheaper 00:05:12.899 --> 00:05:15.740 to run than GPT, and yeah, totally open source. 00:05:15.860 --> 00:05:18.480 Okay, that could be huge. The most capable free 00:05:18.480 --> 00:05:20.720 model out there, potentially. It really could 00:05:20.720 --> 00:05:23.170 be. Early tests, and these are independent tests, 00:05:23.389 --> 00:05:25.589 show it's matching or sometimes even beating 00:05:25.589 --> 00:05:29.170 GPT -5 and CLAWD4 on real -world tasks. Got an 00:05:29.170 --> 00:05:32.250 example. Yeah. On the ATER coding benchmark, 00:05:32.610 --> 00:05:37.149 it scored 71 .6%. Okay. How does that stack up? 00:05:37.290 --> 00:05:40.189 That actually slightly beats CLAWD Opus 4. And 00:05:40.189 --> 00:05:43.290 here's the crazy part. It's apparently 68 times 00:05:43.290 --> 00:05:47.370 cheaper to run. 68 times. Yeah. Wow. Okay, that 00:05:47.370 --> 00:05:49.370 is a big deal for developers, for anyone trying 00:05:49.370 --> 00:05:52.610 to build things like BI. Huge deal. And what's 00:05:52.610 --> 00:05:54.310 interesting is how it does it. You know how some 00:05:54.310 --> 00:05:57.930 open models feel kind of nerfed or just bloated 00:05:57.930 --> 00:05:59.850 and slow? Yeah, sometimes they're not quite ready 00:05:59.850 --> 00:06:02.470 for prime time. Right. But v3 .1 seems to be 00:06:02.470 --> 00:06:04.949 both really high performance and efficient. Fast 00:06:04.949 --> 00:06:06.870 enough for real -time stuff. Yeah, no twisters. 00:06:07.370 --> 00:06:09.550 Any technical tricks? Well, it supports different 00:06:09.550 --> 00:06:13.589 precision formats like BF16 and FP8. Basically, 00:06:13.750 --> 00:06:17.000 Waze the AI handles numbers. Lower precision 00:06:17.000 --> 00:06:19.980 can mean faster speed and less memory, sometimes 00:06:19.980 --> 00:06:22.339 with a tiny hit to accuracy, but often worth 00:06:22.339 --> 00:06:25.139 it. Okay, so more flexibility for developers 00:06:25.139 --> 00:06:27.199 to run it on different kinds of hardware. Makes 00:06:27.199 --> 00:06:29.540 sense. Exactly. Plus, it has some cool new features 00:06:29.540 --> 00:06:32.600 built in, things called thinking tokens and search 00:06:32.600 --> 00:06:35.860 tokens. Thinking tokens. Search tokens. What 00:06:35.860 --> 00:06:38.000 do those do? So thinking tokens, they kind of 00:06:38.000 --> 00:06:40.699 let the model do more internal work, like reasoning 00:06:40.699 --> 00:06:42.920 through a problem before giving the answer, sort 00:06:42.920 --> 00:06:45.300 of like showing its work internally. Oh, okay. 00:06:45.639 --> 00:06:47.540 Like chain of thought prompting, but maybe built 00:06:47.540 --> 00:06:50.060 in. Kind of like that, yeah. And search tokens 00:06:50.060 --> 00:06:52.500 let it pull in live information from the web 00:06:52.500 --> 00:06:54.959 as part of its process. So it can reason and 00:06:54.959 --> 00:06:57.079 access up -to -the -minute data? That seems to 00:06:57.079 --> 00:06:59.839 be the idea. Pretty powerful combo. And the community 00:06:59.839 --> 00:07:02.220 noticed DeepSeek had quietly updated everything, 00:07:02.519 --> 00:07:05.399 removing their older R1 model. Ah, streamlining 00:07:05.399 --> 00:07:09.139 things. Yep. And this V3 .1 just shot to the 00:07:09.139 --> 00:07:11.540 top of the trending models on Hucking Face almost 00:07:11.540 --> 00:07:14.740 instantly after it was uploaded. Whoa. Just imagine 00:07:14.740 --> 00:07:17.720 the potential there. Models that can really integrate 00:07:17.720 --> 00:07:21.759 that deep internal reasoning with live external 00:07:21.759 --> 00:07:25.779 data, all at that kind of scale. Like stacking 00:07:25.779 --> 00:07:28.600 Lego blocks of data and thought together in a 00:07:28.600 --> 00:07:30.300 whole new way. Yeah, the possibilities are kind 00:07:30.300 --> 00:07:32.480 of staggering. So here's the probing question. 00:07:33.279 --> 00:07:37.079 What does this quiet release, this powerful open 00:07:37.079 --> 00:07:40.839 model just appearing, what does that mean for 00:07:40.839 --> 00:07:44.259 AI accessibility? For everyone. It signals a 00:07:44.259 --> 00:07:46.939 powerful, cheaper and definitely more open future 00:07:46.939 --> 00:07:50.019 for advanced AI, I think. Democratizing it a 00:07:50.019 --> 00:07:51.920 bit more. Right. OK, let's shift one more time 00:07:51.920 --> 00:07:54.730 from these huge frontier models back to. More 00:07:54.730 --> 00:07:56.829 immediate stuff. Practical tools and developments 00:07:56.829 --> 00:07:58.370 we're seeing right now. Yeah, the day -to -day 00:07:58.370 --> 00:08:00.810 impact. Exactly. Because beyond the absolute 00:08:00.810 --> 00:08:03.269 cutting edge, AI is weaving itself into the tools 00:08:03.269 --> 00:08:06.410 we use all the time. Like Grammarly just launched 00:08:06.410 --> 00:08:08.670 eight new AI agents. Indeed of them. What do 00:08:08.670 --> 00:08:10.629 they do? Think of them like specific writing 00:08:10.629 --> 00:08:12.670 helpers. They help with grading, checking for 00:08:12.670 --> 00:08:15.050 plagiarism, even spotting AI -generated text, 00:08:15.209 --> 00:08:18.000 helping you brainstorm. across the whole writing 00:08:18.000 --> 00:08:20.339 process. I like specialized assistants. Yeah. 00:08:20.420 --> 00:08:22.360 And I got to admit, you know, I still wrestle 00:08:22.360 --> 00:08:24.560 with prompt drift myself sometimes getting the 00:08:24.560 --> 00:08:26.899 AI to stay on track. Oh yeah, me too. It can 00:08:26.899 --> 00:08:29.019 wander off. So having these kinds of smarter 00:08:29.019 --> 00:08:31.620 assistants built right into where you're already 00:08:31.620 --> 00:08:33.700 working, that's super helpful, actually becoming 00:08:33.700 --> 00:08:36.440 essential, I'd say. Totally agree. And speaking 00:08:36.440 --> 00:08:38.980 of integration, we're also seeing new ways to 00:08:38.980 --> 00:08:43.509 like measure AI's impact. How so? Well, RFs, 00:08:43.590 --> 00:08:47.529 the SEO tool company, they just launched a dashboard 00:08:47.529 --> 00:08:50.769 that tracks web traffic coming from ChatGPT and 00:08:50.769 --> 00:08:53.110 Google's AI overviews. Oh, interesting. So you 00:08:53.110 --> 00:08:55.710 can see how much traffic AI search is sending 00:08:55.710 --> 00:08:58.870 you. Exactly. Across like 44 ,000 sites already. 00:08:58.970 --> 00:09:01.350 So you can finally track AI search as a real 00:09:01.350 --> 00:09:04.009 channel, measure its ROI. That's going to change 00:09:04.009 --> 00:09:06.539 SEO for sure. Yeah, definitely. You need to know 00:09:06.539 --> 00:09:08.200 where your visitors are coming from. And another 00:09:08.200 --> 00:09:11.820 big one, Microsoft Copilot AI. It's now directly 00:09:11.820 --> 00:09:14.919 in Excel. Wait, in Excel itself? How does that 00:09:14.919 --> 00:09:17.740 work? There's literally a new function like sum 00:09:17.740 --> 00:09:20.980 or average, but it's Copilot. You can type plain 00:09:20.980 --> 00:09:23.980 English in it like Copilot, summarize sales data 00:09:23.980 --> 00:09:27.779 in column C, or Copilot, find the Q3 revenue 00:09:27.779 --> 00:09:30.720 for product X from our sales report. Wow. So 00:09:30.720 --> 00:09:32.919 you can pull data, analyze it just by typing 00:09:32.919 --> 00:09:35.620 a formula in natural language. Yeah. Pretty slick, 00:09:35.740 --> 00:09:37.879 right? Especially for people who live in spreadsheets 00:09:37.879 --> 00:09:40.799 all day. That is slick. Okay, and any other quick 00:09:40.799 --> 00:09:42.460 hits from the AI world that caught your eye? 00:09:42.639 --> 00:09:45.279 Yeah, a few rapid -fire ones. Adobe's building 00:09:45.279 --> 00:09:48.100 out its AI features, like a new AI home with 00:09:48.100 --> 00:09:51.080 PDF tools and creation stuff. Integrating it 00:09:51.080 --> 00:09:53.740 deeper into creative workflows. Meta is apparently 00:09:53.740 --> 00:09:56.639 rebuilding its whole AI division midstream to 00:09:56.639 --> 00:09:58.659 try and catch up with OpenAI, a big internal 00:09:58.659 --> 00:10:00.860 push there. The race is definitely on. OpenAI 00:10:00.860 --> 00:10:03.139 says it's now very serious about adding proper 00:10:03.139 --> 00:10:05.919 encryption to chat GPT. Big deal for privacy 00:10:05.919 --> 00:10:07.620 if that happens. Yeah, that would be significant. 00:10:07.899 --> 00:10:10.500 ByteDance they own. TikTok has a new model called 00:10:10.500 --> 00:10:13.360 M3 Agent that can process video and audio together 00:10:13.360 --> 00:10:15.980 in real time. Think smarter content moderation 00:10:15.980 --> 00:10:19.679 or analysis. Okay, multimedia AI. And NVIDIA 00:10:19.679 --> 00:10:23.159 released a new small open model, but it has a 00:10:23.159 --> 00:10:25.879 reasoning toggle. A reasoning toggle. Like you 00:10:25.879 --> 00:10:28.539 can turn its thinking process on or off. Kind 00:10:28.539 --> 00:10:31.379 of seems like that, yeah. Maybe control how much 00:10:31.379 --> 00:10:33.220 effort it puts into reasoning versus just giving 00:10:33.220 --> 00:10:36.519 a quick answer. Cool concept. Lots happening. 00:10:36.960 --> 00:10:40.500 So zooming out slightly, how do all these smaller, 00:10:40.600 --> 00:10:44.720 more specific AI tools, these integrations, how 00:10:44.720 --> 00:10:48.009 do they change our actual daily workflows? They 00:10:48.009 --> 00:10:50.649 make complex tasks simpler, way more efficient, 00:10:50.730 --> 00:10:53.029 and honestly, often more creative, too. Yeah, 00:10:53.049 --> 00:10:54.529 taking the friction out of things, letting you 00:10:54.529 --> 00:10:56.529 focus on the bigger picture. Exactly. They handle 00:10:56.529 --> 00:10:58.690 the drunk work. Okay, so let's try and wrap this 00:10:58.690 --> 00:11:00.629 all together. What's the big picture here for 00:11:00.629 --> 00:11:02.570 you, the listener? Yeah, what does it all mean? 00:11:03.070 --> 00:11:05.789 Well, we've seen AI models jumping into real 00:11:05.789 --> 00:11:08.409 -world prediction markets, making these surprising, 00:11:08.629 --> 00:11:10.909 sometimes really profitable bets. It suggests 00:11:10.909 --> 00:11:13.830 they might be seeing things we don't. Then we 00:11:13.830 --> 00:11:16.149 saw that monster open -source model, DeepSeek 00:11:16.149 --> 00:11:19.960 v3 .1. Just appear quietly, challenging the big 00:11:19.960 --> 00:11:22.759 guys, offering huge power much more affordably, 00:11:22.840 --> 00:11:26.039 democratizing things. Right. And we've seen AI 00:11:26.039 --> 00:11:30.100 weaving itself into everyday tools. Ramily, Excel, 00:11:30.419 --> 00:11:34.500 Adobe. Making us more efficient, maybe more creative. 00:11:34.759 --> 00:11:37.639 So it's clear AI isn't some far -off thing anymore. 00:11:37.799 --> 00:11:40.519 It's here. It's here now, actively changing how 00:11:40.519 --> 00:11:43.120 we learn, how we work, and maybe even how we 00:11:43.120 --> 00:11:45.200 see the future itself. And that leads to kind 00:11:45.200 --> 00:11:47.799 of a provocative thought, maybe. The line between 00:11:47.799 --> 00:11:51.659 our own intuition, human gut feelings, and what 00:11:51.659 --> 00:11:53.559 AI comes up with, it's getting really blurry. 00:11:53.720 --> 00:11:55.539 Yeah. Where does one end and the other begin? 00:11:55.779 --> 00:11:58.740 So consider this. If AI models can consistently 00:11:58.740 --> 00:12:01.879 know something we don't about what's coming next, 00:12:02.059 --> 00:12:04.399 if they consistently beat human predictions in 00:12:04.399 --> 00:12:07.100 some areas, what does that really imply about 00:12:07.100 --> 00:12:08.879 the limits of our own knowledge? And maybe even 00:12:08.879 --> 00:12:10.720 more deeply, what does it mean for what we choose 00:12:10.720 --> 00:12:13.559 to believe? Why we believe it? That is a lot 00:12:13.559 --> 00:12:15.580 to chew on. Definitely something to think about. 00:12:15.720 --> 00:12:17.429 For sure. Well, thank you for diving deep with 00:12:17.429 --> 00:12:19.190 us today. We really hope you found some truly 00:12:19.190 --> 00:12:21.129 important nuggets of knowledge and all that. 00:12:21.210 --> 00:12:22.210 Out to your own music.