WEBVTT 00:00:00.000 --> 00:00:01.419 You've probably heard all the hype, right? AI 00:00:01.419 --> 00:00:04.240 agents automating everything, your calendar, 00:00:04.360 --> 00:00:07.019 your code, promising to totally transform your 00:00:07.019 --> 00:00:09.400 work, maybe even your whole job. But what if 00:00:09.400 --> 00:00:12.220 right now, like in their current form, they're 00:00:12.220 --> 00:00:15.890 actually, well. Kind of useless for most of that. 00:00:16.030 --> 00:00:19.589 Small laugh. Yeah, that's the big question, isn't 00:00:19.589 --> 00:00:22.289 it? The reality on the ground, well, it often 00:00:22.289 --> 00:00:24.769 clashes pretty sharply with the marketing story 00:00:24.769 --> 00:00:26.929 we all keep hearing. Welcome to the Deep Dive. 00:00:26.949 --> 00:00:28.870 Today, we're really going to try and cut through 00:00:28.870 --> 00:00:30.570 that noise. We want to get to the true state 00:00:30.570 --> 00:00:33.630 of AI agents, the nuances, and maybe even more 00:00:33.630 --> 00:00:36.810 surprisingly, how we humans are really using 00:00:36.810 --> 00:00:39.090 artificial intelligence day to day. And this 00:00:39.090 --> 00:00:41.850 Deep Dive, it's built from a whole stack of fresh 00:00:41.850 --> 00:00:43.990 sources. We've got cutting edge academic research. 00:00:44.109 --> 00:00:47.450 The latest industry insights, reports, the works. 00:00:47.729 --> 00:00:49.649 Yeah. So we're going to unpack the current reality 00:00:49.649 --> 00:00:52.729 of these so -called AI agents, you know, strip 00:00:52.729 --> 00:00:54.969 away the PR gloss and see what they can actually 00:00:54.969 --> 00:00:57.530 do. Then we'll do a kind of rapid fire tour through 00:00:57.530 --> 00:01:01.850 some genuinely fascinating and sometimes. honestly 00:01:01.850 --> 00:01:03.890 concerning AI highlights from just the past week 00:01:03.890 --> 00:01:06.310 or so. And finally, we'll dive into this really 00:01:06.310 --> 00:01:08.870 surprising new report. It challenges a lot of 00:01:08.870 --> 00:01:11.530 the conventional wisdom about how we're actually 00:01:11.530 --> 00:01:14.549 using chatbots. It's really separating the marketing. 00:01:15.069 --> 00:01:17.290 you know from the actual magic and figuring out 00:01:17.290 --> 00:01:19.109 what's really making an impact okay so let's 00:01:19.109 --> 00:01:22.269 unpack this this big promise of ai agents for 00:01:22.269 --> 00:01:24.829 a while now the common idea the perception has 00:01:24.829 --> 00:01:26.769 been that these autonomous digital things are 00:01:26.769 --> 00:01:29.250 just around the corner ready to handle all your 00:01:29.250 --> 00:01:32.450 complex daily tasks like you know automating 00:01:32.450 --> 00:01:35.230 your calendar writing whole blocks of production 00:01:35.230 --> 00:01:38.709 ready code sending personalized emails basically 00:01:38.709 --> 00:01:40.709 doing big chunks of your job while you just kind 00:01:40.709 --> 00:01:43.120 of kick back But is that really where we are? 00:01:43.180 --> 00:01:45.599 Is the promise matching the performance? Slight 00:01:45.599 --> 00:01:48.900 pause. Well, the research paints a pretty stark 00:01:48.900 --> 00:01:52.310 picture, honestly. It's quite sobering. Researchers 00:01:52.310 --> 00:01:54.310 over at Carnegie Mellon University working with 00:01:54.310 --> 00:01:57.489 Salesforce recently put some of the leading AI 00:01:57.489 --> 00:02:00.129 models through a really tough test. They weren't 00:02:00.129 --> 00:02:02.129 just testing like isolated schools. They were 00:02:02.129 --> 00:02:05.930 looking at realistic office style tasks. We're 00:02:05.930 --> 00:02:08.169 talking multi -step workflows, things like debugging 00:02:08.169 --> 00:02:11.330 code, searching the web for info, coordinating 00:02:11.330 --> 00:02:14.229 with teammates via messages, and following really 00:02:14.229 --> 00:02:17.110 complex nested instructions to finish a project. 00:02:17.770 --> 00:02:20.069 This isn't just asking it to write a poem, you 00:02:20.069 --> 00:02:22.840 know. It's asking it to act like a, well, like 00:02:22.840 --> 00:02:25.560 a junior employee on a team. Right. And the results, 00:02:25.699 --> 00:02:27.939 I gather they weren't just not great, but maybe 00:02:27.939 --> 00:02:30.219 quite poor, especially given the hype out there. 00:02:30.340 --> 00:02:32.460 Surprisingly poor, might be putting it mildly. 00:02:32.479 --> 00:02:34.780 When you look at the actual numbers, it's stark. 00:02:35.180 --> 00:02:37.439 Gemini 2 .5 Pro, which was actually the best 00:02:37.439 --> 00:02:39.900 performer of the bunch, completed just 30 % of 00:02:39.900 --> 00:02:42.360 these multi -step tasks successfully. 30%. Yeah. 00:02:42.639 --> 00:02:45.919 Claude 3 .5 Sonnet, it managed around 24%. And 00:02:45.919 --> 00:02:49.340 GPT -4 .0. a truly dismal 8 .6%. Wow, single 00:02:49.340 --> 00:02:51.759 digits. Yeah, single digits. Most of the other 00:02:51.759 --> 00:02:53.800 models they tested were under 10%, and some were 00:02:53.800 --> 00:02:56.159 struggling down near 1%. And these were tasks 00:02:56.159 --> 00:02:58.960 that... Frankly, a junior employee should handle, 00:02:59.060 --> 00:03:01.000 you know, maybe day two on the job with a bit 00:03:01.000 --> 00:03:03.319 of guidance. The models really struggled with 00:03:03.319 --> 00:03:06.400 common sense reasoning, keeping context over 00:03:06.400 --> 00:03:09.259 multiple steps and bouncing back from small errors. 00:03:09.340 --> 00:03:11.860 It's like they could build the first few Lego 00:03:11.860 --> 00:03:14.300 blocks maybe, but then totally forgot what the 00:03:14.300 --> 00:03:16.060 final thing was supposed to look like. So it's 00:03:16.060 --> 00:03:18.900 less autonomous super brain and maybe more like 00:03:18.900 --> 00:03:22.099 an overwhelmed intern who needs constant watching. 00:03:22.199 --> 00:03:24.099 I heard CMU even launched something called the 00:03:24.099 --> 00:03:27.280 Agent Company to test this more. in a controlled 00:03:27.280 --> 00:03:29.599 way. Right. It's a brilliant new benchmark, actually. 00:03:29.759 --> 00:03:32.259 The agent company literally simulates a small 00:03:32.259 --> 00:03:34.639 software company. You could basically drop AI 00:03:34.639 --> 00:03:37.080 agents into this fake startup, give them real 00:03:37.080 --> 00:03:39.300 -world problems like develop this new feature 00:03:39.300 --> 00:03:41.719 or fix this bug, and see if they can actually 00:03:41.719 --> 00:03:43.919 survive and contribute meaningfully to the code 00:03:43.919 --> 00:03:46.159 and the team. It's a fantastic way to measure 00:03:46.159 --> 00:03:49.419 genuine multi -step function and collaboration, 00:03:49.599 --> 00:03:52.159 not just isolated skills in a lab. This really 00:03:52.159 --> 00:03:54.319 brings to mind that Gartner concept of agent 00:03:54.319 --> 00:03:56.819 washing. They're suggesting that a lot of these 00:03:56.819 --> 00:03:59.919 so -called AI agents out there right now are 00:03:59.919 --> 00:04:01.960 just, well... fancy AI assistance, they just 00:04:01.960 --> 00:04:04.219 can't plan more than, what, two steps ahead, 00:04:04.419 --> 00:04:07.280 which makes them more reactive than really proactive 00:04:07.280 --> 00:04:09.419 problem solvers. That feels almost misleading, 00:04:09.620 --> 00:04:11.500 doesn't it? It absolutely is misleading. Gartner 00:04:11.500 --> 00:04:15.539 estimates that only about 130 vendors worldwide 00:04:15.539 --> 00:04:18.199 are building anything that's even close to true 00:04:18.199 --> 00:04:21.639 agentic AI right now. Only 130? Yeah. It's AI 00:04:21.639 --> 00:04:25.139 capable of complex multi -step planning and execution 00:04:25.139 --> 00:04:28.560 without... needing constant human help. The vast 00:04:28.560 --> 00:04:30.439 majority of what's being marketed as an agent 00:04:30.439 --> 00:04:32.680 is, well, it's just marketing. And they have 00:04:32.680 --> 00:04:35.699 this pretty bold, almost shocking prediction. 00:04:36.339 --> 00:04:40.040 By 2027, over 40 % of these so -called agentic 00:04:40.040 --> 00:04:41.800 AI projects that are currently being developed, 00:04:41.980 --> 00:04:44.720 they'll be canceled, just scrapped. 40%, wow. 00:04:44.980 --> 00:04:47.100 That's a huge market correction coming. It implies 00:04:47.100 --> 00:04:49.560 a lot of wasted money and dash hopes for many 00:04:49.560 --> 00:04:52.019 companies. That's a serious number of projects 00:04:52.019 --> 00:04:54.759 heading for the bin. But hang on, it's not all 00:04:54.759 --> 00:04:56.740 bad news for agents, is it? there must be some 00:04:56.740 --> 00:04:58.379 bright spots where they do show some promise. 00:04:58.560 --> 00:05:02.680 Not entirely doom and gloom, no. Some agent uses 00:05:02.680 --> 00:05:06.699 do show real promise, but, and this is key, with 00:05:06.699 --> 00:05:09.319 very specific constraints. They can be pretty 00:05:09.319 --> 00:05:11.800 decent at code generation, for example. Although 00:05:11.800 --> 00:05:14.620 the output often needs human review and tweaking, 00:05:14.720 --> 00:05:18.019 it's really perfect out of the box. And they 00:05:18.019 --> 00:05:20.439 can manage workflow automation, but really only 00:05:20.439 --> 00:05:23.759 in very narrow, highly defined, kind of linear 00:05:23.759 --> 00:05:26.319 setups. Okay, so the conditions for success are 00:05:26.319 --> 00:05:29.220 pretty specific, need to be controlled. Precisely. 00:05:29.379 --> 00:05:32.040 If you keep them in a sandbox, you know, a tightly 00:05:32.040 --> 00:05:34.000 controlled environment, monitor their output 00:05:34.000 --> 00:05:36.439 constantly, and make sure the tasks are simple 00:05:36.439 --> 00:05:38.839 and linear, like say... automating a specific 00:05:38.839 --> 00:05:40.899 data entry process with crystal clear rules, 00:05:41.060 --> 00:05:43.459 then yeah, they can be effective tools. But the 00:05:43.459 --> 00:05:45.180 moment you step outside those conditions, ask 00:05:45.180 --> 00:05:47.360 them to handle ambiguity or need them to adapt 00:05:47.360 --> 00:05:49.439 to something unexpected, things tend to fall 00:05:49.439 --> 00:05:51.439 apart pretty fast. It's just not general intelligence 00:05:51.439 --> 00:05:53.959 yet. Not even close. So what does this all mean 00:05:53.959 --> 00:05:56.800 for us then? Why does it matter so much that 00:05:56.800 --> 00:05:59.379 agents aren't really living up to the hype in 00:05:59.379 --> 00:06:01.769 these broader ways? Well, it matters because 00:06:01.769 --> 00:06:04.350 on some level, almost everyone's kind of involved 00:06:04.350 --> 00:06:06.290 in a bit of collective self -deception here. 00:06:06.389 --> 00:06:10.160 You've got vendors racing to show. progress, 00:06:10.160 --> 00:06:12.939 even if it's just surface level or maybe even 00:06:12.939 --> 00:06:17.360 faked, VCs are eager to fund what they hope will 00:06:17.360 --> 00:06:20.639 be the next big platform. Companies desperate 00:06:20.639 --> 00:06:24.459 not to get left behind in the AI race are overbuying 00:06:24.459 --> 00:06:26.540 these tools and maybe under assessing what they 00:06:26.540 --> 00:06:29.540 can actually do. And let's be honest, that enduring 00:06:29.540 --> 00:06:32.920 dream of Jarvis, you know, the fully autonomous, 00:06:33.079 --> 00:06:36.149 all -knowing AI assistant. It's just too appealing 00:06:36.149 --> 00:06:38.069 for people to easily let go of. It's easier to 00:06:38.069 --> 00:06:40.209 believe the hype than face the current limits. 00:06:40.329 --> 00:06:42.550 So distilling that down, what's the core message 00:06:42.550 --> 00:06:45.730 about where AI agents really stand today? Mostly 00:06:45.730 --> 00:06:47.889 hype right now. They're useful, but only in really 00:06:47.889 --> 00:06:50.569 controlled, simple tasks. OK, right. Here's where 00:06:50.569 --> 00:06:52.410 it gets really interesting, though, moving beyond 00:06:52.410 --> 00:06:55.009 just agents. Let's do a kind of rapid fire tour 00:06:55.009 --> 00:06:57.310 of some truly surprising and significant friends 00:06:57.310 --> 00:06:59.649 we're seeing in AI right now. Absolutely. OK, 00:06:59.689 --> 00:07:02.089 kicking it off, you won't believe this, but an 00:07:02.089 --> 00:07:05.579 ex -user. actually created this digital arena 00:07:05.579 --> 00:07:08.480 and pitted the top coding models against each 00:07:08.480 --> 00:07:11.040 other, like in a literal fight to the death. 00:07:11.160 --> 00:07:13.839 Seriously? Yeah. Each model was programmed to 00:07:13.839 --> 00:07:16.199 try and shut down the other's processes while 00:07:16.199 --> 00:07:18.360 defending itself and just trying to stay alive. 00:07:18.540 --> 00:07:20.740 It was this fascinating, like, digital cage match. 00:07:20.879 --> 00:07:23.220 It really showed the adaptive skills and also 00:07:23.220 --> 00:07:25.459 the weaknesses of these models when they face 00:07:25.459 --> 00:07:27.660 a hostile situation. It wasn't just about who 00:07:27.660 --> 00:07:29.500 could code better, but who could outmaneuver 00:07:29.500 --> 00:07:31.819 the others. That's wild. Okay. And then there's 00:07:31.819 --> 00:07:34.420 this Higgs field tool, Sol. It's apparently gone 00:07:34.420 --> 00:07:36.480 viral for incredible realism. Oh, yeah. It's 00:07:36.480 --> 00:07:38.420 making waves. Almost fashion -grade realism. 00:07:38.860 --> 00:07:41.160 in the images it generates. And it has these 00:07:41.160 --> 00:07:44.420 trendy style presets like Y2K or hyper -realistic 00:07:44.420 --> 00:07:46.779 photos. It's absolutely blowing up in creative 00:07:46.779 --> 00:07:49.279 circles. Two secs, silence, whoa. Yeah. Just 00:07:49.279 --> 00:07:51.360 imagine scaling that kind of creative output, 00:07:51.480 --> 00:07:53.959 generating unique fashion content for a whole 00:07:53.959 --> 00:07:57.180 season in like minutes. Or crafting entire virtual 00:07:57.180 --> 00:07:59.500 ad campaigns from scratch for brands. That's 00:07:59.500 --> 00:08:02.079 a really powerful tool. It's truly democratizing 00:08:02.079 --> 00:08:05.759 that high -end visual creation. But on a more... 00:08:06.439 --> 00:08:08.920 Concerning note, there's been this disturbing 00:08:08.920 --> 00:08:12.680 trend spotted on YouTube. We saw, I think it 00:08:12.680 --> 00:08:15.579 was 26 different channels actively pumping out 00:08:15.579 --> 00:08:19.569 fake AI generated videos about the ongoing. Diddy 00:08:19.569 --> 00:08:22.290 trial. Oh, no. Yeah. And these videos, they get 00:08:22.290 --> 00:08:25.389 nearly 70 million views combined across about 00:08:25.389 --> 00:08:28.209 900 videos in a really short time. This isn't 00:08:28.209 --> 00:08:31.110 just, you know, opinion or spin. This is AI being 00:08:31.110 --> 00:08:33.590 used to create totally false narratives, turning 00:08:33.590 --> 00:08:36.250 serious news into sensationalized clickbait. 00:08:36.350 --> 00:08:39.629 It's a real immediate problem for like information 00:08:39.629 --> 00:08:42.039 integrity and public trust. That's worrying. 00:08:42.559 --> 00:08:45.159 OK, also, Google just launched the full version 00:08:45.159 --> 00:08:47.360 of Gemma 3N, right, or their new open model. 00:08:47.480 --> 00:08:49.120 That seems like a pretty big move for the open 00:08:49.120 --> 00:08:51.100 source AI community, making powerful models more 00:08:51.100 --> 00:08:53.100 accessible. Definitely a big deal. And speaking 00:08:53.100 --> 00:08:55.100 of big moves, Meta is reportedly hiring four 00:08:55.100 --> 00:08:57.840 key open AI researchers, poaching them for its 00:08:57.840 --> 00:09:01.259 new super AI team. OpenAI apparently called this 00:09:01.259 --> 00:09:03.379 a side quest, which is kind of funny. But now 00:09:03.379 --> 00:09:05.200 they're scrambling to recalibrate compensation 00:09:05.200 --> 00:09:07.740 to Trump keep the remaining top people. This 00:09:07.740 --> 00:09:09.919 feels very much like a strategic chess game unfolding. 00:09:10.120 --> 00:09:12.639 It does. It really seems like both Elon Musk 00:09:12.639 --> 00:09:16.740 with XAI and Mark Zuckerberg with Meta are aggressively 00:09:16.740 --> 00:09:19.529 trying to position themselves to dominate. directly 00:09:19.529 --> 00:09:22.970 challenging OpenAI's lead by snapping up talent 00:09:22.970 --> 00:09:25.190 and building competing models. And Meta is certainly 00:09:25.190 --> 00:09:26.889 putting its money where its mouth is. They're 00:09:26.889 --> 00:09:28.909 aiming to raise a huge amount. What is it, $29 00:09:28.909 --> 00:09:31.850 billion? Yeah, massive. $3 billion in equity, 00:09:31.990 --> 00:09:34.710 $26 billion in debt from investors like Apollo 00:09:34.710 --> 00:09:37.669 and KKR. And all that cash is specifically going 00:09:37.669 --> 00:09:40.690 towards expanding their AI data centers. They 00:09:40.690 --> 00:09:43.870 want to deploy an astonishing 1 .3 million GPUs 00:09:43.870 --> 00:09:47.649 by 2025. 1 .3 million GPUs. It's an enormous 00:09:47.649 --> 00:09:49.539 investment. It signals a really clear intent 00:09:49.539 --> 00:09:52.440 to be right at the absolute forefront of AI compute 00:09:52.440 --> 00:09:54.720 power. And just a couple of quick hits to round 00:09:54.720 --> 00:09:56.500 out the picture. There was an interview with 00:09:56.500 --> 00:09:59.080 Microsoft CEO Satya Nadella pointing to some 00:09:59.080 --> 00:10:02.500 exciting, maybe unexpected paths for AI's future. 00:10:03.179 --> 00:10:06.399 Also, a very practical warning came out. If you're 00:10:06.399 --> 00:10:08.679 using ChatGPT for certain things, especially 00:10:08.679 --> 00:10:11.139 sensitive stuff like legal advice without human 00:10:11.139 --> 00:10:14.220 review, or definitely diagnosing medical conditions, 00:10:14.480 --> 00:10:17.059 you should probably stop immediately. It's just 00:10:17.059 --> 00:10:19.580 not designed or reliable for that. On a lighter 00:10:19.580 --> 00:10:22.259 note, a useful tip. Remember, OpenAI charges 00:10:22.259 --> 00:10:25.159 by the minute for audio transcription. So speeding 00:10:25.159 --> 00:10:27.279 up your audio before you upload it can actually 00:10:27.279 --> 00:10:29.399 save you some cash. And finally, it's becoming 00:10:29.399 --> 00:10:31.659 really clear that people are now actively looking 00:10:31.659 --> 00:10:34.980 for AI whispers. You know, people with special 00:10:34.980 --> 00:10:37.419 skills in prompt engineering and navigating AI 00:10:37.419 --> 00:10:39.679 tools to guide them through this increasingly 00:10:39.679 --> 00:10:42.779 complex AI world. It's a fascinating new job 00:10:42.779 --> 00:10:45.539 category popping up. Right. So pulling all those 00:10:45.539 --> 00:10:47.799 rapid fire highlights together, what's the big 00:10:47.799 --> 00:10:50.840 picture takeaway? AI is evolving incredibly fast 00:10:50.840 --> 00:10:53.340 with both amazing potential and frankly, some 00:10:53.340 --> 00:10:56.659 pretty concerning impacts. OK, let's let's turn 00:10:56.659 --> 00:10:58.580 our attention now to something maybe even more 00:10:58.580 --> 00:11:02.450 fundamental. How we humans are actually interacting 00:11:02.450 --> 00:11:05.070 with AI day to day. Forget what you might have 00:11:05.070 --> 00:11:06.769 heard about people, you know, falling deeply 00:11:06.769 --> 00:11:08.710 in love with their chatbots or seeing them as 00:11:08.710 --> 00:11:11.649 digital best friends. A new report reveals a 00:11:11.649 --> 00:11:13.509 pretty surprising truth about how we're really 00:11:13.509 --> 00:11:16.320 using them. This last segment looks at that human 00:11:16.320 --> 00:11:19.460 -AI bond. Yeah, this is truly fascinating research. 00:11:20.019 --> 00:11:22.500 Anthropic, you know, the folks behind the Claude 00:11:22.500 --> 00:11:25.399 AI just put up this fresh report looking specifically 00:11:25.399 --> 00:11:28.200 at how people are actually using their AI assistant. 00:11:28.379 --> 00:11:31.059 They did this huge study analyzing, get this, 00:11:31.159 --> 00:11:34.299 4 .5 million conversations with Claude using 00:11:34.299 --> 00:11:37.000 their own in -house tool called Clio and the 00:11:37.000 --> 00:11:39.190 core finding. Despite all the stories we hear 00:11:39.190 --> 00:11:41.370 about loneliness and people wanting AI companions, 00:11:41.769 --> 00:11:44.330 emotional support is barely even a blip for most 00:11:44.330 --> 00:11:47.029 users. Really? That seems to go directly against 00:11:47.029 --> 00:11:49.529 that whole narrative of AI becoming a sort of 00:11:49.529 --> 00:11:51.649 pseudotherapist or even a romantic partner that 00:11:51.649 --> 00:11:53.409 we often hear about, you know, in pop culture 00:11:53.409 --> 00:11:55.929 and some media coverage. It absolutely does. 00:11:56.029 --> 00:11:59.080 They're data. And it's extensive, shows that 00:11:59.080 --> 00:12:03.019 a tiny 2 .9 % of all those millions of chats 00:12:03.019 --> 00:12:05.740 even touched on emotional support topics. Just 00:12:05.740 --> 00:12:09.240 2 .9%. And within that tiny little sliver, things 00:12:09.240 --> 00:12:11.700 like companionship and role play scenarios, they 00:12:11.700 --> 00:12:13.600 made up less than half a percent of the conversations. 00:12:13.919 --> 00:12:17.080 Less than 0 .5%. Anthropic actually characterizes 00:12:17.080 --> 00:12:20.519 the main user relationship with Claude as utilitarian. 00:12:20.759 --> 00:12:23.059 Utilitarian. Basically, yeah. Claude is acting 00:12:23.059 --> 00:12:25.440 more like a polite, super efficient co -worker 00:12:25.440 --> 00:12:27.860 or maybe a digital. research assistant than like 00:12:27.860 --> 00:12:30.799 a pretend soulmate or a deep confidant. People 00:12:30.799 --> 00:12:32.480 are using it to get stuff done, not really to 00:12:32.480 --> 00:12:33.779 share their deepest feelings. But it's a very 00:12:33.779 --> 00:12:36.220 practical, very functional, almost transactional 00:12:36.220 --> 00:12:38.399 relationship then. Precisely. It's about productivity, 00:12:38.600 --> 00:12:40.820 getting information. But here's where it gets 00:12:40.820 --> 00:12:43.860 kind of interesting. Even when users do open 00:12:43.860 --> 00:12:47.519 up emotionally, in those rare cases, the conversations 00:12:47.519 --> 00:12:50.320 usually end on a positive note. The overall sentiment 00:12:50.320 --> 00:12:53.139 tends to get better, not worse, as the chat goes 00:12:53.139 --> 00:12:56.019 on. So yeah, no Ayrshire movie moments where 00:12:56.019 --> 00:12:58.000 someone falls in love with their AI operating 00:12:58.000 --> 00:13:01.039 system. It seems users get the info or the resolution 00:13:01.039 --> 00:13:02.919 they were looking for, even if it was about a 00:13:02.919 --> 00:13:05.460 personal issue. Huh. That's a crucial nuance. 00:13:05.779 --> 00:13:08.240 And Anthropic also adds an important caveat about 00:13:08.240 --> 00:13:10.700 that positivity, don't they? Yes, a very important 00:13:10.700 --> 00:13:13.340 one. They explicitly state that they just don't 00:13:13.340 --> 00:13:15.100 know if that positivity they observed in the 00:13:15.100 --> 00:13:17.860 chat actually translates into real -world well 00:13:17.860 --> 00:13:21.179 -being for the user. There's no long term tracking 00:13:21.179 --> 00:13:23.559 of people's mental state or life outcomes after 00:13:23.559 --> 00:13:26.039 these chats. It's just based on the vibe observed 00:13:26.039 --> 00:13:28.419 in the conversation. So it's still very much 00:13:28.419 --> 00:13:30.039 a tool. And we really don't fully understand 00:13:30.039 --> 00:13:32.759 the psychological impact of these brief, maybe 00:13:32.759 --> 00:13:34.960 positive digital interactions over the long run. 00:13:35.059 --> 00:13:37.000 You know, I still wrestle with crump drift myself 00:13:37.000 --> 00:13:40.299 sometimes. That's where the AI starts to kind 00:13:40.299 --> 00:13:42.539 of. subtly lose the thread of what you originally 00:13:42.539 --> 00:13:45.460 asked it over a long conversation. The outputs 00:13:45.460 --> 00:13:48.179 get less accurate or relevant. Trying to get 00:13:48.179 --> 00:13:50.360 my AI tools to consistently understand what I 00:13:50.360 --> 00:13:52.399 mean turn after turn, it can be a real challenge. 00:13:52.480 --> 00:13:54.179 It's just a constant reminder that these are 00:13:54.179 --> 00:13:57.500 still tools. Powerful, yes, but definitely not 00:13:57.500 --> 00:14:01.039 sentient companions. So what does this anthropic 00:14:01.039 --> 00:14:03.519 report really tell us about the human AI bond 00:14:03.519 --> 00:14:05.779 as it stands right now? It's largely practical. 00:14:05.919 --> 00:14:08.340 Our relationships with AI are, for the most part, 00:14:08.419 --> 00:14:12.299 utilitarian. So trying to bring all these different 00:14:12.299 --> 00:14:14.840 threads together now, it seems pretty clear that 00:14:14.840 --> 00:14:16.879 despite all the immense hype around AI agents, 00:14:16.960 --> 00:14:19.320 their current capabilities, well, they have significant 00:14:19.320 --> 00:14:22.120 limitations. They're largely serving these utilitarian 00:14:22.120 --> 00:14:25.139 roles, right? Like code generation or very narrow 00:14:25.139 --> 00:14:27.519 workflow automation. They're not forming deep 00:14:27.519 --> 00:14:30.019 human connections or running entire departments 00:14:30.019 --> 00:14:32.889 on their own. And this means... Our current interactions 00:14:32.889 --> 00:14:35.590 with AI are far more practical, more functional 00:14:35.590 --> 00:14:37.529 than a lot of the narratives might lead us to 00:14:37.529 --> 00:14:40.509 believe. And this really highlights a critical 00:14:40.509 --> 00:14:43.129 ongoing need for all of us to distinguish between 00:14:43.129 --> 00:14:46.629 the marketing stories and the genuine AI capabilities. 00:14:47.740 --> 00:14:51.179 While AI is undeniably evolving at just breakneck 00:14:51.179 --> 00:14:53.159 speed, understanding its current limitations 00:14:53.159 --> 00:14:55.759 and its specific effective uses, that's really 00:14:55.759 --> 00:14:58.360 key to leveraging it well, both for us as individuals 00:14:58.360 --> 00:15:01.120 and for businesses. We need to cultivate realistic 00:15:01.120 --> 00:15:03.899 expectations to really harness its power and 00:15:03.899 --> 00:15:07.139 avoid costly mistakes or, frankly, just becoming 00:15:07.139 --> 00:15:09.039 a victim of all that agent washing we talked 00:15:09.039 --> 00:15:11.379 about. It's this ongoing tension, isn't it? You 00:15:11.379 --> 00:15:13.259 have the revolutionary potential, like those 00:15:13.259 --> 00:15:16.120 incredible creative tools we saw, that fashion 00:15:16.120 --> 00:15:18.740 -grade realism, balanced against really serious 00:15:18.740 --> 00:15:21.240 risks like the spread of AI -generated misinformation. 00:15:21.320 --> 00:15:23.740 It's just a dynamic, constantly shifting landscape. 00:15:24.139 --> 00:15:26.159 So what does this all mean for you listening 00:15:26.159 --> 00:15:29.120 right now? If current AI is indeed less Jarvis 00:15:29.120 --> 00:15:31.419 and maybe more like a polite, very capable, but 00:15:31.419 --> 00:15:34.019 still fundamentally just a co -worker, how might 00:15:34.019 --> 00:15:36.539 you adjust your own expectations? Or maybe your 00:15:36.539 --> 00:15:38.299 strategies for bringing it into your own work 00:15:38.299 --> 00:15:40.210 or your daily life? life? Do you focus it on 00:15:40.210 --> 00:15:43.009 hyper -specific tasks? Or maybe you remain cautious 00:15:43.009 --> 00:15:45.549 for now? Yeah, definitely stay curious, absolutely. 00:15:45.870 --> 00:15:48.710 But also, maybe critically evaluate the next 00:15:48.710 --> 00:15:51.629 really bold AI claim you hear. Take a moment, 00:15:51.730 --> 00:15:53.470 dig beneath the headlines if you can, look at 00:15:53.470 --> 00:15:55.509 the actual data when it's available, and just 00:15:55.509 --> 00:15:57.990 keep exploring this rapidly changing field for 00:15:57.990 --> 00:16:00.590 yourself. The real story of AI, it's always more 00:16:00.590 --> 00:16:03.129 nuanced, usually more complex, and often, honestly, 00:16:03.190 --> 00:16:05.639 more surprising than the headlines let on. Thank 00:16:05.639 --> 00:16:08.179 you for joining us on this deep dive. Out Hero 00:16:08.179 --> 00:16:08.460 Music.