WEBVTT 00:00:00.000 --> 00:00:03.339 It's sort of the great paradox of January 2026, 00:00:03.560 --> 00:00:05.719 isn't it? It really is. We have these incredible 00:00:05.719 --> 00:00:09.400 models, Dash GPT 5 .2, Gemini 3, and they can 00:00:09.400 --> 00:00:11.419 do amazing things. They can write poetry, they 00:00:11.419 --> 00:00:14.019 can code, they can even, you know, fake empathy 00:00:14.019 --> 00:00:16.260 pretty convincingly. Right. But then you put 00:00:16.260 --> 00:00:17.980 them in a digital office. You give them a Slack 00:00:17.980 --> 00:00:20.820 login, a Google Drive folder, and a basic analysis 00:00:20.820 --> 00:00:23.000 task. You just fall apart. It's like a nervous 00:00:23.000 --> 00:00:25.320 intern on their very first day. That's it. It 00:00:25.320 --> 00:00:28.000 is the intern paradox. And looking at the data 00:00:28.000 --> 00:00:30.050 we have today, it feels like the biggest reality 00:00:30.050 --> 00:00:31.949 check we've had in, what, the last two years? 00:00:32.090 --> 00:00:35.049 Oh, easily. Welcome back to The Deep Dive. It 00:00:35.049 --> 00:00:39.770 is Thursday, January 22nd, 2026. We're here to 00:00:39.770 --> 00:00:41.750 slow down, take a breath, and just try to make 00:00:41.750 --> 00:00:43.310 sense of the noise. And I'm here to help connect 00:00:43.310 --> 00:00:45.530 some of those dots. So today is really about 00:00:45.530 --> 00:00:47.789 taking a hard, honest look at where we actually 00:00:47.789 --> 00:00:50.869 are. We're going to unpack why this idea of the 00:00:50.869 --> 00:00:53.530 autonomous office agent just isn't working yet. 00:00:53.770 --> 00:00:56.450 We'll be looking at a new benchmark called Apex 00:00:56.450 --> 00:00:59.450 that is, frankly, a little embarrassing for the 00:00:59.450 --> 00:01:01.649 big labs. But we're also looking at the other 00:01:01.649 --> 00:01:04.870 side of that coin. Because while the brains of 00:01:04.870 --> 00:01:08.290 the AI are struggling with paperwork, the senses, 00:01:08.569 --> 00:01:11.409 I mean, vision, audio, they've just taken this 00:01:11.409 --> 00:01:14.390 quantum leap forward. Right. We're talking about 00:01:14.390 --> 00:01:17.769 Google's new 4D vision. Yeah. Pretty surreal 00:01:17.769 --> 00:01:20.390 story involving Liza Minnelli of all people. 00:01:20.510 --> 00:01:24.090 It is such a strange week. The tech is stalling 00:01:24.090 --> 00:01:26.510 in some areas and just hitting warp speed in 00:01:26.510 --> 00:01:28.569 others. Before we jumped into the data, though, 00:01:28.670 --> 00:01:30.810 I kind of have to make a vulnerable admission 00:01:30.810 --> 00:01:34.709 here. I honestly still wrestle with prompt drift 00:01:34.709 --> 00:01:37.349 myself. I was at my desk this morning trying 00:01:37.349 --> 00:01:39.969 to get an agent to format a report from three 00:01:39.969 --> 00:01:42.890 different documents. And I must have spent 45 00:01:42.890 --> 00:01:46.159 minutes just. tweaking instructions, you know, 00:01:46.180 --> 00:01:48.219 correcting its errors, telling it to stop making 00:01:48.219 --> 00:01:51.040 updates. And then it hit me. What's that? I could 00:01:51.040 --> 00:01:52.500 have just done the work myself in 10 minutes. 00:01:52.599 --> 00:01:54.760 Oh, absolutely. That is the automation tax. You 00:01:54.760 --> 00:01:56.900 pay it in patience just to prove the machine 00:01:56.900 --> 00:01:59.579 can do it. Exactly. And that feeling, it's not 00:01:59.579 --> 00:02:02.120 just me being impatient. It turns out there's 00:02:02.120 --> 00:02:04.920 hard data to back it up now. So let's start with 00:02:04.920 --> 00:02:07.920 segment one, the Apex failure. Yeah, this is 00:02:07.920 --> 00:02:09.419 a really significant report. It's called the 00:02:09.419 --> 00:02:13.060 AI gap. For, I'd say, the last 18 months, the 00:02:13.060 --> 00:02:16.680 narrative From OpenAI, from Google, from Anthropic 00:02:16.680 --> 00:02:20.000 has been pretty clear. AI is ready for real jobs. 00:02:20.280 --> 00:02:22.199 Exactly. We've been told they can be analysts, 00:02:22.560 --> 00:02:25.479 paralegals, executive assistants. Yeah, just 00:02:25.479 --> 00:02:27.840 give them the keys. Right. But we haven't really 00:02:27.840 --> 00:02:31.360 had a standardized way to test that claim. Not 00:02:31.360 --> 00:02:34.180 in a messy, real -world setting. That's the key. 00:02:34.280 --> 00:02:37.740 We had tests for write a poem or solve this math 00:02:37.740 --> 00:02:40.099 problem. We did not have tests for be an employee. 00:02:40.650 --> 00:02:42.669 And that's where the Apex Agents Benchmark comes 00:02:42.669 --> 00:02:45.629 in. So this is not a multiple choice test. No, 00:02:45.750 --> 00:02:47.990 not at all. They literally dropped the top models. 00:02:48.030 --> 00:02:50.810 We're talking Gemini 3 Flash, GPT 5 .2, Opus 00:02:50.810 --> 00:02:53.909 4 .5 into these simulated white collar workflows. 00:02:54.439 --> 00:02:57.139 So things like read the Slack thread, then cross 00:02:57.139 --> 00:02:59.099 -reference it with the PDF and Google Drive, 00:02:59.240 --> 00:03:01.840 and draft a reply to the client explaining why 00:03:01.840 --> 00:03:03.960 we can't do the refund. Precisely. The stuff 00:03:03.960 --> 00:03:06.000 that's actually, you know, 90 % of knowledge 00:03:06.000 --> 00:03:08.539 work. Yeah. It isn't just generating text. It's 00:03:08.539 --> 00:03:11.240 context switching. It's reasoning across different 00:03:11.240 --> 00:03:13.639 sources that might even contradict each other. 00:03:13.780 --> 00:03:16.879 And the results were? Well, I'm looking at this 00:03:16.879 --> 00:03:20.400 chart and shocking feels like an understatement. 00:03:20.539 --> 00:03:23.800 They were abysmal. So the winner was Gemini 3 00:03:23.800 --> 00:03:26.879 Flash. You want to take a guess at the accuracy 00:03:26.879 --> 00:03:30.099 rate? I mean, I would hope for at least a passing 00:03:30.099 --> 00:03:35.240 grade, maybe, what, 65 %? 24. 24%. That was the 00:03:35.240 --> 00:03:39.159 gold medal. GPT 5 .2 came in at 23%. And the 00:03:39.159 --> 00:03:41.979 others, like Opus 4 .5 and the standard Gemini 00:03:41.979 --> 00:03:44.740 3 Pro, were all hovering around 18%. That is 00:03:44.740 --> 00:03:47.460 just incredibly low. That means three out of 00:03:47.460 --> 00:03:50.120 four times the employee gets the task wrong. 00:03:50.360 --> 00:03:53.580 If I had a human intern with a 24 % success rate, 00:03:53.740 --> 00:03:55.340 I mean, I'd have to let them go before lunch. 00:03:55.919 --> 00:03:58.659 or they'd just quit. The report notes that in 00:03:58.659 --> 00:04:00.680 most cases, the models just couldn't piece the 00:04:00.680 --> 00:04:02.939 information together, they'd hallucinate a policy 00:04:02.939 --> 00:04:05.360 that didn't exist, or they'd miss a key detail 00:04:05.360 --> 00:04:07.419 in the Slack thread because it contradicted the 00:04:07.419 --> 00:04:09.500 PDF. It really feels like a connective tissue 00:04:09.500 --> 00:04:12.099 problem. They can process the individual data 00:04:12.099 --> 00:04:14.039 points, right? They can read the PDF, they can 00:04:14.039 --> 00:04:16.800 read the chat, but stringing them together into 00:04:16.800 --> 00:04:20.100 a coherent chain of logic, that's where it all 00:04:20.100 --> 00:04:22.259 breaks down. It is the perfect analogy for an 00:04:22.259 --> 00:04:24.100 intern. You ask an intern, what's the capital 00:04:24.100 --> 00:04:27.199 of France? They nail it. That's retrieval. Right. 00:04:27.300 --> 00:04:29.399 But if you ask them, look at these three messy 00:04:29.399 --> 00:04:31.699 files and tell me if we can legally fire this 00:04:31.699 --> 00:04:34.540 vendor, they panic. That's reasoning. It really 00:04:34.540 --> 00:04:37.540 makes you wonder about the timeline. We've all 00:04:37.540 --> 00:04:39.879 been preparing for this wave of autonomous agents 00:04:39.879 --> 00:04:44.259 to take over our inboxes in 2026. So looking 00:04:44.259 --> 00:04:47.000 at these numbers, I guess I have to ask, is the 00:04:47.000 --> 00:04:49.480 dream of the autonomous AI employee officially 00:04:49.480 --> 00:04:53.060 dead for now? Not dead, just delayed. They lack 00:04:53.060 --> 00:04:55.620 connective tissue reasoning. That distinction 00:04:55.620 --> 00:04:58.199 is key. But while the current models are struggling, 00:04:58.379 --> 00:05:00.040 the labs aren't just sitting still, are they? 00:05:00.139 --> 00:05:02.620 No, not at all. And this brings us to the leaks 00:05:02.620 --> 00:05:05.000 and the new tools emerging this week. The rumor 00:05:05.000 --> 00:05:07.639 mill is really spinning. We've got a leak about 00:05:07.639 --> 00:05:11.639 GPT -5 .3, which is codenamed Garlic. Garlic? 00:05:11.639 --> 00:05:13.899 That's an interesting choice. I know, right? 00:05:14.300 --> 00:05:17.230 But the leak suggests a pivot. Instead of just 00:05:17.230 --> 00:05:19.490 bigger is better, which has kind of been the 00:05:19.490 --> 00:05:22.089 strategy for five years, this next version is 00:05:22.089 --> 00:05:25.610 supposedly all about cost, speed, and crucially 00:05:25.610 --> 00:05:28.290 reasoning. Which would be a direct answer to 00:05:28.290 --> 00:05:31.170 the Apex failure. Exactly. If they can crack 00:05:31.170 --> 00:05:33.089 that reasoning bottleneck without making the 00:05:33.089 --> 00:05:36.569 model 10 times more expensive to run, well... 00:05:36.829 --> 00:05:38.870 That changes the entire equation. But while we're 00:05:38.870 --> 00:05:41.930 waiting for garlic to maybe save the day, there 00:05:41.930 --> 00:05:44.529 are some tools right now that are quietly solving 00:05:44.529 --> 00:05:47.189 this workflow problem just by changing how we 00:05:47.189 --> 00:05:49.170 interact with the info. You mentioned Notebook 00:05:49.170 --> 00:05:51.889 LM earlier. Yeah, Notebook LM really feels like 00:05:51.889 --> 00:05:53.930 a sleeper hit to me. Everyone is chasing the 00:05:53.930 --> 00:05:56.670 big agent dream. But Notebook LM has evolved 00:05:56.670 --> 00:05:58.970 into this amazing all -in -one researcher. It's 00:05:58.970 --> 00:06:01.269 gone from just a summarizer to a synthesizer. 00:06:01.629 --> 00:06:03.769 It has. The new workflows are really impressive. 00:06:03.910 --> 00:06:06.149 You can just dump in raw research papers, your 00:06:06.149 --> 00:06:08.629 notes, transcripts, and it doesn't just chat 00:06:08.629 --> 00:06:10.550 with you about them. It converts them. Right. 00:06:10.689 --> 00:06:13.610 I saw that workflow where it takes all that raw 00:06:13.610 --> 00:06:16.370 data and turns it into a slide deck outline, 00:06:16.730 --> 00:06:19.889 generates an audio overview, which sounds creepily 00:06:19.889 --> 00:06:22.350 human, and then it builds out comparison tables 00:06:22.350 --> 00:06:25.009 all in minutes. And that connects to this broader 00:06:25.009 --> 00:06:27.810 trend we're seeing, this idea of no PowerPoint. 00:06:28.250 --> 00:06:30.589 Please tell me that means what I think it means. 00:06:30.750 --> 00:06:32.810 Well. It means we stop fighting with PowerPoint. 00:06:33.170 --> 00:06:36.269 The shift is towards creating consistent, full 00:06:36.269 --> 00:06:39.430 slides using scripts. So you write the narrative 00:06:39.430 --> 00:06:42.089 and the AI builds the visual container for you. 00:06:42.149 --> 00:06:44.490 No more dragging text boxes or aligning fonts. 00:06:44.970 --> 00:06:47.230 Exactly. That's the dream. We become editors 00:06:47.230 --> 00:06:49.209 -in -chief instead of slide designers. Yeah. 00:06:49.290 --> 00:06:51.209 But there was another tool leaked from OpenAI 00:06:51.209 --> 00:06:54.790 that sounds a little more managerial. Salute. 00:06:54.949 --> 00:06:57.069 Salute is fascinating. It's an internal tool 00:06:57.069 --> 00:06:58.709 they're testing, and it's basically a project 00:06:58.709 --> 00:07:01.569 manager AI. You upload your files, you assign 00:07:01.569 --> 00:07:04.490 tasks to the AI, and it actually tracks the progress. 00:07:04.730 --> 00:07:07.149 So it's not just do this one thing. It's here's 00:07:07.149 --> 00:07:09.490 the project, now you manage the steps. It's trying 00:07:09.490 --> 00:07:12.629 to bridge that gap we saw in Apex. If the model 00:07:12.629 --> 00:07:15.589 can't reason across tasks on its own, maybe we 00:07:15.589 --> 00:07:17.910 just need a dedicated manager layer of software 00:07:17.910 --> 00:07:20.449 to force it to stay on track. Feels like a real 00:07:20.449 --> 00:07:22.850 shift in our role, you know? With tools like 00:07:22.850 --> 00:07:25.689 Salute and Notebook LM, are we actually working 00:07:25.689 --> 00:07:29.209 less or are we just managing the AI more? Managing 00:07:29.209 --> 00:07:31.769 more, but the output quality is exponentially 00:07:31.769 --> 00:07:35.009 higher. That feels like the trade -off of 2026. 00:07:35.329 --> 00:07:37.589 You're the conductor now, not the first violin. 00:07:37.790 --> 00:07:39.730 But let's talk about the business side of this. 00:07:39.889 --> 00:07:41.870 Because running these reasoning models, these 00:07:41.870 --> 00:07:45.259 manager layers, it is not free. Far from it. 00:07:45.319 --> 00:07:47.939 And the economics are getting pretty ugly. We 00:07:47.939 --> 00:07:50.220 saw a report on Anthropic, the makers of Claude. 00:07:50.379 --> 00:07:52.560 Their margins just took a really significant 00:07:52.560 --> 00:07:55.379 hit. Dropping from 50 percent down to 40. That's 00:07:55.379 --> 00:07:58.079 a huge slice of profit. It is. And the culprit 00:07:58.079 --> 00:08:01.459 is exactly who you'd expect. The cloud providers. 00:08:02.139 --> 00:08:04.980 Google and Amazon hiked their server costs by 00:08:04.980 --> 00:08:07.819 23 percent. It's the unsexy reality of the AI 00:08:07.819 --> 00:08:09.860 ecosystem, isn't it? You can have the smartest 00:08:09.860 --> 00:08:11.360 model in the world, but if you have to pay a 00:08:11.360 --> 00:08:13.889 landlord to run it, you're at their mercy. It's 00:08:13.889 --> 00:08:16.610 the infrastructure squeeze, and it's driving 00:08:16.610 --> 00:08:19.610 these massive capital deals. We just saw the 00:08:19.610 --> 00:08:22.850 Saudi infrastructure fund team up with Humane 00:08:22.850 --> 00:08:27.410 for a $1 .2 billion deal. Just to build data 00:08:27.410 --> 00:08:30.129 centers. $1 .2 billion just for the plumbing. 00:08:30.269 --> 00:08:32.870 For the plumbing. But, you know, while the infrastructure 00:08:32.870 --> 00:08:34.710 is getting more expensive and the reasoning is 00:08:34.710 --> 00:08:37.289 hitting a wall, the creative output is finding 00:08:37.289 --> 00:08:39.629 these really interesting new business models. 00:08:39.750 --> 00:08:42.730 This Eleven Labs story really caught my eye. 00:08:42.950 --> 00:08:44.929 It's a landmark moment, I think. They dropped 00:08:44.929 --> 00:08:48.970 a 13 -track AI music album. And we're not talking 00:08:48.970 --> 00:08:51.210 about some, you know, anonymous generated lo 00:08:51.210 --> 00:08:54.299 -fi beats. This features Liza Minnelli and Art 00:08:54.299 --> 00:08:56.659 Garfunkel. Now, to be clear, because the ethics 00:08:56.659 --> 00:08:58.879 here usually get messy, Liza Minnelli did not 00:08:58.879 --> 00:09:00.960 go into a booth and record this, correct? No, 00:09:00.980 --> 00:09:03.039 she did not. But her state and her team signed 00:09:03.039 --> 00:09:05.899 off on it. And that's the key. Yes. Full royalties 00:09:05.899 --> 00:09:07.879 are being paid. The labels are actually involved. 00:09:08.100 --> 00:09:10.139 It's a licensed collaboration. They're essentially 00:09:10.139 --> 00:09:13.399 treating her voice as an instrument, like a Stradivarius, 00:09:13.580 --> 00:09:16.360 that can be played by the AI with the artist's 00:09:16.360 --> 00:09:19.730 permission. That's a complete 180 from all the 00:09:19.730 --> 00:09:23.350 lawsuits we saw back in 2024 and 2025. It suggests 00:09:23.350 --> 00:09:26.350 a future where an artist's prime voice can become 00:09:26.350 --> 00:09:29.950 immortal. Exactly. And Spotify is leaning into 00:09:29.950 --> 00:09:32.669 this whole vibe culture, too. Their new prompted 00:09:32.669 --> 00:09:35.629 playlists are getting shockingly good. You don't 00:09:35.629 --> 00:09:37.669 search for a genre anymore. You just type in, 00:09:37.710 --> 00:09:40.110 make me a playlist that feels like rainy Tokyo 00:09:40.110 --> 00:09:42.889 nights. And it actually understands the semantic 00:09:42.889 --> 00:09:45.830 texture of that request. It's moving from metadata 00:09:45.830 --> 00:09:49.409 like this is a rock song to emotional data. This 00:09:49.409 --> 00:09:52.110 song feels like heartbreak. That's it. So does 00:09:52.110 --> 00:09:54.330 the Eleven Labs album prove that artists and 00:09:54.330 --> 00:09:57.710 AI can actually coexist profitably? Yes. It turns 00:09:57.710 --> 00:10:00.409 legacy voices into a scalable, renewable asset. 00:10:00.629 --> 00:10:02.990 A renewable asset. That is a wild way to think 00:10:02.990 --> 00:10:04.940 about a human voice. But speaking of sensing 00:10:04.940 --> 00:10:06.639 the world, we have to talk about the breakthrough 00:10:06.639 --> 00:10:08.200 of the week. This is the one that actually made 00:10:08.200 --> 00:10:10.259 me just stop and stare at my screen. Google's 00:10:10.259 --> 00:10:13.919 D4RT. Right. Dynamic 4D reconstruction and tracking. 00:10:14.279 --> 00:10:16.879 Now, we've had computer vision for a while. My 00:10:16.879 --> 00:10:19.600 car sees lane lines. My phone recognizes my face. 00:10:19.940 --> 00:10:22.559 How is this different? This is the difference 00:10:22.559 --> 00:10:24.600 between looking at a photograph and actually 00:10:24.600 --> 00:10:27.279 stepping inside the room. Traditional computer 00:10:27.279 --> 00:10:30.720 vision looks at a 2D image and draws boxes around 00:10:30.720 --> 00:10:34.840 things. Cats, cars. Right. D4RT watches a video 00:10:34.840 --> 00:10:37.559 and builds a world model. A world model. Yes. 00:10:37.639 --> 00:10:40.019 It tracks every single pixel across time. That's 00:10:40.019 --> 00:10:42.320 the fourth dimension here, time. And from that 00:10:42.320 --> 00:10:45.179 flat video, it reconstructs a full 3D scene. 00:10:45.419 --> 00:10:47.940 So if I showed a video of me walking around my 00:10:47.940 --> 00:10:50.559 kitchen, it's not just seeing a man in a kitchen. 00:10:50.600 --> 00:10:53.730 It's actually... Building the 3D geometry of 00:10:53.730 --> 00:10:56.629 the fridge, the table, the coffee cup. And it 00:10:56.629 --> 00:10:58.470 understands exactly where the camera is moving 00:10:58.470 --> 00:11:01.070 in that space. Correct. And it handles the messy 00:11:01.070 --> 00:11:03.509 stuff that usually breaks these models. Motion 00:11:03.509 --> 00:11:06.009 blur, things getting blocked. What we call occlusion. 00:11:06.250 --> 00:11:08.389 Ah. Like if you walk behind the fridge, old models 00:11:08.389 --> 00:11:11.789 would think you just vanished. D4NT knows you 00:11:11.789 --> 00:11:13.289 didn't disappear. It knows you're just behind 00:11:13.289 --> 00:11:15.889 an object. Okay. And the speed. That is the whoa 00:11:15.889 --> 00:11:18.669 moment. Previous models, and this was cutting 00:11:18.669 --> 00:11:21.320 edge just six months ago. would take maybe 10 00:11:21.320 --> 00:11:23.320 minutes to parse a video like that. 10 minutes, 00:11:23.440 --> 00:11:26.039 okay. I4RT does a one -minute video in five seconds. 00:11:26.299 --> 00:11:29.159 Two -sec silence, five seconds. Whoa. That's 00:11:29.159 --> 00:11:31.860 basically real time. It's anywhere from 18 to 00:11:31.860 --> 00:11:34.539 300 times faster than anything else out there. 00:11:35.139 --> 00:11:37.860 It uses this unified query system. So instead 00:11:37.860 --> 00:11:39.960 of having one model track objects, another one 00:11:39.960 --> 00:11:43.179 gets the depth, another one map the room, D4RT 00:11:43.179 --> 00:11:45.740 does it all in a single pass. I mean, imagine 00:11:45.740 --> 00:11:48.879 the implications for... Well, for everything. 00:11:49.259 --> 00:11:51.679 Robotics, obviously. If a robot can understand 00:11:51.679 --> 00:11:54.480 the 3D geometry of a room in milliseconds, it 00:11:54.480 --> 00:11:57.019 can navigate like a human. But also surveillance, 00:11:57.259 --> 00:11:59.240 content creation. It's like the AI isn't just 00:11:59.240 --> 00:12:01.299 watching a movie anymore. It's building the movie. 00:12:01.659 --> 00:12:04.519 set in its head instantly. It's superhuman perception. 00:12:04.720 --> 00:12:06.440 We're giving machines a spatial understanding 00:12:06.440 --> 00:12:08.840 of reality that might actually be faster than 00:12:08.840 --> 00:12:11.059 our own ability to process a scene. It's a fundamental 00:12:11.059 --> 00:12:13.600 shift. So if machines can reconstruct reality 00:12:13.600 --> 00:12:17.000 this fast, what happens to truth in video evidence? 00:12:17.379 --> 00:12:19.419 We stop trusting our eyes and start trusting 00:12:19.419 --> 00:12:21.779 digital watermarks. That is a chilling thought. 00:12:22.240 --> 00:12:24.659 We're going to take a quick breather here. We've 00:12:24.659 --> 00:12:26.659 covered the failures of the office intern and 00:12:26.659 --> 00:12:29.559 the superhuman vision of Google. When we come 00:12:29.559 --> 00:12:32.600 back, we are going to look at the Empowered Utility 00:12:32.600 --> 00:12:35.360 Belt, the specific tools you can use right now 00:12:35.360 --> 00:12:38.379 to make your life a little easier. Stay with 00:12:38.379 --> 00:12:41.799 us, and we are back. Okay, let's get tactical. 00:12:42.080 --> 00:12:44.639 We've talked about the big models, the high -level 00:12:44.639 --> 00:12:47.460 concepts, but what about the tools that actually 00:12:47.460 --> 00:12:50.419 save you time on a Tuesday afternoon? We called 00:12:50.419 --> 00:12:53.539 this section the Empowered Utility Belt. Yeah, 00:12:53.580 --> 00:12:55.860 there are some great ones this week. First up 00:12:55.860 --> 00:12:58.639 is something called Clawed Cowork. This one is 00:12:58.639 --> 00:13:00.340 interesting because it's not a new app. It's 00:13:00.340 --> 00:13:02.820 more of a workflow, right? Right. It's a 13 -minute 00:13:02.820 --> 00:13:05.779 exercise, and it's designed to turn Claude from 00:13:05.779 --> 00:13:08.000 just a chatbot into a real thinking partner. 00:13:08.399 --> 00:13:10.759 It's all about priming the model with your context 00:13:10.759 --> 00:13:13.200 so you don't have to explain your job every single 00:13:13.200 --> 00:13:15.860 time you open a new chat. I love that. It's like 00:13:15.860 --> 00:13:18.200 onboarding your AI colleague one time instead 00:13:18.200 --> 00:13:20.639 of every single morning. What else you got? ChartGen 00:13:20.639 --> 00:13:23.519 AI. This one is for anyone who just drowns in 00:13:23.519 --> 00:13:26.259 data. It takes raw data from pretty much anywhere. 00:13:26.700 --> 00:13:30.600 Facebook ad exports, TikTok analytics, just messy 00:13:30.600 --> 00:13:33.220 Excel sheets. And it turns them into professional 00:13:33.220 --> 00:13:37.320 charts in seconds. So no more fighting with Excel 00:13:37.320 --> 00:13:39.559 pivot tables. There are more pivot tables. And 00:13:39.559 --> 00:13:41.960 then there's Locate Store. It's very specific, 00:13:42.039 --> 00:13:44.720 but very cool. You have a Google sheet full of 00:13:44.720 --> 00:13:47.519 addresses. It instantly turns that into an interactive 00:13:47.519 --> 00:13:50.500 map with search filters. That's huge for logistics 00:13:50.500 --> 00:13:52.720 or even just planning a trip. But the one that 00:13:52.720 --> 00:13:55.379 really caught my eye for pure automation is Demonstrate. 00:13:55.850 --> 00:13:57.769 Okay, how does that work? You record a browser 00:13:57.769 --> 00:14:00.769 task one time. So say, log into this portal, 00:14:00.929 --> 00:14:03.289 download the invoice, rename it, and upload it 00:14:03.289 --> 00:14:06.029 to Dropbox. You just do it once, Demonstrate 00:14:06.029 --> 00:14:08.870 records it, and then this is the kicker, it deploys 00:14:08.870 --> 00:14:11.110 it as serverless code. So you don't just get 00:14:11.110 --> 00:14:12.889 a macro, you get a piece of software that runs 00:14:12.889 --> 00:14:14.990 in the cloud for you. Exactly. You are literally 00:14:14.990 --> 00:14:17.970 programming by doing. And finally, we have Calum. 00:14:18.320 --> 00:14:20.740 Right, the AI calendar assistant. But it handles 00:14:20.740 --> 00:14:23.019 the hard stuff like multi -person meetings, shared 00:14:23.019 --> 00:14:25.100 availability, all the real world constraints. 00:14:25.840 --> 00:14:28.460 It's really trying to kill the email ping pong 00:14:28.460 --> 00:14:31.840 of scheduling. Of all of these tools, if you 00:14:31.840 --> 00:14:34.480 had to pick one, which of these actually saves 00:14:34.480 --> 00:14:36.779 you an hour of sleep tonight? Demonstrate automating 00:14:36.779 --> 00:14:39.179 browser tasks is the ultimate cure for boredom. 00:14:39.259 --> 00:14:41.080 I am with you on that one. Anything that can 00:14:41.080 --> 00:14:43.100 fill out a web form for me is a friend of mine. 00:14:43.240 --> 00:14:46.960 Amen to that. So let's zoom out. What does all 00:14:46.960 --> 00:14:49.820 of this really mean for you, the listener, sitting 00:14:49.820 --> 00:14:53.580 here in January 2026? I think we're in what you 00:14:53.580 --> 00:14:56.320 could call a gap year. A gap year. Explain that. 00:14:56.500 --> 00:14:59.440 Well, the text -based agents, the brains, they're 00:14:59.440 --> 00:15:02.779 struggling. GPT -5 .2, Gemini 3, they're getting 00:15:02.779 --> 00:15:05.639 barely 24 % accuracy on real work. The brain 00:15:05.639 --> 00:15:07.659 is still learning how to file paperwork. Okay. 00:15:07.929 --> 00:15:10.269 But while the brain is stumbling, the senses 00:15:10.269 --> 00:15:13.889 are just exploding. Exactly. Google's D4RT vision 00:15:13.889 --> 00:15:16.309 is seeing the world 300 times faster than before. 00:15:17.210 --> 00:15:19.470 Eleven Labs has solved the legal and creative 00:15:19.470 --> 00:15:22.610 puzzle of AI music. So the synthesis here is, 00:15:22.750 --> 00:15:26.570 the AI might not be ready to be your lawyer or 00:15:26.570 --> 00:15:29.210 your project manager just yet, but it can see 00:15:29.210 --> 00:15:31.710 your world and sing your songs better than ever 00:15:31.710 --> 00:15:35.129 before. The brain is lagging, but the eyes and 00:15:35.129 --> 00:15:37.409 ears have become superhuman. That's a really 00:15:37.409 --> 00:15:39.820 powerful place to be. It means we have to stop 00:15:39.820 --> 00:15:42.639 waiting for some magic general manager AI and 00:15:42.639 --> 00:15:45.399 just start using the super sensory AI tools we 00:15:45.399 --> 00:15:47.740 have right now. Use the vision, use the voice, 00:15:47.879 --> 00:15:50.460 use the automation. Don't wait for the reasoning 00:15:50.460 --> 00:15:52.440 to be perfect. And if you want to start somewhere 00:15:52.440 --> 00:15:54.600 practical, I would highly encourage you to try 00:15:54.600 --> 00:15:57.159 that Claude Cowork 13 -minute exercise we mentioned. 00:15:57.360 --> 00:15:59.659 It's a small investment that really pays off 00:15:59.659 --> 00:16:02.120 every time you open that chat window. It really 00:16:02.120 --> 00:16:03.799 does change the dynamic. We're going to leave 00:16:03.799 --> 00:16:06.279 you with one final thought to mull over. Yeah, 00:16:06.320 --> 00:16:09.980 we talked about Google's D4RT, how it reconstructs 00:16:09.980 --> 00:16:12.879 a 3D world from a flat video in just five seconds. 00:16:12.980 --> 00:16:15.159 It's incredible. So here's the question. If an 00:16:15.159 --> 00:16:17.940 AI can reconstruct a 3D world from a flat video 00:16:17.940 --> 00:16:21.600 in five seconds, how long until it can reconstruct 00:16:21.600 --> 00:16:23.919 a better version of your workflow than you can? 00:16:24.139 --> 00:16:26.039 Something to think about. Thanks for diving in 00:16:26.039 --> 00:16:27.139 with us. We'll see you next time.