WEBVTT 00:00:00.000 --> 00:00:02.779 We often talk about AI, like it's just one huge 00:00:02.779 --> 00:00:06.099 thing, right? One big brain. But there's actually 00:00:06.099 --> 00:00:09.119 a massive difference, a really functional difference 00:00:09.119 --> 00:00:12.740 between, say, an AI that's read everything and 00:00:12.740 --> 00:00:15.519 can only talk about a solution versus one that 00:00:15.519 --> 00:00:18.780 can actually do the task for you. Right. It's 00:00:18.780 --> 00:00:21.399 pure theory versus action. It's like the shift 00:00:21.399 --> 00:00:24.420 from asking a large language model, an LLM, how 00:00:24.420 --> 00:00:27.339 to book a cheap flight to telling a language 00:00:27.339 --> 00:00:30.059 action model. a lamb, hey, book the cheapest 00:00:30.059 --> 00:00:33.079 flight, email me the receipt. Beat. That's the 00:00:33.079 --> 00:00:35.579 jump from talking to actually doing. Welcome 00:00:35.579 --> 00:00:37.840 to the deep dive. Yeah. And look, if you're starting 00:00:37.840 --> 00:00:40.280 to feel a bit lost in all the acronyms flying 00:00:40.280 --> 00:00:43.340 around, you're definitely not alone. Our mission 00:00:43.340 --> 00:00:46.619 today really is to cut through that noise, give 00:00:46.619 --> 00:00:49.000 you a kind of shortcut to understanding the different 00:00:49.000 --> 00:00:52.119 specializations. We need the AI toolbox. Exactly. 00:00:52.219 --> 00:00:54.810 Think about it, right? You wouldn't use a hammer 00:00:54.810 --> 00:00:57.270 to turn a screw. Doesn't matter how great the 00:00:57.270 --> 00:00:59.929 hammer is. We need to know the right tool for 00:00:59.929 --> 00:01:03.350 the job. So today we're diving into eight key 00:01:03.350 --> 00:01:06.450 AI model types. We'll look at what they do, their 00:01:06.450 --> 00:01:09.090 main uses, and yeah, the important watch outs 00:01:09.090 --> 00:01:11.549 for each. Okay. Let's start with the foundation. 00:01:11.769 --> 00:01:14.430 Probably the one most people know. The LM. Okay, 00:01:14.450 --> 00:01:16.750 let's unpack this. The LLM, the Large Language 00:01:16.750 --> 00:01:19.400 Model. This is... like the definition model. 00:01:19.519 --> 00:01:21.400 It's the most famous one out there. Think of 00:01:21.400 --> 00:01:23.239 it like that super smart friend who's basically 00:01:23.239 --> 00:01:26.019 read every book, every website, you know, Chet, 00:01:26.120 --> 00:01:29.780 GPT, Gemini, Claude, those ones. Yeah, and they're 00:01:29.780 --> 00:01:32.319 large because the amount of text they train on 00:01:32.319 --> 00:01:35.060 is just astronomical. They absorb all these patterns 00:01:35.060 --> 00:01:37.480 of language, but how they work, it's actually 00:01:37.480 --> 00:01:40.219 pretty simple at its core. They're mostly just 00:01:40.219 --> 00:01:41.939 next -word guessers. They don't actually know 00:01:41.939 --> 00:01:44.459 truth. Right. They break down sentences into 00:01:44.459 --> 00:01:46.939 tokens, sort of like pieces of words. Exactly. 00:01:47.040 --> 00:01:49.159 Pieces of words or even punctuation. And then 00:01:49.159 --> 00:01:51.870 they just do the math. calculating the most likely 00:01:51.870 --> 00:01:55.069 next token, the next piece, based on, well, billions 00:01:55.069 --> 00:01:56.870 of examples they've seen. If you start with, 00:01:56.969 --> 00:01:58.450 I need to put this on there. I guess a shelf 00:01:58.450 --> 00:02:01.010 or table or whatever fits statistically. Yeah. 00:02:01.290 --> 00:02:04.250 And this guessing ability makes them really good 00:02:04.250 --> 00:02:07.390 for tech stuff. Writing content, sure. Blog posts, 00:02:07.629 --> 00:02:11.169 emails, summarizing huge articles down to like 00:02:11.169 --> 00:02:13.430 three bullet points. Oh, yeah. Summarizing is 00:02:13.430 --> 00:02:16.909 huge and helping programmers, right? Explaining 00:02:16.909 --> 00:02:20.810 tricky code, debugging. Yeah. Even translation 00:02:20.810 --> 00:02:23.629 or customer service chat bots. But that's also 00:02:23.629 --> 00:02:25.949 where prompting gets tricky. Yeah. And honestly, 00:02:26.490 --> 00:02:28.849 I still wrestle with getting problems just right 00:02:28.849 --> 00:02:31.409 myself sometimes. It's not always easy. You really 00:02:31.409 --> 00:02:33.629 see the difference between just asking vaguely. 00:02:33.759 --> 00:02:36.240 versus being really specific? Totally. Like, 00:02:36.259 --> 00:02:38.419 a bad prompt is just, write an email to work 00:02:38.419 --> 00:02:41.259 from home, that's it. Vague. Super vague. A good 00:02:41.259 --> 00:02:43.300 prompt gives the goal, maybe the specific days 00:02:43.300 --> 00:02:45.120 you want off, the reasons why, like, needing 00:02:45.120 --> 00:02:48.060 deep work focus, and even the tone. Make it polite, 00:02:48.139 --> 00:02:50.159 professional, but ask firmly. You've got to guide 00:02:50.159 --> 00:02:52.060 the guesser. And this leads to the big watch 00:02:52.060 --> 00:02:54.180 out, right? The whole hallucination problem. 00:02:54.479 --> 00:02:56.800 Because they only guess what looks right statistically. 00:02:56.960 --> 00:02:59.400 They can just confidently make stuff up. Fake 00:02:59.400 --> 00:03:02.159 studies, incorrect historical facts, citations 00:03:02.159 --> 00:03:04.759 that lead nowhere. pure invention, but sounds 00:03:04.759 --> 00:03:08.240 plausible. So if an LLM is fundamentally a great 00:03:08.240 --> 00:03:10.840 guesser, what's the single most critical thing 00:03:10.840 --> 00:03:13.219 a user has to remember about its output? You 00:03:13.219 --> 00:03:16.219 absolutely must check important facts. The AI 00:03:16.219 --> 00:03:18.719 only guesses what looks right. It doesn't verify. 00:03:19.080 --> 00:03:22.219 Good. OK, so LLMs are great text guessers, but 00:03:22.219 --> 00:03:25.919 they can be slow and Well, inventive. Yeah. What 00:03:25.919 --> 00:03:28.360 if we need pictures and we need them fast? Ah, 00:03:28.400 --> 00:03:30.819 speed. That takes us to the LCM, right? Exactly. 00:03:31.159 --> 00:03:34.400 The latent consistency model, LCM. Think of it 00:03:34.400 --> 00:03:37.080 as the visual cousin to the LLM, but its whole 00:03:37.080 --> 00:03:40.099 reason for being is speed in making images. Much 00:03:40.099 --> 00:03:42.259 faster than older models. Yeah. If you've ever 00:03:42.259 --> 00:03:45.539 waited like 30 seconds or more for an AI image, 00:03:46.319 --> 00:03:48.840 that old model like stable diffusion was working 00:03:48.840 --> 00:03:52.669 slowly. taking maybe 50, even 100 steps to refine 00:03:52.669 --> 00:03:54.469 the image from noise like a really carol paint 00:03:54.469 --> 00:03:57.189 party adding tiny details. Right. The LCM, though, 00:03:57.270 --> 00:03:58.830 is more like a master painter who's done this 00:03:58.830 --> 00:04:01.270 a million times. It learns these underlying patterns 00:04:01.270 --> 00:04:03.750 of the latent consistency and figures out how 00:04:03.750 --> 00:04:06.150 to jump huge steps. Goes from step one, maybe 00:04:06.150 --> 00:04:08.509 straight, to step 25. Whoa. So it can generate 00:04:08.509 --> 00:04:10.750 really decent images in just like two to eight 00:04:10.750 --> 00:04:14.689 steps total. Two to eight. Compared to 50. That's 00:04:14.689 --> 00:04:17.189 a huge difference. Massive difference. And that 00:04:17.189 --> 00:04:19.470 speed opens up doors for anything real time. 00:04:19.709 --> 00:04:22.769 Think AR, VR apps. The scene needs to update 00:04:22.769 --> 00:04:24.490 instantly when you turn your head, right? Right, 00:04:24.670 --> 00:04:26.610 yeah. No lag. Or making pictures right there 00:04:26.610 --> 00:04:28.709 on your phone without needing to send data to 00:04:28.709 --> 00:04:31.839 the cloud and wait. fast design tools, improving 00:04:31.839 --> 00:04:34.579 video call quality in real time. OK, so speed's 00:04:34.579 --> 00:04:36.779 the big win. What's the catch? Well, the trade 00:04:36.779 --> 00:04:38.699 -off for skipping all those steps is sometimes 00:04:38.699 --> 00:04:41.620 the images might be a bit less detailed, maybe 00:04:41.620 --> 00:04:44.120 a little too smooth compared to the slower methods. 00:04:44.500 --> 00:04:46.879 But for many uses, that near -instant result 00:04:46.879 --> 00:04:50.139 is what matters most. So why is that master painter 00:04:50.139 --> 00:04:52.920 approach, that speed, so crucial for something 00:04:52.920 --> 00:04:55.769 like a VR headset? Because everything has to 00:04:55.769 --> 00:04:58.329 update instantly as you move. Lag breaks the 00:04:58.329 --> 00:05:01.009 whole illusion of immersion. Makes sense. Okay, 00:05:01.009 --> 00:05:03.569 so we have LLMs for talking, LCMs for fast painting, 00:05:03.889 --> 00:05:05.949 but they're both creating things. What if we 00:05:05.949 --> 00:05:08.829 need the AI to actually, you know, do a task? 00:05:09.110 --> 00:05:11.730 A multi -step task. Ah, now we get to the LAM. 00:05:11.930 --> 00:05:14.430 The LAM. Language Action Model. This is that 00:05:14.430 --> 00:05:16.110 critical shift you mentioned from talking to 00:05:16.110 --> 00:05:18.470 doing. Exactly. If the LLM is the smart friend 00:05:18.470 --> 00:05:21.350 who just talks, the LAM is like the project manager. 00:05:21.689 --> 00:05:24.439 It takes action. How does it do that? Is it just 00:05:24.439 --> 00:05:28.180 a better LLM? Not quite. It's more like a system. 00:05:28.339 --> 00:05:30.399 It starts with an LLM. Yeah, that's the brain 00:05:30.399 --> 00:05:32.480 for understanding your request. But then it adds 00:05:32.480 --> 00:05:35.620 memory like a notebook, a planner to map out 00:05:35.620 --> 00:05:39.319 the steps. And the really crucial part, the ability 00:05:39.319 --> 00:05:41.899 to use tools. This means it can connect to other 00:05:41.899 --> 00:05:44.339 things, your email, a web browser, calendars, 00:05:44.439 --> 00:05:47.639 using APIs. Ah, APIs. That's the key difference 00:05:47.639 --> 00:05:49.879 then, the connection to other services. That's 00:05:49.879 --> 00:05:52.000 the game changer. So an LLM tells you how to 00:05:52.000 --> 00:05:55.420 book a flight. A LAM connects to, say, Expedia's 00:05:55.420 --> 00:05:57.620 API, searches flights based on your criteria, 00:05:58.100 --> 00:06:00.779 finds the cheapest one, books it, and then asks 00:06:00.779 --> 00:06:04.040 you to confirm. Wow. OK. That sounds like the 00:06:04.040 --> 00:06:06.220 future of automation, really. And it is. Think, 00:06:06.319 --> 00:06:09.100 AI agents, these things can handle complex sequences. 00:06:09.399 --> 00:06:11.620 Like, you could tell it. Find 10 potential customers 00:06:11.620 --> 00:06:13.959 in the software industry in California, draft 00:06:13.959 --> 00:06:16.639 a personalized intro email based on our new product 00:06:16.639 --> 00:06:20.170 launch, and show me the drafts. A lot more sophisticated 00:06:20.170 --> 00:06:22.769 than just asking a question. Definitely. Advanced 00:06:22.769 --> 00:06:25.069 customer support, too, actually processing a 00:06:25.069 --> 00:06:27.310 refund, not just answering questions about the 00:06:27.310 --> 00:06:30.329 policy or just digital assistance that can actually 00:06:30.329 --> 00:06:32.410 turn on your lights or set alarms because they 00:06:32.410 --> 00:06:35.509 can connect to those systems. So what core function 00:06:35.509 --> 00:06:38.910 lets a lamb move beyond just conversation and 00:06:38.910 --> 00:06:41.389 actually access, say, your calendar or email? 00:06:41.670 --> 00:06:43.810 It's that ability to connect directly to other 00:06:43.810 --> 00:06:46.569 tools and services using APIs. Those are the 00:06:46.569 --> 00:06:49.379 hands and feet. Got it. But building one giant 00:06:49.379 --> 00:06:51.800 lamb that can do everything sounds incredibly 00:06:51.800 --> 00:06:55.839 complex and frankly expensive to run. It absolutely 00:06:55.839 --> 00:06:58.240 is. Which leads us perfectly to the next model 00:06:58.240 --> 00:07:00.819 type, which is all about efficiency. The MOE. 00:07:01.259 --> 00:07:03.540 MOE, mixture of experts. Yeah, mixture of experts. 00:07:03.639 --> 00:07:06.500 Now, this isn't one single giant model. It's 00:07:06.500 --> 00:07:09.319 actually a whole system. Think of lots of smaller... 00:07:09.399 --> 00:07:11.680 specialized models, the experts. And they're 00:07:11.680 --> 00:07:13.680 all managed by another AI called the router. 00:07:13.980 --> 00:07:15.839 OK, so instead of one generalist, it's like having 00:07:15.839 --> 00:07:18.100 a team of specialists. Exactly, like asking a 00:07:18.100 --> 00:07:21.300 specific group of pros. The router looks at your 00:07:21.300 --> 00:07:23.819 request, your prompt, and figures out, OK, this 00:07:23.819 --> 00:07:25.860 is about coding, and sends it only to the coding 00:07:25.860 --> 00:07:29.860 expert. Or this needs history knowledge, so it 00:07:29.860 --> 00:07:32.000 activates the history expert. Usually just one 00:07:32.000 --> 00:07:34.079 or two experts get activated for any given task. 00:07:34.180 --> 00:07:35.819 Right, so you're not running the whole giant 00:07:35.819 --> 00:07:37.980 brain all the time. Precisely. And here's the 00:07:37.980 --> 00:07:40.240 really cool part. the efficiency and size benefit. 00:07:40.540 --> 00:07:43.899 You can build a massive model overall, let's 00:07:43.899 --> 00:07:46.079 say, a trillion parameters worth of knowledge. 00:07:46.259 --> 00:07:50.120 But because the router only turns on maybe 5 00:07:50.120 --> 00:07:53.240 % or 10 % of those experts for any single query, 00:07:54.120 --> 00:07:56.180 you're only paying the computational cost of 00:07:56.180 --> 00:07:59.079 running a much smaller model, maybe just 100 00:07:59.079 --> 00:08:00.980 billion parameters in that example. Wait, hold 00:08:00.980 --> 00:08:03.259 on. So you get the knowledge of a potentially 00:08:03.259 --> 00:08:06.389 enormous model. but the running cost of something 00:08:06.389 --> 00:08:09.329 much smaller. That's the magic. Wow. I mean, 00:08:09.329 --> 00:08:11.290 that just fundamentally changes the economics 00:08:11.290 --> 00:08:13.769 of building top -tier AI, doesn't it? That scaling 00:08:13.769 --> 00:08:16.730 efficiency is incredible. It's huge. That's why 00:08:16.730 --> 00:08:19.730 this architecture, MoE, is believed to be used 00:08:19.730 --> 00:08:22.149 in the really big, powerful models like Mixed 00:08:22.149 --> 00:08:25.490 Roll and probably GPT -4, too. It's also great 00:08:25.490 --> 00:08:27.430 for companies wanting to personalize. They can 00:08:27.430 --> 00:08:29.629 add their own private experts, train just on 00:08:29.629 --> 00:08:31.829 their internal documents or specific industry 00:08:31.829 --> 00:08:34.840 regulations. So you could ask it to explain I 00:08:34.840 --> 00:08:37.299 don't know, inflation using general knowledge. 00:08:37.340 --> 00:08:40.100 Yeah. And write Python code to calculate CPI 00:08:40.100 --> 00:08:42.639 using a coding expert, all in one go. Exactly. 00:08:42.860 --> 00:08:45.240 Dual knowledge handled efficiently. OK, the efficiency 00:08:45.240 --> 00:08:48.039 is amazing. But beyond cost, what's the core 00:08:48.039 --> 00:08:50.960 risk if that router AI messes up its job in a 00:08:50.960 --> 00:08:53.440 Moe system? If the router picks the wrong expert, 00:08:53.679 --> 00:08:55.980 the whole answer will likely be poor quality 00:08:55.980 --> 00:08:58.940 or just plain wrong. Bad robbing means bad results. 00:08:59.100 --> 00:09:01.399 Makes sense. Garbage in, garbage out. Or rather, 00:09:01.600 --> 00:09:03.940 wrong expert in, garbage out. Pretty much. And 00:09:03.940 --> 00:09:06.019 training them is complex, plus you need enough 00:09:06.019 --> 00:09:08.399 memory VRAM to hold all those experts, even if 00:09:08.399 --> 00:09:10.720 most are idle. OK, so Moe's give us efficient 00:09:10.720 --> 00:09:13.179 expertise. But all these models so far, they've 00:09:13.179 --> 00:09:15.200 been dealing with text or maybe creating images. 00:09:15.309 --> 00:09:18.110 based on text, what if the AI needs to actually 00:09:18.110 --> 00:09:21.450 see the world? Ah, giving the AI eyes. That must 00:09:21.450 --> 00:09:23.870 be the VLM vision language model. You got it. 00:09:24.110 --> 00:09:26.990 V -L -M. It can see and talk. It understands 00:09:26.990 --> 00:09:30.850 both pictures and text at the same time. LLMs, 00:09:31.250 --> 00:09:34.009 fundamentally, were blind. VLMs have eyes. How 00:09:34.009 --> 00:09:36.750 does that work internally? Is it merging two 00:09:36.750 --> 00:09:39.190 different kinds of models? Kind of, yeah. It 00:09:39.190 --> 00:09:42.779 uses an image encoder. Think of that as the eyes, 00:09:42.860 --> 00:09:46.019 which looks at a picture and translates it into 00:09:46.019 --> 00:09:48.480 a sort of numerical description, like orange 00:09:48.480 --> 00:09:51.179 cat sitting on a green chair. OK. A number string 00:09:51.179 --> 00:09:53.779 that represents the image. Right. And then that 00:09:53.779 --> 00:09:56.039 numerical string gets fed into a language model, 00:09:56.179 --> 00:09:59.210 the mouth. which can then understand it and talk 00:09:59.210 --> 00:10:01.330 about it, answer questions about it. And that 00:10:01.330 --> 00:10:03.629 opens up some really intuitive uses. The classic 00:10:03.629 --> 00:10:05.830 example is taking a picture of your fridge and 00:10:05.830 --> 00:10:07.909 asking, OK, what can I make for dinner with this 00:10:07.909 --> 00:10:10.169 stuff? That's the famous one. But it goes way 00:10:10.169 --> 00:10:13.230 beyond that. Think about visual search. Take 00:10:13.230 --> 00:10:15.110 a picture of some cool shoes you see someone 00:10:15.110 --> 00:10:17.450 wearing and ask, where can I buy these exact 00:10:17.450 --> 00:10:21.190 ones? Ooh, dangerous for my wallet. Huh. Tell 00:10:21.190 --> 00:10:24.870 me about it. Or analyzing video content. and 00:10:24.870 --> 00:10:27.809 really powerfully, tools for the visually impaired. 00:10:28.570 --> 00:10:30.509 Imagine it describing the world around them, 00:10:30.750 --> 00:10:33.549 reading signs, maybe even recognizing faces or 00:10:33.549 --> 00:10:35.870 expressions. That's incredible, a truly helpful 00:10:35.870 --> 00:10:38.429 application. But I guess the same watch out applies 00:10:38.429 --> 00:10:41.990 here as with LLMs. Can VLMs hallucinate about 00:10:41.990 --> 00:10:44.210 images? Absolutely. They can misinterpret an 00:10:44.210 --> 00:10:46.330 image, describe an object that isn't there, or 00:10:46.330 --> 00:10:48.710 get an action wrong. And of course, privacy becomes 00:10:48.710 --> 00:10:50.509 a bigger concern when you're uploading images, 00:10:50.610 --> 00:10:52.830 especially of people or inside your home. Right. 00:10:53.090 --> 00:10:55.529 So besides the fridge example, what's a really 00:10:55.529 --> 00:10:57.970 powerful, maybe less obvious application of a 00:10:57.970 --> 00:11:00.470 VLM? I think describing the world in real time 00:11:00.470 --> 00:11:02.509 for visually impaired people is one of the most 00:11:02.509 --> 00:11:05.669 profound uses. Yeah, definitely. OK, VLMs are 00:11:05.669 --> 00:11:08.389 powerful. They see and talk. But they sound like 00:11:08.389 --> 00:11:10.250 they need a lot of computing power, just like 00:11:10.250 --> 00:11:13.690 the big LLMs. What if you need AI on a smaller 00:11:13.690 --> 00:11:16.789 device, like your phone, and privacy is paramount? 00:11:16.950 --> 00:11:19.389 Great question. That's where the SLM comes in 00:11:19.389 --> 00:11:21.389 the small language model. While everyone's chasing 00:11:21.389 --> 00:11:25.470 bigger and bigger, these SLMs, like Microsoft's 00:11:25.470 --> 00:11:28.629 Fi3, are quietly becoming super important for 00:11:28.629 --> 00:11:31.029 on -device AI. They're designed specifically 00:11:31.029 --> 00:11:33.750 to run efficiently on phones, laptops, maybe 00:11:33.750 --> 00:11:36.720 even smart appliances. Small. So fewer parameters, 00:11:36.779 --> 00:11:38.679 I guess. How small are we talking? Just maybe 00:11:38.679 --> 00:11:41.059 fewer, maybe just one to three billion parameters 00:11:41.059 --> 00:11:43.299 compared to hundreds or even over a trillion 00:11:43.299 --> 00:11:45.440 for the big LLMs. Okay, one to three billion 00:11:45.440 --> 00:11:48.480 versus a trillion. How do they stay capable then? 00:11:48.679 --> 00:11:51.279 How aren't they just dumb? The secret sauce is 00:11:51.279 --> 00:11:53.720 the training data. Instead of just scraping the 00:11:53.720 --> 00:11:56.720 entire messy internet like LLMs often do. Right. 00:11:56.909 --> 00:12:00.210 Researchers focus on quality over quantity. They 00:12:00.210 --> 00:12:02.789 use what they call textbook quality data, highly 00:12:02.789 --> 00:12:05.429 curated information, specific examples of reasoning, 00:12:05.750 --> 00:12:08.210 logic problems. It's like feeding it a very well 00:12:08.210 --> 00:12:09.889 -structured education instead of just letting 00:12:09.889 --> 00:12:12.289 it browse the web randomly. Textbook quality. 00:12:12.429 --> 00:12:14.879 Yeah. That's interesting. So why is the training 00:12:14.879 --> 00:12:17.799 data for an SLM described that way? Because they 00:12:17.799 --> 00:12:20.440 prioritize high -quality curated knowledge and 00:12:20.440 --> 00:12:22.779 reasoning examples over just sheer volume from 00:12:22.779 --> 00:12:25.480 the internet. Quality over quantity. But doesn't 00:12:25.480 --> 00:12:28.460 that curated approach risk making them less robust? 00:12:29.220 --> 00:12:31.340 Like, they only know what's in the textbook. 00:12:31.460 --> 00:12:34.139 They can't handle the messy, unpredictable real 00:12:34.139 --> 00:12:36.980 world as well. Are they trading robustness for 00:12:36.980 --> 00:12:38.940 that efficiency? That is the trade -off, yeah. 00:12:39.100 --> 00:12:42.299 They will know less about really news or uncommon 00:12:42.299 --> 00:12:44.360 topics. Their conversational memory might be 00:12:44.360 --> 00:12:48.059 shorter too, but the payoff is huge. True on 00:12:48.059 --> 00:12:50.059 -device AI. Meaning it runs right there on your 00:12:50.059 --> 00:12:52.700 phone. Exactly. Your phone's assistant can work 00:12:52.700 --> 00:12:54.759 instantly without needing a cloud connection. 00:12:55.139 --> 00:12:57.580 That means better privacy. Your data doesn't 00:12:57.580 --> 00:13:00.039 have to leave your device. Think about instant 00:13:00.039 --> 00:13:02.360 code completion in your programming editor or 00:13:02.360 --> 00:13:04.820 voice commands in your car that just work, even 00:13:04.820 --> 00:13:07.490 if you have no signal. Smart microwaves, maybe. 00:13:07.629 --> 00:13:10.289 OK, so SLMs are great for efficient, private, 00:13:10.470 --> 00:13:12.990 on -device tasks. But they're still fundamentally 00:13:12.990 --> 00:13:15.269 about generating text or understanding commands, 00:13:15.429 --> 00:13:18.070 like LLMs. What if the main goal isn't generation 00:13:18.070 --> 00:13:20.929 at all, but really deep understanding of language 00:13:20.929 --> 00:13:23.809 meaning? For deep understanding, we need to look 00:13:23.809 --> 00:13:26.610 back at a slightly older but foundational model 00:13:26.610 --> 00:13:30.710 type, the MLM Masked Language Model. Masked Language 00:13:30.710 --> 00:13:33.029 Model. OK, how is that different from an LLM 00:13:33.029 --> 00:13:36.620 guessing the next word? So LLMs play the predict 00:13:36.620 --> 00:13:39.779 the next word game. MLMs, like Google's famous 00:13:39.779 --> 00:13:42.600 BERT model, play fill in the blank. Fill in the 00:13:42.600 --> 00:13:45.220 blank? Yeah. During training, they take a sentence 00:13:45.220 --> 00:13:48.200 and just hide or mask about 15 % of the words. 00:13:48.700 --> 00:13:51.279 Then they force the model to guess the missing 00:13:51.279 --> 00:13:54.820 words. Crucially, to do that well, the model 00:13:54.820 --> 00:13:57.279 has to look at the words before the blank and 00:13:57.279 --> 00:14:00.220 the words after the blank. Ah, so it needs context 00:14:00.220 --> 00:14:02.909 from both directions. bidirectional context. 00:14:03.350 --> 00:14:05.389 Exactly. That forces it to develop a much deeper 00:14:05.389 --> 00:14:07.669 understanding of the sentence's actual meaning 00:14:07.669 --> 00:14:10.450 and structure, not just predicting the most likely 00:14:10.450 --> 00:14:13.110 next word in sequence. OK, so where is that deep 00:14:13.110 --> 00:14:15.669 understanding most useful if it's not writing 00:14:15.669 --> 00:14:18.409 poems? Right. They're analysis models, not creative 00:14:18.409 --> 00:14:20.629 writers. Their big strength is understanding 00:14:20.629 --> 00:14:22.909 meaning, which is critical for search engines. 00:14:23.269 --> 00:14:25.450 When you search something like, best camera is 00:14:25.450 --> 00:14:27.889 A to Z, the MLM helps the engine understand A 00:14:27.889 --> 00:14:30.649 to Z means a complete guide, not literally the 00:14:30.649 --> 00:14:33.159 letter. Uh, understanding the intent behind the 00:14:33.159 --> 00:14:36.220 search. Precisely. They're also great for sentiment 00:14:36.220 --> 00:14:39.059 analysis, quickly sorting thousands of customer 00:14:39.059 --> 00:14:42.279 reviews into positive, negative, neutral, and 00:14:42.279 --> 00:14:45.059 for something called Named Entity Recognition, 00:14:45.240 --> 00:14:48.139 or N -E -R, that's pulling out specific pieces 00:14:48.139 --> 00:14:50.779 of info from text, like finding all the company 00:14:50.779 --> 00:14:53.179 names or dollar amounts mentioned in a long report. 00:14:53.580 --> 00:14:55.519 So since MLMs aren't really creative writers, 00:14:56.100 --> 00:14:57.940 what's their crucial function for something like 00:14:57.940 --> 00:15:00.639 a search engine? They understand the deep context 00:15:00.639 --> 00:15:03.039 and the actual meaning behind your search query, 00:15:03.220 --> 00:15:06.039 not just the keywords. Got it. Pure contextual 00:15:06.039 --> 00:15:09.059 depth. OK, one more model to go. We've covered 00:15:09.059 --> 00:15:12.860 text, images, actions, efficiency, seeing, small 00:15:12.860 --> 00:15:16.460 size, deep understanding, what's left. Precision 00:15:16.460 --> 00:15:19.539 in vision, specifically identifying exact boundaries. 00:15:19.740 --> 00:15:21.600 We need the SAM, the Segment Anything model. 00:15:21.779 --> 00:15:24.889 Segment Anything. SAM. Okay, this sounds specialized. 00:15:25.070 --> 00:15:27.029 Highly specialized. It's a computer vision model, 00:15:27.090 --> 00:15:29.289 but its focus isn't on telling you what an object 00:15:29.289 --> 00:15:31.909 is. Its job is to draw an incredibly precise 00:15:31.909 --> 00:15:34.269 outline around everything it sees in a picture. 00:15:34.730 --> 00:15:36.690 It segments objects right down to the individual 00:15:36.690 --> 00:15:38.769 pixel. Down to the pixel. Wow, how does it learn 00:15:38.769 --> 00:15:41.860 to do that for... Well, anything. It was trained 00:15:41.860 --> 00:15:44.779 on a massive data set, something like 11 million 00:15:44.779 --> 00:15:48.059 images, containing over a billion segmentation 00:15:48.059 --> 00:15:51.059 masks, those precise outlines. So it learned 00:15:51.059 --> 00:15:53.980 the general concept of what is an object and 00:15:53.980 --> 00:15:56.000 what is its boundary, even if it doesn't know 00:15:56.000 --> 00:15:58.500 the object's name. You usually activate it just 00:15:58.500 --> 00:16:01.080 by clicking on an object or drawing a rough box 00:16:01.080 --> 00:16:03.669 around it. So it knows This boundary belongs 00:16:03.669 --> 00:16:06.429 to object one, but not necessarily that object 00:16:06.429 --> 00:16:09.409 one is a dog or a car. Exactly. It just finds 00:16:09.409 --> 00:16:12.110 the edges. Okay. Where do you use that kind of 00:16:12.110 --> 00:16:15.220 pixel perfect? outlining. It's actually a foundational 00:16:15.220 --> 00:16:18.340 tool for lots of applications. Think photo editing, 00:16:18.600 --> 00:16:21.480 that remove background feature. Sam is likely 00:16:21.480 --> 00:16:24.399 powering the precise selection. Or critically, 00:16:24.759 --> 00:16:27.460 medical imaging. Imagine a surgeon analyzing 00:16:27.460 --> 00:16:30.679 an MRI. Sam can draw the exact outline around 00:16:30.679 --> 00:16:33.100 a tumor or an organ, which allows for really 00:16:33.100 --> 00:16:35.080 accurate measurements. Right. Precision is key 00:16:35.080 --> 00:16:37.299 there. Absolutely. And it's vital for robotics 00:16:37.299 --> 00:16:39.759 and self -driving cars too. They need to know 00:16:39.759 --> 00:16:42.200 the exact edges of objects in the real world 00:16:42.200 --> 00:16:44.730 to nap and interact safely. So why would a surgeon 00:16:44.730 --> 00:16:48.149 looking at an MRI need SAM, the segmenter, maybe 00:16:48.149 --> 00:16:51.490 more than a VLM, the describer? SAM provides 00:16:51.490 --> 00:16:54.629 those precise pixel level outlines needed to 00:16:54.629 --> 00:16:57.269 accurately measure biological structures, not 00:16:57.269 --> 00:16:59.370 just a general description. Measurement needs 00:16:59.370 --> 00:17:02.389 precision. Makes perfect sense. Is there a watch 00:17:02.389 --> 00:17:04.640 out for SAM? The main thing is just remembering 00:17:04.640 --> 00:17:07.279 what it is. It's brilliant at finding outlines, 00:17:07.700 --> 00:17:10.119 but it won't tell you what the object is. It 00:17:10.119 --> 00:17:12.119 often needs to be paired with other models, maybe 00:17:12.119 --> 00:17:15.019 a VLM, to get both the precise outline and the 00:17:15.019 --> 00:17:17.859 identification. Okay, wow. That's eight different 00:17:17.859 --> 00:17:20.420 types. Quite the toolbox. It really is. So let's 00:17:20.420 --> 00:17:22.039 try and pull this all together. What's the big 00:17:22.039 --> 00:17:24.539 idea here for you, the listener? I think the 00:17:24.539 --> 00:17:27.400 core takeaway is just realizing that AI isn't 00:17:27.400 --> 00:17:30.099 one thing, it's this specialized toolbox. And 00:17:30.099 --> 00:17:32.259 now hopefully knowing the model name and the 00:17:32.259 --> 00:17:34.700 acronym actually tells you something concrete 00:17:34.700 --> 00:17:36.960 about its core function. You know which tool 00:17:36.960 --> 00:17:39.039 to think about for which job. Yeah, let's do 00:17:39.039 --> 00:17:41.500 a quick recap. If you need writing help, content 00:17:41.500 --> 00:17:44.960 generation, chatbots, you're thinking LLM. Right, 00:17:45.119 --> 00:17:47.680 needs super fast images, especially for real 00:17:47.680 --> 00:17:49.859 -time stuff like AR or on your phone, that's 00:17:49.859 --> 00:17:52.880 the LCM. Want to automate a complex task? Have 00:17:52.880 --> 00:17:56.059 the AI actually do things by connecting to other 00:17:56.059 --> 00:17:58.680 tools. That's the LAM, the action model. Need 00:17:58.680 --> 00:18:01.299 maximum efficiency and the ability to build huge 00:18:01.299 --> 00:18:03.759 models cost -effectively using specialist parts. 00:18:03.799 --> 00:18:06.079 That's the MoE architecture. Need an AI that 00:18:06.079 --> 00:18:08.359 can see and understand images and talk about 00:18:08.359 --> 00:18:11.160 them. That's the VLM. Want AI that runs directly 00:18:11.160 --> 00:18:14.000 on your device, prioritizing privacy and efficiency. 00:18:13.799 --> 00:18:17.319 over knowing absolutely everything. That's the 00:18:17.319 --> 00:18:20.140 SLM, the small one. Need deep understanding of 00:18:20.140 --> 00:18:22.799 language meaning for search or analysis, not 00:18:22.799 --> 00:18:25.259 generation. That's the MLM. And finally, need 00:18:25.259 --> 00:18:27.779 to draw those perfect pixel -level outlines around 00:18:27.779 --> 00:18:30.400 objects in images, like for background removal 00:18:30.400 --> 00:18:33.819 or medical scans. That's the SAM. Exactly. So 00:18:33.819 --> 00:18:35.619 now you have this framework, right? We really 00:18:35.619 --> 00:18:37.119 hope you can use this knowledge next time you 00:18:37.119 --> 00:18:39.359 hear about a new AI tool. Don't just ask, is 00:18:39.359 --> 00:18:42.380 it AI? Ask yourself, okay. Based on what it does, 00:18:42.539 --> 00:18:44.619 is this primarily analyzing things, creating 00:18:44.619 --> 00:18:47.079 things, or acting on things? Which tool from 00:18:47.079 --> 00:18:49.039 the toolbox does it sound like? Yeah, that's 00:18:49.039 --> 00:18:50.519 a great way to approach it. And the pace things 00:18:50.519 --> 00:18:52.460 are moving, these tools are starting to combine 00:18:52.460 --> 00:18:54.640 in really interesting ways. We've covered models 00:18:54.640 --> 00:18:56.880 that talk, models that act, models that see. 00:18:57.539 --> 00:18:59.880 So here's a thought to leave you with. If the 00:18:59.880 --> 00:19:03.180 next big step... combines that incredible efficiency 00:19:03.180 --> 00:19:06.680 of the Moe's specialized experts with the Lamb's 00:19:06.680 --> 00:19:09.220 ability to connect to tools and automate real 00:19:09.220 --> 00:19:12.160 -world tasks, well, that creates the potential 00:19:12.160 --> 00:19:16.640 for truly autonomous, hyper -specialized AI agents. 00:19:17.599 --> 00:19:19.920 So thinking about that future, what's the single 00:19:19.920 --> 00:19:23.099 most complex task, maybe in your personal life, 00:19:23.099 --> 00:19:25.220 maybe in your business, that you would finally 00:19:25.220 --> 00:19:27.980 trust an autonomous AI agent to handle completely? 00:19:28.319 --> 00:19:29.980 Something to chew on. Definitely something to 00:19:29.980 --> 00:19:31.529 think about. We'll see you next time on the Deep 00:19:31.529 --> 00:19:31.730 Dive.