WEBVTT 00:00:00.000 --> 00:00:02.120 For years, if you wanted to get into serious 00:00:02.120 --> 00:00:05.839 AI modeling or deep learning, you hit this wall, 00:00:06.299 --> 00:00:08.480 a really expensive one. Yeah, the hardware wall. 00:00:08.660 --> 00:00:11.500 You needed these specialized, powerful GPUs or 00:00:11.500 --> 00:00:13.419 just a massive cloud budget. It felt kind of 00:00:13.419 --> 00:00:15.359 out of reach for most people. Right, or like 00:00:15.359 --> 00:00:17.640 you needed a grant from a university or something 00:00:17.640 --> 00:00:21.420 just to get started. Exactly. It was a real barrier. 00:00:21.699 --> 00:00:23.879 Your cool idea could just stall because you didn't 00:00:23.879 --> 00:00:26.260 have the cash for the hardware or even the time 00:00:26.260 --> 00:00:28.559 to figure out the setup. Pretty demoralizing. 00:00:28.579 --> 00:00:31.140 Well, that wall. The one that basically defined 00:00:31.140 --> 00:00:35.039 AI development economics for like a decade. Yeah. 00:00:35.100 --> 00:00:37.329 It's kind of gone now. Remarkably. And welcome 00:00:37.329 --> 00:00:39.549 to The Deep Dive. We've been digging into sources 00:00:39.549 --> 00:00:42.950 that give a really solid guide to Google CoLab. 00:00:43.210 --> 00:00:45.869 It's this free tool, runs right in your browser, 00:00:46.009 --> 00:00:49.869 and it honestly puts a high -powered AI lab right 00:00:49.869 --> 00:00:52.189 at your fingertips. This isn't just, you know, 00:00:52.210 --> 00:00:55.369 some handy utility. It's a direct shortcut to 00:00:55.369 --> 00:00:57.829 building AI stuff no matter what your budget 00:00:57.829 --> 00:00:59.789 looks like. So our mission today is pretty clear. 00:00:59.929 --> 00:01:01.969 We want to understand how CoLab is structured, 00:01:02.210 --> 00:01:05.049 how its sort of unique cell system works. Yeah, 00:01:05.049 --> 00:01:07.090 and then we'll hit the key. key features that 00:01:07.090 --> 00:01:10.129 make it so powerful. And maybe most importantly 00:01:10.129 --> 00:01:13.030 for you listening, we're going to cover the common 00:01:13.030 --> 00:01:15.909 mistakes, the things that catch pretty much every 00:01:15.909 --> 00:01:19.670 new user so you can avoid losing time or worse, 00:01:19.810 --> 00:01:22.650 your work. Right. But get these few things down 00:01:22.650 --> 00:01:25.129 and you can basically jump straight into building 00:01:25.129 --> 00:01:27.670 without those early headaches. Okay. So let's 00:01:27.670 --> 00:01:30.510 start right here. What is Colab exactly? So Google 00:01:30.510 --> 00:01:33.159 Colab, it's... short for Collaboratory People, 00:01:33.379 --> 00:01:35.980 often call it the Google Docs for Code. That's 00:01:35.980 --> 00:01:37.700 a great analogy. Yeah, it's collaborative. You 00:01:37.700 --> 00:01:40.640 can share it easily. And crucially, zero installation 00:01:40.640 --> 00:01:43.700 needed. No setup on your own computer. You just 00:01:43.700 --> 00:01:45.939 log into Google and boom, you've got a coding 00:01:45.939 --> 00:01:48.099 environment. And it is like Google Docs in that 00:01:48.099 --> 00:01:50.780 way. Under the hood, it's basically a cloud -hosted 00:01:50.780 --> 00:01:53.280 Jupyter notebook. But the really revolutionary 00:01:53.280 --> 00:01:56.840 part, the game changer, is the free access it 00:01:56.840 --> 00:01:59.799 gives you to some seriously powerful hardware. 00:01:59.959 --> 00:02:02.959 We're talking GPUs and... tpus okay so gpus and 00:02:02.959 --> 00:02:04.959 tpus if you've got a new to this these are special 00:02:04.959 --> 00:02:06.900 processors they're designed for the kind of map 00:02:06.900 --> 00:02:08.900 deep learning needs tons of calculations all 00:02:08.900 --> 00:02:12.280 at once they speed things up a lot yeah think 00:02:12.280 --> 00:02:15.080 of it like this your regular computer chip the 00:02:15.080 --> 00:02:17.699 cpu it's like a really skilled chef carefully 00:02:17.699 --> 00:02:22.110 making one complex dish okay a gpu That's like 00:02:22.110 --> 00:02:24.789 an army of line cooks all making thousands of 00:02:24.789 --> 00:02:27.650 simple things simultaneously. The speed difference 00:02:27.650 --> 00:02:30.389 is honestly pretty wild. Our sources mention 00:02:30.389 --> 00:02:33.349 a model that might take, say, eight hours on 00:02:33.349 --> 00:02:36.030 a decent laptop could finish in maybe 15 minutes 00:02:36.030 --> 00:02:39.389 on colab using these free accelerators wow that 00:02:39.389 --> 00:02:41.409 kind of accessibility it really does level the 00:02:41.409 --> 00:02:44.310 playing field doesn't it totally students founders 00:02:44.310 --> 00:02:46.990 trying to bootstrap something researchers anywhere 00:02:46.990 --> 00:02:50.030 they can all jump in learning suddenly just costs 00:02:50.030 --> 00:02:53.110 your time not a pile of cash so what would you 00:02:53.110 --> 00:02:55.050 say is the single biggest impact of that free 00:02:55.050 --> 00:02:57.030 access for someone just starting out it just 00:02:57.030 --> 00:02:59.830 removes the money barrier completely Anyone can 00:02:59.830 --> 00:03:01.729 start building AI right away. All right. So to 00:03:01.729 --> 00:03:03.789 actually use that power, you need to get the 00:03:03.789 --> 00:03:06.129 hang of the basic structure first, which is the 00:03:06.129 --> 00:03:08.009 Jupyter notebook format. It's built out of these 00:03:08.009 --> 00:03:10.830 little independent blocks or cells. Right. And 00:03:10.830 --> 00:03:14.270 each cell hold either code, usually Python or 00:03:14.270 --> 00:03:16.909 text like notes or explanations using something 00:03:16.909 --> 00:03:20.469 called Markdown. And this cell based approach 00:03:20.469 --> 00:03:23.530 is really key for efficiency. Especially with 00:03:23.530 --> 00:03:26.330 data work. Like imagine you have a big 10 gigabyte 00:03:26.330 --> 00:03:28.430 data set you need to analyze. Okay. Cell one, 00:03:28.590 --> 00:03:30.250 maybe you load the data, takes five minutes. 00:03:30.629 --> 00:03:32.770 Cell two, you clean it up, do some processing. 00:03:32.930 --> 00:03:35.289 Cell three, you make a chart, a visualization. 00:03:35.689 --> 00:03:37.590 Got it. Now, let's say you just want to change 00:03:37.590 --> 00:03:40.509 the title on that chart. You only need to rerun 00:03:40.509 --> 00:03:43.430 cell three. Ah, right. The data loaded in cell 00:03:43.430 --> 00:03:46.009 one and processed in cell two, it's still there 00:03:46.009 --> 00:03:48.129 loaded in memory for that session. You don't 00:03:48.129 --> 00:03:50.030 have to wait another five minutes just to change 00:03:50.030 --> 00:03:52.530 a label. which saves a ton of time when you're 00:03:52.530 --> 00:03:55.169 tweaking things. Okay. But that flexibility also 00:03:55.169 --> 00:03:57.509 leads to maybe the biggest point of confusion 00:03:57.509 --> 00:04:00.090 for beginners, right? Oh, yeah. Execution order 00:04:00.090 --> 00:04:03.090 versus cell position. This is crucial. The cells 00:04:03.090 --> 00:04:05.449 look like they're arranged top to bottom, like 00:04:05.449 --> 00:04:08.409 a document. Yeah. But they actually run in whatever 00:04:08.409 --> 00:04:10.469 order you tell them to run. You see those little 00:04:10.469 --> 00:04:12.250 numbers and brackets next to the cells, like 00:04:12.250 --> 00:04:14.650 one, two. That's the actual execution order for 00:04:14.650 --> 00:04:16.689 your current session. And this is where it gets 00:04:16.689 --> 00:04:20.839 messy. You could, like... scroll way down to 00:04:20.839 --> 00:04:23.079 cell five and define some important variable 00:04:23.079 --> 00:04:25.420 there then you scroll back up to cell three and 00:04:25.420 --> 00:04:28.160 try to use that variable but if you haven't actually 00:04:28.160 --> 00:04:31.259 run cell five yet in this set exactly cell three 00:04:31.259 --> 00:04:33.660 crashes because the variable doesn't exist yet 00:04:33.660 --> 00:04:35.800 as far as the execution kernel is concerned even 00:04:35.800 --> 00:04:37.540 though you can see it defined lower down on the 00:04:37.540 --> 00:04:40.459 page it's like trying to use step five ingredients 00:04:40.459 --> 00:04:43.180 back in step three of a recipe perfect analogy 00:04:43.180 --> 00:04:45.860 it hasn't happened yet in the process so Why 00:04:45.860 --> 00:04:48.040 is getting your head around this execution order 00:04:48.040 --> 00:04:52.319 thing so critical for debugging? Because getting 00:04:52.319 --> 00:04:54.300 cells out of order is probably the number one 00:04:54.300 --> 00:04:56.720 reason for weird results and frustration when 00:04:56.720 --> 00:04:58.959 you're starting out. Getting going is super simple, 00:04:59.000 --> 00:05:01.019 though. All you need is a Google account. Go 00:05:01.019 --> 00:05:03.939 to colab .research .google .com. You can run 00:05:03.939 --> 00:05:06.540 a quick test like print. Hello, Colab. Right. 00:05:06.860 --> 00:05:09.420 And that confirms everything's ready. Python's 00:05:09.420 --> 00:05:12.680 there. NumPy, Pandas, TensorFlow. The whole data 00:05:12.680 --> 00:05:15.180 science toolkit is pre -installed, ready to go. 00:05:15.550 --> 00:05:17.750 Okay, let's talk features. The things that make 00:05:17.750 --> 00:05:19.529 this more than just a simple notebook, number 00:05:19.529 --> 00:05:22.230 one have to be getting that free supercomputer 00:05:22.230 --> 00:05:24.569 access. Definitely. You just go to the runtime 00:05:24.569 --> 00:05:27.949 menu, click change runtime type, and pick GPU. 00:05:28.490 --> 00:05:30.910 Usually the free one they offer is something 00:05:30.910 --> 00:05:34.250 like a Tesla T4 GPU, which is pretty powerful. 00:05:34.610 --> 00:05:36.350 Yeah, tell us about the T4. What's the deal with 00:05:36.350 --> 00:05:39.480 that specific chip? So the T4 is really good, 00:05:39.579 --> 00:05:41.740 especially for what's called inference. That's 00:05:41.740 --> 00:05:44.579 running a model after it's already trained. It's 00:05:44.579 --> 00:05:47.660 also decent for medium sized training tasks. 00:05:48.060 --> 00:05:50.740 It's not designed for like training a massive 00:05:50.740 --> 00:05:54.100 model for days on end, but for learning, prototyping, 00:05:54.139 --> 00:05:56.920 getting things working. It's fantastic. It's 00:05:56.920 --> 00:05:59.220 a game changer for free access. And they've also 00:05:59.220 --> 00:06:01.620 got that AI assistant built right in now, Gemini. 00:06:01.740 --> 00:06:03.639 Yeah. And it's not just some generic chatbot. 00:06:03.720 --> 00:06:06.319 It's context aware. It actually knows about the 00:06:06.319 --> 00:06:09.079 code. in your notebook. Ah, that's clever. It 00:06:09.079 --> 00:06:11.500 knows your variables, the libraries you've imported, 00:06:11.680 --> 00:06:13.620 even the error message you just got. You can 00:06:13.620 --> 00:06:16.699 highlight some code that makes no sense and ask 00:06:16.699 --> 00:06:19.100 it, like, explain this like I'm five. And because 00:06:19.100 --> 00:06:20.920 it has the context, the explanation is usually 00:06:20.920 --> 00:06:23.819 spot on. It really helps you learn faster. That 00:06:23.819 --> 00:06:25.680 sounds incredibly useful. Another big one is 00:06:25.680 --> 00:06:27.899 GitHub integration, right? Yeah. That feels important 00:06:27.899 --> 00:06:30.399 for more serious work. Oh, huge. It connects 00:06:30.399 --> 00:06:32.379 Colab directly to how professionals actually 00:06:32.379 --> 00:06:34.519 manage code. You can open notebooks straight 00:06:34.519 --> 00:06:36.660 from a GitHub repository, make changes, save 00:06:36.660 --> 00:06:39.220 them back, commit them all without leaving Colab. 00:06:39.259 --> 00:06:41.879 Nice. And visualization, too. Libraries like 00:06:41.879 --> 00:06:44.620 Matplotlib or Plotly, they just work out of the 00:06:44.620 --> 00:06:46.839 box. You run a code cell to make a chart, and 00:06:46.839 --> 00:06:48.920 boom, the interactive chart appears right below 00:06:48.920 --> 00:06:51.519 it. Instant feedback. Yeah, seeing the result 00:06:51.519 --> 00:06:54.199 immediately like that must really speed up exploring 00:06:54.199 --> 00:06:57.240 data. So beyond just the speed from the GPU, 00:06:57.459 --> 00:07:00.319 what feature really makes Colab great for sharing 00:07:00.319 --> 00:07:03.220 your work or research? The mix of markdown text 00:07:03.220 --> 00:07:05.939 and runnable code. It lets you create these documents 00:07:05.939 --> 00:07:08.879 that both explain and demonstrate fully reproducible. 00:07:09.079 --> 00:07:11.040 Okay, let's shift gears to the warnings. The 00:07:11.040 --> 00:07:13.180 common traps. We talked about the time traveler 00:07:13.180 --> 00:07:16.439 problem, that execution order confusion. Right. 00:07:16.839 --> 00:07:19.100 And the quick fix, maybe the slightly brute force 00:07:19.100 --> 00:07:22.839 fix, if things get weird, is runtime, run all. 00:07:23.680 --> 00:07:26.040 Clears everything and runs top to bottom. But 00:07:26.040 --> 00:07:28.459 there's a bigger potential tragedy, isn't there? 00:07:28.560 --> 00:07:32.060 The great disconnect. Ah, yes. This one hurts. 00:07:32.300 --> 00:07:34.439 Colab runtimes, the virtual machines you're using, 00:07:34.500 --> 00:07:37.500 they're temporary. How temporary? Well... If 00:07:37.500 --> 00:07:39.959 you're idle for about 90 minutes, it might disconnect. 00:07:40.259 --> 00:07:42.740 And there's usually a hard limit, maybe 12 hours 00:07:42.740 --> 00:07:46.180 total session time on the free tier when it disconnects. 00:07:46.180 --> 00:07:48.300 Everything's gone. Everything that was only in 00:07:48.300 --> 00:07:50.879 the machine's memory, your variables, your loaded 00:07:50.879 --> 00:07:53.139 data, that model you were halfway through training, 00:07:53.300 --> 00:07:57.279 poof, gone instantly. Oh, man. Six hours of training 00:07:57.279 --> 00:07:59.500 just vanished. Yeah. Okay, so what's the absolute 00:07:59.500 --> 00:08:02.060 must -do defense against that? You have to save 00:08:02.060 --> 00:08:04.680 anything important somewhere permanent. And the 00:08:04.680 --> 00:08:06.720 standard way is mounting Google Drive. How does 00:08:06.720 --> 00:08:08.339 that work? It's just a little code snippet you 00:08:08.339 --> 00:08:11.199 run. It securely links your temporary Colab session 00:08:11.199 --> 00:08:14.220 to your personal Google Drive storage. Think 00:08:14.220 --> 00:08:17.220 of Colab like a shared desk that gets wiped clean 00:08:17.220 --> 00:08:20.160 every night. Google Drive is your personal locked 00:08:20.160 --> 00:08:23.100 filing cabinet. Anything you save only to the 00:08:23.100 --> 00:08:26.360 Colab machine's temporary disk, the content directory 00:08:26.360 --> 00:08:29.019 will disappear when the session ends. Mount Drive, 00:08:29.319 --> 00:08:33.220 save there. Yeah. I got to admit, even now, sometimes 00:08:33.220 --> 00:08:35.480 I'll tweak a variable in one cell, forget to 00:08:35.480 --> 00:08:37.559 rerun it, then spend like five minutes debugging 00:08:37.559 --> 00:08:39.919 something else entirely. Happens to the best 00:08:39.919 --> 00:08:43.230 of us. Only to realize the change I made never 00:08:43.230 --> 00:08:45.570 actually executed. It's just part of the notebook 00:08:45.570 --> 00:08:48.269 life, I guess. And related to that is the temporary 00:08:48.269 --> 00:08:51.409 storage trap. That content folder seems like 00:08:51.409 --> 00:08:53.549 a place to put files. Right. You download a data 00:08:53.549 --> 00:08:55.710 set there. Exactly. Maybe using a shell command. 00:08:56.110 --> 00:08:58.870 But if you don't immediately copy or move that 00:08:58.870 --> 00:09:01.149 data set to your mounted Google Drive, it's going 00:09:01.149 --> 00:09:04.049 to vanish when the session resets. Poof. So if 00:09:04.049 --> 00:09:06.250 I've got a really important long training job 00:09:06.250 --> 00:09:09.289 running, what's my absolute minimum safety net 00:09:09.289 --> 00:09:11.620 against losing work from a disconnect? Definitely 00:09:11.620 --> 00:09:13.600 set up checkpointing. Save your model's progress 00:09:13.600 --> 00:09:16.440 frequently to Google Drive. Okay, enough about 00:09:16.440 --> 00:09:18.000 avoiding disaster. Let's talk about creating 00:09:18.000 --> 00:09:19.960 cool stuff. What can you actually build with 00:09:19.960 --> 00:09:22.240 this for free? Well, it's great for training 00:09:22.240 --> 00:09:25.720 reasonably sized models, like taking a pre -trained 00:09:25.720 --> 00:09:28.159 image classifier and fine -tuning it for a specific 00:09:28.159 --> 00:09:30.460 task you have. Makes sense. And it's perfect 00:09:30.460 --> 00:09:32.940 for prototyping. Got an idea for, I don't know, 00:09:33.019 --> 00:09:36.379 a slogan generator app using a language model? 00:09:36.519 --> 00:09:38.399 Yeah. You could probably fine -tune a model in 00:09:38.399 --> 00:09:41.639 Colab, wrap it in a simple interface, and have 00:09:41.639 --> 00:09:44.519 a working demo in an afternoon, all free. That's 00:09:44.519 --> 00:09:46.600 pretty cool. What about other areas, like audio? 00:09:46.899 --> 00:09:49.120 Oh, yeah. You could use something like the Whisper 00:09:49.120 --> 00:09:52.080 Library right inside Colab. Take an hour -long 00:09:52.080 --> 00:09:55.259 audio file, transcribe it, get a summary, maybe 00:09:55.259 --> 00:09:57.700 analyze the sentiment, all in one notebook file 00:09:57.700 --> 00:10:00.779 you can share. Okay, so where's the line between 00:10:00.779 --> 00:10:03.100 the free tier and when you need to pay? So the 00:10:03.100 --> 00:10:05.019 free tier has limits. You don't get guaranteed 00:10:05.019 --> 00:10:07.139 GPU access. Sometimes they might not be available. 00:10:07.299 --> 00:10:08.980 Usually you can only run maybe two notebooks 00:10:08.980 --> 00:10:11.799 at the same time. And RAM is typically around 00:10:11.799 --> 00:10:15.620 12, 13 gigabytes. Which is decent for learning, 00:10:15.659 --> 00:10:18.419 but maybe not for huge projects. Right. That's 00:10:18.419 --> 00:10:21.580 where Colab Pro comes in. It's about $10 a month 00:10:21.580 --> 00:10:24.340 last I checked. And what does that get you? Priority 00:10:24.340 --> 00:10:27.059 access to faster GPUs, things like the A100 or 00:10:27.059 --> 00:10:30.379 V100, which are much more powerful. Also, longer 00:10:30.379 --> 00:10:33.960 run times, like up to 24 hours, and significantly 00:10:33.960 --> 00:10:36.980 more RAM, sometimes up to 50 gigs or more. So 00:10:36.980 --> 00:10:39.360 the A100 versus the T4 we mentioned earlier, 00:10:39.580 --> 00:10:42.220 what's the real difference? Think architecture. 00:10:42.340 --> 00:10:46.690 The A100 is just... a beast built for massive 00:10:46.690 --> 00:10:49.649 parallel processing really designed for speeding 00:10:49.649 --> 00:10:52.789 up those huge multi -day training jobs you pay 00:10:52.789 --> 00:10:54.929 for guaranteed access to that kind of premium 00:10:54.929 --> 00:10:57.820 horsepower So when should a startup or maybe 00:10:57.820 --> 00:11:00.320 a solo founder absolutely make the jump from 00:11:00.320 --> 00:11:02.179 free to pro? I'd say when your training really 00:11:02.179 --> 00:11:04.139 needs to run for more than 12 hours consistently 00:11:04.139 --> 00:11:07.100 or if getting access to those top tier GPUs becomes 00:11:07.100 --> 00:11:09.879 critical for hitting deadlines. OK, let's get 00:11:09.879 --> 00:11:11.740 into the collab ninja stuff. Advanced tricks 00:11:11.740 --> 00:11:14.399 for people who use it a lot. OK, ninja tips. 00:11:14.580 --> 00:11:16.559 Ninjas love magic commands. These are special 00:11:16.559 --> 00:11:18.279 instructions in a cell starting with percent 00:11:18.279 --> 00:11:20.279 or percent, like just putting percent time at 00:11:20.279 --> 00:11:22.059 the start of a cell. Yeah. It'll tell you exactly 00:11:22.059 --> 00:11:24.779 how long that one cell took to run. Super useful 00:11:24.779 --> 00:11:27.009 for finding. where your code is slow. Ah, finding 00:11:27.009 --> 00:11:29.710 bottlenecks. Nice. What else? Shell commands. 00:11:30.169 --> 00:11:33.149 Using an exclamation mark at the start of a line 00:11:33.149 --> 00:11:35.710 lets you run Linux shell commands directly on 00:11:35.710 --> 00:11:39.100 the cloud machine. Like what? like wgeturl. This 00:11:39.100 --> 00:11:41.159 downloads a file directly to the Colab machine 00:11:41.159 --> 00:11:43.860 super fast, bypassing your own internet connection. 00:11:44.320 --> 00:11:47.379 If you're grabbing a 50 gigabyte data set, huge 00:11:47.379 --> 00:11:49.559 time saver. Wow, okay, that's a good one. And 00:11:49.559 --> 00:11:51.620 security, this is crucial. Use the secrets feature. 00:11:51.759 --> 00:11:54.559 There's a little key icon in the sidebar. Use 00:11:54.559 --> 00:11:57.419 that to store sensitive stuff like API keys, 00:11:57.779 --> 00:12:01.100 your OpenIAPI key or whatever. Never, ever just 00:12:01.100 --> 00:12:03.440 paste keys directly into your code cells, especially 00:12:03.440 --> 00:12:05.220 if you might share that notebook on GitHub later. 00:12:05.419 --> 00:12:07.720 Big no -no. Right. Keep those secrets secret. 00:12:07.840 --> 00:12:11.059 It seems like all this integration, GitHub, Drive, 00:12:11.399 --> 00:12:14.460 the shell, really smooths out the workflow, doesn't 00:12:14.460 --> 00:12:16.519 it? Totally. And it leads to this idea, Colab 00:12:16.519 --> 00:12:18.700 is the lab where you experiment and develop. 00:12:18.860 --> 00:12:21.120 It's not usually the factory where you run a 00:12:21.120 --> 00:12:23.320 massive production service. So the process is 00:12:23.320 --> 00:12:27.159 develop in Colab, train the model, save the important 00:12:27.159 --> 00:12:29.919 results, the trained weights. Exactly. Save those 00:12:29.919 --> 00:12:32.700 weights securely to drive. Then you take the 00:12:32.700 --> 00:12:35.460 core logic, rewrite it cleanly, maybe as a standard 00:12:35.460 --> 00:12:38.220 Python script, and deploy that script to a proper 00:12:38.220 --> 00:12:41.200 scalable server for users. Gotcha. Lab first, 00:12:41.259 --> 00:12:45.039 then factory. Whoa. But just imagine, though, 00:12:45.120 --> 00:12:49.070 scaling something up to handle, like... a billion 00:12:49.070 --> 00:12:52.009 queries a month. And the core model was developed 00:12:52.009 --> 00:12:54.850 entirely for free, maybe just six months after 00:12:54.850 --> 00:12:57.230 you first learned Python on this very platform. 00:12:57.409 --> 00:13:00.309 That's kind of wild. That's the power here. Okay, 00:13:00.370 --> 00:13:03.529 quick ninja question. What's the one keyboard 00:13:03.529 --> 00:13:05.649 shortcut that saves the most time day to day 00:13:05.649 --> 00:13:08.350 in Colab? Oh, easy. Got to be shift plus enter. 00:13:08.470 --> 00:13:10.649 Runs the current cell and immediately moves your 00:13:10.649 --> 00:13:12.370 cursor to the next one. Keeps the flow going. 00:13:12.669 --> 00:13:15.269 So wrapping up, Colab really has knocked down 00:13:15.269 --> 00:13:17.409 those big barriers, hasn't it? Cost and complexity. 00:13:17.649 --> 00:13:19.909 It really has. It's the great equalizer. Puts 00:13:19.909 --> 00:13:22.529 powerful AI tools in everyone's hands right in 00:13:22.529 --> 00:13:25.230 their browser. Anyone, anywhere. And the notebook 00:13:25.230 --> 00:13:27.649 format itself, that mix of text and code, it 00:13:27.649 --> 00:13:30.169 encourages reproducible work. You can share not 00:13:30.169 --> 00:13:32.549 just your result, but your entire process. It's 00:13:32.549 --> 00:13:35.409 the ultimate show, don't tell. So for everyone 00:13:35.409 --> 00:13:38.740 listening. You basically have a free supercomputer 00:13:38.740 --> 00:13:42.320 available right now. Here's a thought. What's 00:13:42.320 --> 00:13:44.679 the smallest, maybe weirdest, but interesting 00:13:44.679 --> 00:13:47.679 data set you can find online right now? Something 00:13:47.679 --> 00:13:49.860 small enough to just try loading and making one 00:13:49.860 --> 00:13:52.519 simple chart from in your very first Colab notebook. 00:13:52.820 --> 00:13:55.159 Yeah, take what we've talked about, open up colab 00:13:55.159 --> 00:13:57.440 .research .google .com and just start building 00:13:57.440 --> 00:13:59.240 something. Give it a try. We'll catch you on 00:13:59.240 --> 00:13:59.940 the next deep dive.