WEBVTT 00:00:00.000 --> 00:00:03.220 Imagine an AI model, right, built to be helpful, 00:00:03.419 --> 00:00:06.040 then secretly giving out dangerous instructions. 00:00:06.219 --> 00:00:08.619 Yeah. Now picture a really clever fix that makes 00:00:08.619 --> 00:00:11.279 it, like, forget those bad habits, even on your 00:00:11.279 --> 00:00:13.960 phone. Pretty wild stuff. Welcome to the Deep 00:00:13.960 --> 00:00:16.519 Dive. Today, we're going to cut through the noise 00:00:16.519 --> 00:00:18.940 and unpack some truly fascinating developments 00:00:18.940 --> 00:00:22.579 in AI. We're talking surprising breakthroughs 00:00:22.579 --> 00:00:24.960 and making open source models actually safe, 00:00:25.120 --> 00:00:28.519 innovative tools reshaping how we create and 00:00:28.519 --> 00:00:31.160 learn. And even AI landing right in your pocket 00:00:31.160 --> 00:00:33.299 with some serious power. Exactly. It's really 00:00:33.299 --> 00:00:36.060 a snapshot of how fast this whole landscape is 00:00:36.060 --> 00:00:39.200 shifting. The focus now seems to be on trust, 00:00:39.439 --> 00:00:41.500 practical use, and just making advanced tech 00:00:41.500 --> 00:00:43.960 accessible. We've got some really cool insights 00:00:43.960 --> 00:00:46.380 today. Great. Let's dive in, starting with something 00:00:46.380 --> 00:00:49.359 pretty fundamental. Yeah. AI safety. And this 00:00:49.359 --> 00:00:52.049 idea some are calling benevolent hacking. Okay, 00:00:52.090 --> 00:00:54.070 so what's really fascinating here is this paradox. 00:00:54.570 --> 00:00:57.630 Developers take these huge open source AI models. 00:00:57.890 --> 00:01:00.670 Right. And they simplify them, strip them down 00:01:00.670 --> 00:01:03.850 to make them faster, or run on devices without 00:01:03.850 --> 00:01:06.370 much power like your phone. Makes sense. Efficiency. 00:01:06.689 --> 00:01:10.629 But the catch is... This process, this compression, 00:01:10.950 --> 00:01:13.750 often accidentally takes out these internal layers 00:01:13.750 --> 00:01:15.629 that are crucial for safety. It's kind of like, 00:01:15.650 --> 00:01:17.430 whoops, just lost the guardrail. So it's not 00:01:17.430 --> 00:01:19.849 intentional, but in trying to make it lean, they 00:01:19.849 --> 00:01:23.269 cut out vital parts. We're learning these skip 00:01:23.269 --> 00:01:25.810 layers are actually what block harmful stuff. 00:01:26.010 --> 00:01:29.959 Yeah, exactly. Things like hate speech. explicit 00:01:29.959 --> 00:01:32.840 content even instructions for building weapons 00:01:32.840 --> 00:01:35.859 they're basically the model's internal ethical 00:01:35.859 --> 00:01:40.340 compass and losing that is Well, risky. It really 00:01:40.340 --> 00:01:42.900 is. And the scope is huge because these open 00:01:42.900 --> 00:01:45.060 source models are everywhere. Anyone can grab 00:01:45.060 --> 00:01:47.420 them, change them, run them offline. So if those 00:01:47.420 --> 00:01:49.799 safety layers are gone, the risk just shoots 00:01:49.799 --> 00:01:52.260 up. You've got a powerful tool without its safety 00:01:52.260 --> 00:01:54.239 checks. Okay, so here's where it gets interesting. 00:01:54.379 --> 00:01:56.540 The fix you mentioned. Right, the ingenious fix. 00:01:56.700 --> 00:01:59.079 It's a new method that basically retrains the 00:01:59.079 --> 00:02:01.219 compressed model to, you know, remember how to 00:02:01.219 --> 00:02:03.430 behave ethically. Retrains it. and here's the 00:02:03.430 --> 00:02:06.390 kicker it does this without needing the original 00:02:06.390 --> 00:02:09.729 training data oh wow yeah that's massive not 00:02:09.729 --> 00:02:12.330 needing that original data makes it super privacy 00:02:12.330 --> 00:02:15.389 friendly you're not exposing sensitive info just 00:02:15.389 --> 00:02:17.449 to patch up the safety they essentially teach 00:02:17.449 --> 00:02:19.409 it ethical recall then pretty much they give 00:02:19.409 --> 00:02:22.830 this concrete example before the fix a model 00:02:22.830 --> 00:02:25.189 could be prompted maybe with an image and some 00:02:25.189 --> 00:02:28.849 text to give bomb making instructions yikes yeah 00:02:29.439 --> 00:02:32.099 But after retraining, even the compressed version 00:02:32.099 --> 00:02:34.900 just flat out refused. It had learned its lesson, 00:02:35.020 --> 00:02:37.699 so to speak. So this approach kind of bakes safety 00:02:37.699 --> 00:02:39.960 back in, even after the model's been heavily 00:02:39.960 --> 00:02:42.680 modified. Exactly. It's safe by design, again, 00:02:42.780 --> 00:02:45.099 and it's lighter than adding external filters, 00:02:45.180 --> 00:02:47.259 plus it's more robust because the model itself 00:02:47.259 --> 00:02:50.400 gets the risk. The team calls it benevolent hacking. 00:02:50.780 --> 00:02:53.460 It's quite an elegant solution, really. You know, 00:02:53.479 --> 00:02:55.419 I still wrestle with prompt drift sometimes, 00:02:55.599 --> 00:02:57.759 just trying to get consistent, useful outputs 00:02:57.759 --> 00:03:00.039 from these. Oh, yeah, it's tricky. So the idea 00:03:00.039 --> 00:03:03.580 that models could just forget their ethical guardrails 00:03:03.580 --> 00:03:07.110 through optimization. lose that alignment that's 00:03:07.110 --> 00:03:10.129 a really profound concern this solution feels 00:03:10.129 --> 00:03:13.069 well significant it does and they define some 00:03:13.069 --> 00:03:16.050 terms for us too like open source models that's 00:03:16.050 --> 00:03:18.509 ai with public code anyone can use and tweak 00:03:18.509 --> 00:03:21.810 right and on device ai that's ai running right 00:03:21.810 --> 00:03:24.789 on your gadget your phone your laptop no internet 00:03:24.789 --> 00:03:27.870 needed okay stepping back then what's the biggest 00:03:27.870 --> 00:03:31.129 implication of this benevolent hacking idea, 00:03:31.330 --> 00:03:34.310 this approach to AI, learning to behave reliably, 00:03:34.669 --> 00:03:37.409 even after being drastically changed. It means 00:03:37.409 --> 00:03:40.909 private, powerful AI can be trustworthy everywhere, 00:03:41.229 --> 00:03:44.750 even offline. Trustworthy everywhere. That focus 00:03:44.750 --> 00:03:47.169 on internal safety. It leads us nicely into the 00:03:47.169 --> 00:03:49.569 wider landscape, right? Where innovation is just 00:03:49.569 --> 00:03:51.449 popping up all over. Oh yeah, it's moving fast. 00:03:51.669 --> 00:03:53.009 Let's do some rapid fire highlights from the 00:03:53.009 --> 00:03:55.979 industry. Okay, first up, education. Anthropic 00:03:55.979 --> 00:03:59.159 is launching a free AI fluency curriculum, K 00:03:59.159 --> 00:04:02.460 -12, higher ed. Free. That's big. Totally. And 00:04:02.460 --> 00:04:04.740 it's designed to work with any AI model. Plus, 00:04:04.800 --> 00:04:07.819 it's fully remixable. So no vendor lock in. Great 00:04:07.819 --> 00:04:10.520 for getting AI literacy into school. That is 00:04:10.520 --> 00:04:12.449 good news. OK, what else? for creatives check 00:04:12.449 --> 00:04:14.689 this out an ex -user got fed up with existing 00:04:14.689 --> 00:04:16.990 tools and built a self -correcting image agent 00:04:16.990 --> 00:04:19.810 self -correcting yeah it like creates an image 00:04:19.810 --> 00:04:22.129 tests it against the prompt refines it until 00:04:22.129 --> 00:04:24.810 it's perfect no more endless fiddling getting 00:04:24.810 --> 00:04:27.529 the exact visual you want like that and speaking 00:04:27.529 --> 00:04:30.310 of practical stuff a chat gpt user shared this 00:04:30.310 --> 00:04:33.529 super comprehensive prompt for learning new topics 00:04:33.529 --> 00:04:38.000 it went viral almost a million views wow It just 00:04:38.000 --> 00:04:40.920 shows people are really hungry for structured 00:04:40.920 --> 00:04:44.800 ways to use AI for deep learning, not just quick 00:04:44.800 --> 00:04:47.800 answers. Makes sense. We want tools, not just 00:04:47.800 --> 00:04:49.819 toys. What about content creation? There's a 00:04:49.819 --> 00:04:52.660 new tool called Mirage. You type a prompt and 00:04:52.660 --> 00:04:56.060 boom, a fully edited TikTok or Instagram reel 00:04:56.060 --> 00:04:58.259 pops out. Seriously? Yeah, and it's supposedly 00:04:58.259 --> 00:05:00.920 designed to hook viewers on the For You page, 00:05:01.139 --> 00:05:04.019 like an AI viral video generator. Okay, that's 00:05:04.019 --> 00:05:06.569 something. Anything more serious? well yeah on 00:05:06.569 --> 00:05:08.350 the security front there was an alert attackers 00:05:08.350 --> 00:05:12.350 found a way to bypass x's ad protections oh embedding 00:05:12.350 --> 00:05:15.990 malware links and then tricking grok x's ai into 00:05:15.990 --> 00:05:18.610 amplifying them they're calling it grokking oh 00:05:18.610 --> 00:05:21.129 man yeah big warning definitely do not click 00:05:21.129 --> 00:05:23.329 those links shows how fast the bad actors adapt 00:05:23.329 --> 00:05:27.379 you know definitely okay shifting gears Hiring. 00:05:27.459 --> 00:05:30.079 Big moves there, too. OpenAI is apparently building 00:05:30.079 --> 00:05:33.480 a full stack AI hiring platform. Better matching 00:05:33.480 --> 00:05:35.939 for companies and workers. Interesting. Competing 00:05:35.939 --> 00:05:38.279 with LinkedIn. Seems like it. And LinkedIn's 00:05:38.279 --> 00:05:39.819 not standing still. They're rolling out their 00:05:39.819 --> 00:05:44.040 own AI hiring assistant. Yeah. So, yeah, AI looks 00:05:44.040 --> 00:05:46.100 set to really change recruitment. All right. 00:05:46.139 --> 00:05:49.019 One more. Big money news. Brett Taylor's AI startup, 00:05:49.160 --> 00:05:54.199 Sierra, just raised a massive $350 million at 00:05:54.199 --> 00:05:58.160 a $10 million valuation. And get this, in just 00:05:58.160 --> 00:06:00.519 18 months, they've signed up hundreds of clients. 00:06:01.279 --> 00:06:05.639 SoFi, Brex, total raised is now $635 million. 00:06:06.139 --> 00:06:09.040 That is incredible momentum. Serious enterprise 00:06:09.040 --> 00:06:11.100 folks. For sure. Okay, so looking across all 00:06:11.100 --> 00:06:12.680 these different things, education, creation, 00:06:12.899 --> 00:06:16.709 security, hiring, funding. Which one really signals 00:06:16.709 --> 00:06:19.029 a shift, making AI more practical for everyday 00:06:19.029 --> 00:06:21.569 people? I think it's AI tools becoming part of 00:06:21.569 --> 00:06:23.389 daily learning and creative workflows, right? 00:06:23.470 --> 00:06:25.189 Yeah, that integration feels key. Those quick 00:06:25.189 --> 00:06:27.250 hits show the breadth. But let's dive a bit deeper 00:06:27.250 --> 00:06:29.410 into some core AI tools and concepts that are 00:06:29.410 --> 00:06:31.250 really shaping where things are going. Okay. 00:06:31.790 --> 00:06:34.810 So one thing that caught my eye was this guide 00:06:34.810 --> 00:06:38.470 introducing the chat GPT agent model. And this 00:06:38.470 --> 00:06:40.610 isn't just your standard chatbot. What is it, 00:06:40.629 --> 00:06:43.230 then? It's talking about a general -purpose digital 00:06:43.230 --> 00:06:47.199 worker AI. an agent that can actually like operate 00:06:47.199 --> 00:06:50.300 a computer it moves way beyond just simple api 00:06:50.300 --> 00:06:53.519 automations operate a computer yeah so like using 00:06:53.519 --> 00:06:55.699 software interfaces the way a human does kind 00:06:55.699 --> 00:06:58.120 of yeah it's a conceptual leap not just calling 00:06:58.120 --> 00:07:00.160 a specific function but actually interacting 00:07:00.160 --> 00:07:04.560 with the system more broadly real ai agency okay 00:07:04.560 --> 00:07:06.540 that's definitely something to watch and then 00:07:06.540 --> 00:07:09.370 there's google ai studio right Presented as this 00:07:09.370 --> 00:07:12.069 powerful free tool. And it's a good window into 00:07:12.069 --> 00:07:15.649 the future of multimodal AI. Multimodal, meaning 00:07:15.649 --> 00:07:17.550 it handles different types of data. Exactly. 00:07:17.670 --> 00:07:20.449 Text, images, audio, all at once. It shows the 00:07:20.449 --> 00:07:23.189 industry is moving way beyond just text. Chatbots 00:07:23.189 --> 00:07:25.329 were just the start. But with all this power, 00:07:25.529 --> 00:07:28.670 how do we get reliable results? Avoid the noise. 00:07:29.050 --> 00:07:30.990 Good question. And that leads to another key 00:07:30.990 --> 00:07:33.660 point. Using structured prompts. There's a guide 00:07:33.660 --> 00:07:35.759 emphasizing this for succeeding in the, quote, 00:07:35.860 --> 00:07:39.439 AI revolution. Structured prompts. Like giving 00:07:39.439 --> 00:07:42.079 clearer instructions. Yeah, specifically using 00:07:42.079 --> 00:07:45.319 JSON prompts. That means using a specific standardized 00:07:45.319 --> 00:07:48.740 beta format JavaScript object notation to tell 00:07:48.740 --> 00:07:51.319 the AI exactly what you want. Oh, okay. It leads 00:07:51.319 --> 00:07:54.060 to more consistent, high -quality results. Helps 00:07:54.060 --> 00:07:56.120 you cut through what they call AI junk. It's 00:07:56.120 --> 00:07:59.100 about precision. Giving the AI clear guardrails. 00:07:59.160 --> 00:08:02.629 Got it. So just to recap the jargon, API automations 00:08:02.629 --> 00:08:05.750 are those specific limited tasks where software 00:08:05.750 --> 00:08:09.209 talks to software. Right. And JSON prompts are 00:08:09.209 --> 00:08:11.529 that structured way of giving instructions for 00:08:11.529 --> 00:08:13.689 more predictable, better outputs. You got it. 00:08:13.769 --> 00:08:15.870 So looking at these foundational shifts agents, 00:08:16.110 --> 00:08:19.430 multimodal structured prompts, what's the core 00:08:19.430 --> 00:08:21.550 benefit of using something like JSON for these 00:08:21.550 --> 00:08:24.089 advanced AI interactions? Predictable, high quality 00:08:24.089 --> 00:08:27.050 AI outputs, reducing that junk significantly. 00:08:27.569 --> 00:08:29.680 Makes sense. Precision matters more as capability 00:08:29.680 --> 00:08:31.879 grows. Okay, let's pivot again to another round 00:08:31.879 --> 00:08:34.860 of quick hits, industry movements, deals, what's 00:08:34.860 --> 00:08:37.000 happening. All right, quick hits round two. First, 00:08:37.399 --> 00:08:40.799 OpenAI significantly expanded its employee secondary 00:08:40.799 --> 00:08:44.179 sale. We're talking around $10 .3 billion now. 00:08:44.340 --> 00:08:47.340 Wow, that valuation just keeps climbing. Tells 00:08:47.340 --> 00:08:49.620 you something, right. And then a huge integration 00:08:49.620 --> 00:08:53.110 deal, Google and Apple. Oh, yeah. What about? 00:08:53.330 --> 00:08:56.309 Google's going to power Siri's AI search upgrade. 00:08:56.470 --> 00:08:59.230 Think about that, Reach. Google AI, potentially 00:08:59.230 --> 00:09:02.690 inside every iPhone Siri. That is massive. That 00:09:02.690 --> 00:09:04.970 could touch almost everyone. Okay, what else 00:09:04.970 --> 00:09:07.629 from Google? Google Photos beefed up its image 00:09:07.629 --> 00:09:10.269 -to -video feature, using VO3 now, apparently 00:09:10.269 --> 00:09:12.649 giving it more advanced capabilities. Turning 00:09:12.649 --> 00:09:15.549 snaps into clips gets fancier. Cool. Any drama. 00:09:15.830 --> 00:09:19.049 Always some drama. Scale AI is suing a former 00:09:19.049 --> 00:09:22.370 employee and a rival company, Merkur. Alleging 00:09:22.370 --> 00:09:24.809 customer theft. Standard growing pains in a hot 00:09:24.809 --> 00:09:27.309 sector, maybe? Could be. And hardware. Interesting 00:09:27.309 --> 00:09:30.490 move here. OpenAI is teaming up with Broadcom. 00:09:30.629 --> 00:09:33.830 Why? To make its own custom AI chips. Ah, vertical 00:09:33.830 --> 00:09:36.169 integration. Like Apple does. Exactly. Trying 00:09:36.169 --> 00:09:38.289 to control more of the hardware stack. Probably 00:09:38.289 --> 00:09:40.490 optimize performance right from the silicon up. 00:09:40.590 --> 00:09:43.169 Okay. Lots of moves. Out of these quick hits, 00:09:43.389 --> 00:09:47.360 OpenAI's valuation, the... Google -Apple -Siri 00:09:47.360 --> 00:09:51.019 deal, photos upgrade, the lawsuit, OpenAI's chip 00:09:51.019 --> 00:09:53.539 plans. Which one do you think points to the biggest 00:09:53.539 --> 00:09:56.440 future tech integration for most people, the 00:09:56.440 --> 00:09:58.419 one that will just blend into daily life? I got 00:09:58.419 --> 00:10:01.899 to say, Google and Apple teaming on Siri's AI 00:10:01.899 --> 00:10:05.350 search that hints at widespread. invisible impact. 00:10:05.590 --> 00:10:07.169 Yeah, I agree. That feels like one that could 00:10:07.169 --> 00:10:09.190 just happen to millions of users without them 00:10:09.190 --> 00:10:11.289 even thinking about it. Now, for a segment that 00:10:11.289 --> 00:10:13.690 really feels like it brings AI power right to 00:10:13.690 --> 00:10:16.470 your fingertips, and it ties back perfectly to 00:10:16.470 --> 00:10:18.509 our first chat about on -device capabilities. 00:10:19.049 --> 00:10:21.309 Ah, you're talking about Google's embedding Gemma. 00:10:21.370 --> 00:10:24.309 Exactly. This compact embedding model. Powerful 00:10:24.309 --> 00:10:26.929 private AI right on your local device. Yeah, 00:10:26.990 --> 00:10:29.129 the dream here is pretty awesome. Imagine running 00:10:29.129 --> 00:10:31.889 a full ARG pipeline. Remind us what ROG is again. 00:10:32.070 --> 00:10:34.879 Right. Retrieval Augmented Generation. So the 00:10:34.879 --> 00:10:37.019 AI finds relevant info from a knowledge base 00:10:37.019 --> 00:10:38.919 before it generates the answer. Think of it like 00:10:38.919 --> 00:10:41.980 giving the AI cheat sheets. Now imagine running 00:10:41.980 --> 00:10:44.460 that whole process, finding info, generating 00:10:44.460 --> 00:10:47.000 an answer directly on your phone. No internet 00:10:47.000 --> 00:10:50.279 needed, no cloud servers, no API calls. That's 00:10:50.279 --> 00:10:53.179 huge for privacy and speed. Total game changer. 00:10:53.320 --> 00:10:57.840 And Google, kind of quietly, just dropped embedding 00:10:57.840 --> 00:11:00.580 Gemma to make this possible. It's small, but 00:11:00.580 --> 00:11:03.149 apparently really mighty. How small are we talking? 00:11:03.370 --> 00:11:06.950 It's a 308 million parameter model. Tiny compared 00:11:06.950 --> 00:11:09.889 to the big guys, but specifically optimized for 00:11:09.889 --> 00:11:12.990 local devices, phones, laptops, desktops. And 00:11:12.990 --> 00:11:15.009 it's for embedding tasks. What does that mean 00:11:15.009 --> 00:11:17.330 in practice? Embeddings basically turn complex 00:11:17.330 --> 00:11:20.330 data, like words or sentences, into numbers vectors 00:11:20.330 --> 00:11:22.409 that capture their meaning. Makes it easy for 00:11:22.409 --> 00:11:24.889 computers to compare stuff. So embedding Gemma 00:11:24.889 --> 00:11:26.879 is really good at boosting search. Understanding 00:11:26.879 --> 00:11:29.240 what you mean, not just the keywords you type. 00:11:29.379 --> 00:11:31.919 And it works across languages. Get this. Trained 00:11:31.919 --> 00:11:34.639 on over 100 languages. Multilingual right out 00:11:34.639 --> 00:11:36.740 of the box. Plus, you can customize its output 00:11:36.740 --> 00:11:39.379 dimensions. Very versatile. Impressive. And the 00:11:39.379 --> 00:11:41.559 privacy aspect is built in because it's offline. 00:11:41.860 --> 00:11:44.840 Exactly. Runs completely offline. User privacy 00:11:44.840 --> 00:11:47.480 is baked in. And performance, it actually topped 00:11:47.480 --> 00:11:50.200 the MTEB leaderboard. That's the ranking system 00:11:50.200 --> 00:11:52.399 for these models. Yeah, for models under 500 00:11:52.399 --> 00:11:55.639 million parameters. It beat out competitors from... 00:11:55.820 --> 00:11:59.259 So here, Mistral, even OpenAI's smaller embedding 00:11:59.259 --> 00:12:01.740 options, packs a punch for its size. So what 00:12:01.740 --> 00:12:04.279 does that top performance mean for you, the user? 00:12:04.539 --> 00:12:07.740 Better semantic search, understanding the meaning 00:12:07.740 --> 00:12:10.460 behind your query, higher accuracy if you're 00:12:10.460 --> 00:12:13.559 using it for RAG, and crucially, fewer garbage 00:12:13.559 --> 00:12:16.820 answers. Just more relevant, precise results 00:12:16.820 --> 00:12:19.970 faster. And it's open source. Yep. Anyone can 00:12:19.970 --> 00:12:22.250 grab it, plug it into their local setup, customize 00:12:22.250 --> 00:12:25.070 it. This is especially big for enterprises. They 00:12:25.070 --> 00:12:27.129 lean heavily on RAG but haven't had great small 00:12:27.129 --> 00:12:30.269 models for on -device use until now. Embedding 00:12:30.269 --> 00:12:33.830 Gemma fills that gap. Two sec silence. Whoa. 00:12:34.429 --> 00:12:37.490 Okay, just thinking. Imagine scaling this power. 00:12:37.629 --> 00:12:40.269 A billion queries, maybe, on billions of devices, 00:12:40.529 --> 00:12:42.909 all running privately using this little model. 00:12:43.070 --> 00:12:44.950 It's kind of wild to think about the potential 00:12:44.950 --> 00:12:48.580 scale here. It really is. Boiling it down, how 00:12:48.580 --> 00:12:50.779 does Google's local embedding model fundamentally 00:12:50.779 --> 00:12:53.759 change what's possible for, say, phone -based 00:12:53.759 --> 00:12:56.460 AI applications? It enables genuinely powerful 00:12:56.460 --> 00:12:58.980 private AI functions directly on your device, 00:12:59.299 --> 00:13:01.440 sponsor. So let's try and connect the dots here. 00:13:01.539 --> 00:13:03.730 We've covered a lot of ground. What does this 00:13:03.730 --> 00:13:06.529 all mean for us looking at these different pieces? 00:13:06.809 --> 00:13:09.769 Well, it feels like this deep dive really highlighted 00:13:09.769 --> 00:13:13.169 a major theme, doesn't it? AI is evolving incredibly 00:13:13.169 --> 00:13:16.789 fast and there's this dual push. Dual push. Yeah. 00:13:16.830 --> 00:13:19.909 On one hand, making it more powerful and accessible, 00:13:20.070 --> 00:13:22.950 like with embedding Gemma on your phone or tools 00:13:22.950 --> 00:13:26.429 like Mirage for creation. But on the other hand, 00:13:26.470 --> 00:13:29.509 a really strong focus on addressing core safety 00:13:29.509 --> 00:13:32.590 and ethical concerns like that benevolent hack. 00:13:32.679 --> 00:13:35.639 to keep models behaving properly even after you 00:13:35.639 --> 00:13:37.759 shrink them down right so it's not just about 00:13:37.759 --> 00:13:40.480 capability it's about responsibility too exactly 00:13:40.480 --> 00:13:43.960 we're seeing this drive towards making AI more 00:13:43.960 --> 00:13:46.360 trustworthy more useful in everyday life whether 00:13:46.360 --> 00:13:49.019 it's helping you learn powering your phone search 00:13:49.019 --> 00:13:52.340 or transforming business operations it feels 00:13:52.340 --> 00:13:54.799 like building a foundation of responsible innovation 00:13:54.799 --> 00:13:57.519 that makes sense it's been an exciting look into 00:13:57.519 --> 00:13:59.940 where AI is heading showing just how quickly 00:13:59.940 --> 00:14:01.740 things are moving we definitely encourage you 00:14:01.740 --> 00:14:03.860 to explore these topics more, see how they connect 00:14:03.860 --> 00:14:06.399 to the technology you use every day? Yeah, and 00:14:06.399 --> 00:14:09.039 maybe a thought to leave you with. As AI gets 00:14:09.039 --> 00:14:11.500 smarter, as it weaves itself more deeply into 00:14:11.500 --> 00:14:14.340 our lives, what new responsibilities do we pick 00:14:14.340 --> 00:14:17.549 up? We as users. Developers? All of us, really. 00:14:17.669 --> 00:14:20.470 Users, developers, citizens. How do we actively 00:14:20.470 --> 00:14:23.629 shape its future safely and ethically? It's something 00:14:23.629 --> 00:14:26.289 worth mulling over as these tools become so central 00:14:26.289 --> 00:14:28.909 to everything. A really important question. Thank 00:14:28.909 --> 00:14:30.769 you for joining us for this deep dive. We appreciate 00:14:30.769 --> 00:14:32.809 you lending us your curiosity. Keep learning. 00:14:32.990 --> 00:14:33.830 Out Hero Music.