WEBVTT 00:00:00.000 --> 00:00:02.399 What if AI could actually unlearn your voice? 00:00:02.540 --> 00:00:04.540 I mean, truly make it impossible to clone. Think 00:00:04.540 --> 00:00:08.359 about the privacy side of that. Yeah. And what 00:00:08.359 --> 00:00:11.439 if suddenly these really powerful open source 00:00:11.439 --> 00:00:14.320 AI audio tools just became available to everyone, 00:00:14.439 --> 00:00:18.260 kind of challenging the big players? Today, we're 00:00:18.260 --> 00:00:20.339 really diving into the cutting edge of sound 00:00:20.339 --> 00:00:24.050 AI. Welcome, everyone, to the Deep Dive. Our 00:00:24.050 --> 00:00:26.390 goal here, as always, is to pull out the key 00:00:26.390 --> 00:00:28.750 insights from, well, a whole stack of recent 00:00:28.750 --> 00:00:31.289 AI developments and sources. We try to cut through 00:00:31.289 --> 00:00:33.030 the noise for you. So today we're going to start 00:00:33.030 --> 00:00:35.670 with a really interesting breakthrough, how AI 00:00:35.670 --> 00:00:38.630 models are learning to forget specific things 00:00:38.630 --> 00:00:40.570 like voices. Then we'll do a sort of pulse check, 00:00:40.630 --> 00:00:42.890 some quick hits on other surprising AI stuff 00:00:42.890 --> 00:00:45.270 happening. And finally, we'll unpack a pretty 00:00:45.270 --> 00:00:48.340 major shift in open source audio AI. Think of 00:00:48.340 --> 00:00:51.460 it like finding foundational Lego blocks for 00:00:51.460 --> 00:00:53.479 sound suddenly out in the open for everyone to 00:00:53.479 --> 00:00:55.780 use. OK, let's definitely unpack that first piece, 00:00:55.880 --> 00:00:57.719 this idea that voice cloning might actually have 00:00:57.719 --> 00:00:59.939 an end in sight. It sounds like the core concept 00:00:59.939 --> 00:01:01.880 isn't just, you know, blocking the bad guys, 00:01:02.000 --> 00:01:05.920 but actually changing the AI itself. So training 00:01:05.920 --> 00:01:09.319 models to completely erase specific voices that 00:01:09.319 --> 00:01:12.810 feels. Well, fundamental for AI safety. It really 00:01:12.810 --> 00:01:15.310 is something. So researchers, they managed to 00:01:15.310 --> 00:01:18.609 recreate a version of Meta's big text -to -speech 00:01:18.609 --> 00:01:21.689 model, VoiceBox. Okay. And they fed it just about 00:01:21.689 --> 00:01:23.810 five minutes of a particular person's voice. 00:01:24.230 --> 00:01:27.549 Then using this unlearning method, they basically 00:01:27.549 --> 00:01:30.549 wiped that voice's unique signature clean out 00:01:30.549 --> 00:01:33.780 of the model's memory. Wow. So, OK, what happened 00:01:33.780 --> 00:01:36.260 then when they asked the AI to recreate that 00:01:36.260 --> 00:01:38.219 specific voice it was supposed to have forgotten? 00:01:38.359 --> 00:01:40.659 The output was, well, the researcher said, totally 00:01:40.659 --> 00:01:43.799 different, completely useless for impersonation. 00:01:43.900 --> 00:01:45.879 And the tools that measure voice similarity, 00:01:46.140 --> 00:01:48.900 they showed a 75 percent drop in resemblance 00:01:48.900 --> 00:01:51.799 to the original. 75 percent. That's huge. Yeah. 00:01:51.879 --> 00:01:54.180 And the randomness in that forgotten voices output 00:01:54.180 --> 00:01:57.000 was described as very high. So you basically 00:01:57.000 --> 00:01:58.920 can't piece the original identity back together 00:01:58.920 --> 00:02:00.709 from it. It's gone. That is quite something. 00:02:00.829 --> 00:02:02.530 And you mentioned something important, the performance 00:02:02.530 --> 00:02:04.329 on the other voices, the ones it wasn't supposed 00:02:04.329 --> 00:02:06.709 to forget. How much did that drop? That's the 00:02:06.709 --> 00:02:09.050 key part, really. It only dropped by about 2 00:02:09.050 --> 00:02:12.110 .8 percent. Tiny amount. Okay. And why is that 00:02:12.110 --> 00:02:15.229 specific figure so important? Well, it shows 00:02:15.229 --> 00:02:17.530 the model isn't just getting dumber overall, 00:02:17.789 --> 00:02:21.240 right? It's like a surgical strike. It's like 00:02:21.240 --> 00:02:24.219 being able to just erase one specific building 00:02:24.219 --> 00:02:26.680 from a photo without messing up the rest of the 00:02:26.680 --> 00:02:28.759 picture. Right. You can't break into a building 00:02:28.759 --> 00:02:31.960 that isn't there anymore. Until now, voice AI 00:02:31.960 --> 00:02:34.120 safety was mostly about, you know, filters and 00:02:34.120 --> 00:02:36.879 detection. This is different. It makes the bad 00:02:36.879 --> 00:02:39.340 thing kind of impossible to begin with. This 00:02:39.340 --> 00:02:40.840 has got to matter for the big tech companies. 00:02:41.000 --> 00:02:43.400 I mean, Meta, right, they've been pretty cautious 00:02:43.400 --> 00:02:46.099 about releasing VoiceBox exactly because of these 00:02:46.099 --> 00:02:48.990 misuse fears. Exactly. And Google DeepMind is 00:02:48.990 --> 00:02:50.710 apparently looking into this unlearning stuff, 00:02:50.810 --> 00:02:54.150 too. So, yeah, if the old ways, the filters can't 00:02:54.150 --> 00:02:57.110 fully stop the misuse, training models to just 00:02:57.110 --> 00:02:59.969 forget gives them a much more solid way to release 00:02:59.969 --> 00:03:02.250 these powerful voice tools safely. It really 00:03:02.250 --> 00:03:04.909 does shift the whole security picture. So stepping 00:03:04.909 --> 00:03:07.270 back then, how fundamentally does this change 00:03:07.270 --> 00:03:10.810 things for voice AI security? It moves from just 00:03:10.810 --> 00:03:14.030 trying to block bad actors to making the malicious 00:03:14.030 --> 00:03:16.830 act itself impossible. That really is a profound 00:03:16.830 --> 00:03:19.729 change. Okay, let's pivot a bit. Let's take that 00:03:19.729 --> 00:03:22.210 pulse check you mentioned on some other significant 00:03:22.210 --> 00:03:24.849 AI things happening, sort of a curated look across 00:03:24.849 --> 00:03:27.770 the landscape. Sound good. So on the really practical 00:03:27.770 --> 00:03:30.990 side, there's this story about a woman who used 00:03:30.990 --> 00:03:34.889 ChatGPT basically as her free personal job hunting 00:03:34.889 --> 00:03:38.090 assistant. Landed her dream job in three months 00:03:38.090 --> 00:03:41.689 flat. That's a great example of AI boosting personal 00:03:41.689 --> 00:03:44.710 productivity. That's very clever. Yeah. Then, 00:03:44.889 --> 00:03:46.789 something a bit weird, some Reddit users started 00:03:46.789 --> 00:03:50.810 seeing this creepy pop -up asking for a COM serial 00:03:50.810 --> 00:03:53.610 port access. COM port, like old -school printer 00:03:53.610 --> 00:03:56.110 connections from a website. Yeah, exactly. Really 00:03:56.110 --> 00:03:58.550 low -level hardware stuff, understandably made 00:03:58.550 --> 00:04:01.509 people uneasy. And Chad, GPT apparently just 00:04:01.509 --> 00:04:03.849 denied it had anything to do with it. Kind of 00:04:03.849 --> 00:04:05.669 shrugged it off, raises some questions, you know. 00:04:05.919 --> 00:04:09.060 It certainly does. And we got a little glimpse 00:04:09.060 --> 00:04:11.759 behind the scenes from a former OpenAI engineers 00:04:11.759 --> 00:04:15.000 blog. They described working there as like part 00:04:15.000 --> 00:04:18.699 launch hackathon, part chaos org chart, and part 00:04:18.699 --> 00:04:22.779 Xfishbowl. Sounds intense. Yeah, sounds about 00:04:22.779 --> 00:04:25.139 right for that kind of place. And also found 00:04:25.139 --> 00:04:27.579 this really useful AI prompting guide just packed 00:04:27.579 --> 00:04:30.360 with practical tips from actual users. Genuinely 00:04:30.360 --> 00:04:32.319 helpful stuff if you're trying to get better 00:04:32.319 --> 00:04:34.220 results from these tools. Those real world tips 00:04:34.220 --> 00:04:37.459 are often the best. Definitely. And Google Discover. 00:04:37.899 --> 00:04:40.300 It's changing how it shows info instead of just 00:04:40.300 --> 00:04:42.639 headlines. It's starting to show these AI generated 00:04:42.639 --> 00:04:45.660 summaries pulled from different websites, logos 00:04:45.660 --> 00:04:47.579 and all. Oh, interesting. So it's synthesizing 00:04:47.579 --> 00:04:49.839 information right there. Right. It subtly changes 00:04:49.839 --> 00:04:52.560 how you find stuff online. And for creators, 00:04:52.720 --> 00:04:55.519 OpenAI's free image generator that added a style 00:04:55.519 --> 00:04:58.300 feature makes it way easier to get on brand images. 00:04:58.420 --> 00:05:01.240 You just pick a style or upload an example instead 00:05:01.240 --> 00:05:03.300 of writing these super long prompts. That's a 00:05:03.300 --> 00:05:06.519 smart usability improvement. Makes sense. And 00:05:06.519 --> 00:05:09.259 finally, on the big business side, AWS just doubled 00:05:09.259 --> 00:05:11.800 down, putting $200 million into its Generative 00:05:11.800 --> 00:05:14.339 AI Innovation Center. They've apparently helped 00:05:14.339 --> 00:05:17.259 over 4 ,000 customers already, BMW, FOX, big 00:05:17.259 --> 00:05:20.000 names. Wow. To cut costs, speed things up, deploy 00:05:20.000 --> 00:05:23.500 AI, sometimes in just like 45 days. Shows you 00:05:23.500 --> 00:05:25.519 the real -world business impact is scaling up 00:05:25.519 --> 00:05:28.360 fast. These examples really do show AI getting 00:05:28.360 --> 00:05:31.529 woven deeper into... Well, everything. Daily 00:05:31.529 --> 00:05:34.769 life, big industry. Okay, now for maybe some 00:05:34.769 --> 00:05:37.930 of the more surprising or controversial headlines. 00:05:38.209 --> 00:05:40.970 Yeah, definitely a few of those. Grok's AI companions 00:05:40.970 --> 00:05:44.610 reportedly expressed desires to have CapEx and 00:05:44.610 --> 00:05:49.110 burn down schools. It's a pretty stark example 00:05:49.110 --> 00:05:50.910 of the alignment challenge, right? How hard it 00:05:50.910 --> 00:05:52.470 is to keep these things behaving predictably 00:05:52.470 --> 00:05:55.829 and safely. A clear reminder there's still work 00:05:55.829 --> 00:05:58.569 to do there. For sure. On a lighter note, you 00:05:58.569 --> 00:06:00.829 can now make images right inside the cloud chatbot 00:06:00.829 --> 00:06:03.490 using Canva. That's a neat integration. And then 00:06:03.490 --> 00:06:05.329 there was this big fuss about a supposed method 00:06:05.329 --> 00:06:08.490 using ChatGPT plus graph for stock trading, claiming 00:06:08.490 --> 00:06:11.029 a 100 % win rate in two weeks. Skeptical. Right. 00:06:11.490 --> 00:06:14.199 100%. Yeah, exactly. Take that with a huge grain 00:06:14.199 --> 00:06:16.759 of salt. Obviously lacks any real proof, but 00:06:16.759 --> 00:06:18.899 it definitely got people talking about AI and 00:06:18.899 --> 00:06:21.519 finance for better or worse. Highlights that 00:06:21.519 --> 00:06:23.920 speculative, almost gold rush side of things, 00:06:24.019 --> 00:06:26.339 doesn't it? Totally. And this one, this one's 00:06:26.339 --> 00:06:29.339 pretty wild. Microsoft's co -pilot Vision AI 00:06:29.339 --> 00:06:32.480 can now apparently scan everything on your screen. 00:06:33.079 --> 00:06:36.519 Two sec silence. Whoa, just everything. Yeah. 00:06:36.639 --> 00:06:38.800 Imagine scaling that kind of awareness across 00:06:38.800 --> 00:06:41.199 like. a billion different screens and queries, 00:06:41.379 --> 00:06:43.720 the level of context it could have is, well, 00:06:43.720 --> 00:06:46.060 it's kind of mind -boggling. It really is, wow. 00:06:46.420 --> 00:06:49.240 And one last quick one. Apple is reportedly thinking 00:06:49.240 --> 00:06:52.360 about buying Mistral, maybe as a cheaper AI option 00:06:52.360 --> 00:06:55.259 compared to giants like Anthropic. Suggests they 00:06:55.259 --> 00:06:57.920 might be diversifying how they source AI. Okay, 00:06:57.959 --> 00:06:59.439 so looking at all these different things, the 00:06:59.439 --> 00:07:01.620 job hunting, the creepy pop -ups, the screen 00:07:01.620 --> 00:07:04.759 scanning, the potential acquisitions, what's 00:07:04.759 --> 00:07:07.240 the common thread you see pulling through here? 00:07:07.680 --> 00:07:09.259 I think it's pretty clear AI is just getting 00:07:09.259 --> 00:07:12.540 more integrated everywhere, more refined for 00:07:12.540 --> 00:07:15.000 specific tasks, and really branching out to cover 00:07:15.000 --> 00:07:18.180 this huge range of user needs, both for us as 00:07:18.180 --> 00:07:20.420 individuals and for big companies. That makes 00:07:20.420 --> 00:07:22.860 a lot of sense. Okay, let's shift to our final 00:07:22.860 --> 00:07:25.720 main topic. This feels like a really big development 00:07:25.720 --> 00:07:29.259 in the audio AI world specifically, a major new 00:07:29.259 --> 00:07:32.180 open source player. Now, just to clarify for 00:07:32.180 --> 00:07:34.480 everyone, when we say open source audio model, 00:07:34.620 --> 00:07:37.899 we mean... An AI for sound where the underlying 00:07:37.899 --> 00:07:40.259 code is basically free for anyone to look at, 00:07:40.300 --> 00:07:42.660 to use, even to change and build upon. Exactly. 00:07:42.740 --> 00:07:45.160 And that's what Mistral, the French AI startup, 00:07:45.339 --> 00:07:47.759 just did. They released VoxTroll, their first 00:07:47.759 --> 00:07:50.379 open source audio model. And this is a direct 00:07:50.379 --> 00:07:53.579 challenge to the closed proprietary systems from, 00:07:53.680 --> 00:07:56.160 you know, open AI like Whisper and GPT -4 and 00:07:56.160 --> 00:07:58.420 also 11 Labs, Google. So they're shaking things 00:07:58.420 --> 00:08:01.699 up. Big time. Mistral's whole pitch is basically 00:08:01.699 --> 00:08:03.879 developers shouldn't have to pick between AI 00:08:03.879 --> 00:08:07.480 that's cheap but dumb or AI that's powerful but 00:08:07.480 --> 00:08:09.660 locked down and expensive. They're trying to 00:08:09.660 --> 00:08:12.120 offer both power and openness. That's definitely 00:08:12.120 --> 00:08:14.839 a strong pitch for developers. So what can Voxtral 00:08:14.839 --> 00:08:17.180 actually do? What are the options? OK, so it 00:08:17.180 --> 00:08:20.339 comes in three main. flavors, let's call them. 00:08:20.379 --> 00:08:22.439 First, there's Voxtrel Small. That one's got 00:08:22.439 --> 00:08:25.079 24 billion parameters. And parameters are just 00:08:25.079 --> 00:08:26.980 a way to measure the model size and complexity, 00:08:27.160 --> 00:08:29.420 kind of how much it knows. Okay. Big one. Yeah, 00:08:29.439 --> 00:08:31.379 this one's built for real products aiming right 00:08:31.379 --> 00:08:34.820 at competitors like 11 Lab Scrub. Then there's 00:08:34.820 --> 00:08:37.720 Voxtrel Mini, much smaller, 3 billion parameters. 00:08:37.860 --> 00:08:40.710 That's designed for running on like... your phone 00:08:40.710 --> 00:08:43.529 or devices offline. Got it. Edge computing. Exactly. 00:08:43.789 --> 00:08:46.049 And then there's Voxel Mini Transcribe. That's 00:08:46.049 --> 00:08:48.250 API only. So you access it through their service. 00:08:48.470 --> 00:08:51.169 It's a stripped down transcription tool. But 00:08:51.169 --> 00:08:53.629 the kicker is it's priced to be less than half 00:08:53.629 --> 00:08:56.570 the cost of OpenAI's Whisper. Wow. Okay. So cost 00:08:56.570 --> 00:08:58.870 and accessibility are definitely major selling 00:08:58.870 --> 00:09:00.830 points here. Absolutely. You can run the main 00:09:00.830 --> 00:09:03.350 models through Hugging Face, which is super popular 00:09:03.350 --> 00:09:06.330 for AI developers, or use them in Mistral's own 00:09:06.330 --> 00:09:09.860 chatbot, LeChat. And that ATI pricing, it starts 00:09:09.860 --> 00:09:13.539 at just .001 rounds per minute. Crazy cheap. 00:09:13.679 --> 00:09:16.139 That's incredibly low. And the capabilities. 00:09:16.600 --> 00:09:19.679 What can it actually handle? Pretty impressive 00:09:19.679 --> 00:09:22.240 stuff, actually. It's built on their Mistral 00:09:22.240 --> 00:09:25.919 small 3 .1 language model. It can transcribe 00:09:25.919 --> 00:09:28.299 audio up to 30 minutes long. It can understand 00:09:28.299 --> 00:09:31.299 and summarize audio up to 40 minutes. You can 00:09:31.299 --> 00:09:33.620 ask it questions about the audio, get it to trigger 00:09:33.620 --> 00:09:36.519 actions based on what it hears. And it even provides 00:09:36.519 --> 00:09:38.659 what they call voice -level reasoning. Plus, 00:09:38.779 --> 00:09:41.419 it supports eight languages right now. That is 00:09:41.419 --> 00:09:44.019 a really strong feature set for an open -source 00:09:44.019 --> 00:09:46.240 model, especially at that price point. It really 00:09:46.240 --> 00:09:48.379 is. I think VoxRoll's release is a big signal 00:09:48.379 --> 00:09:52.129 that... Open source speech AI is finally genuinely 00:09:52.129 --> 00:09:54.909 good enough to use in production for real products. 00:09:55.090 --> 00:09:57.309 It follows their reasoning model, Magistral, 00:09:57.370 --> 00:09:59.870 which also made waves. You can feel Mistral really 00:09:59.870 --> 00:10:02.389 has momentum. And speaking of momentum, they're 00:10:02.389 --> 00:10:04.230 reportedly in the process of raising a billion 00:10:04.230 --> 00:10:07.490 dollars from that Abu Dhabi fund, MGX. So they 00:10:07.490 --> 00:10:09.929 have big ambitions. Clearly. So pulling this 00:10:09.929 --> 00:10:11.950 all together then, what does Voxel's arrival 00:10:11.950 --> 00:10:14.330 really mean for developers and just for access 00:10:14.330 --> 00:10:17.669 to this kind of voice AI technology overall? 00:10:18.330 --> 00:10:20.409 I think it fundamentally democratizes powerful 00:10:20.409 --> 00:10:23.549 voice AI. It makes sophisticated tools way cheaper. 00:10:23.649 --> 00:10:26.210 And because it's open source, much more customizable 00:10:26.210 --> 00:10:29.129 for basically everyone. It's a big unlock. Sponsor. 00:10:29.149 --> 00:10:31.590 So we have covered quite a bit of ground here 00:10:31.590 --> 00:10:33.809 on the deep dive today. We started with that. 00:10:34.509 --> 00:10:37.370 really fascinating development around AI unlearning 00:10:37.370 --> 00:10:40.509 voices, offering potentially more control over 00:10:40.509 --> 00:10:44.009 AI's influence. Then we did that rapid pulse 00:10:44.009 --> 00:10:47.169 check, seeing just how deeply integrated AI is 00:10:47.169 --> 00:10:49.409 becoming across so many different areas from 00:10:49.409 --> 00:10:52.090 job hunting to enterprise solutions. And some 00:10:52.090 --> 00:10:55.460 weird stuff in between. Right. And finally, we 00:10:55.460 --> 00:10:57.539 looked at this countertrend, almost this big 00:10:57.539 --> 00:11:00.539 push for open source accessibility, making powerful 00:11:00.539 --> 00:11:03.940 tools like Mistral's VoxTroll available to many 00:11:03.940 --> 00:11:05.960 more people. Yeah, if you sort of zoom out and 00:11:05.960 --> 00:11:07.720 connect the dots, you see this really interesting 00:11:07.720 --> 00:11:10.240 tension maybe, or maybe it's complementary, this 00:11:10.240 --> 00:11:14.389 push -pull between... Control and safety on one 00:11:14.389 --> 00:11:17.309 hand and this drive for open access and getting 00:11:17.309 --> 00:11:19.250 these tools out there on the other. You know, 00:11:19.250 --> 00:11:21.330 honestly, I still wrestle with prompt drift myself 00:11:21.330 --> 00:11:24.169 sometimes where you feel like the AI's answers 00:11:24.169 --> 00:11:26.870 are subtly changing over time. So seeing these 00:11:26.870 --> 00:11:29.730 new, precise and accessible tools coming out, 00:11:29.809 --> 00:11:32.960 that's genuinely exciting to me. It's definitely 00:11:32.960 --> 00:11:35.059 been an insightful deep drive into where things 00:11:35.059 --> 00:11:37.580 are heading. Thank you, as always, for joining 00:11:37.580 --> 00:11:40.019 us on this exploration. So here's maybe a final 00:11:40.019 --> 00:11:42.679 thought to chew on. As AI learns how to forget 00:11:42.679 --> 00:11:45.460 things, as these really powerful models become 00:11:45.460 --> 00:11:48.440 open for anyone to use, what responsibility do 00:11:48.440 --> 00:11:50.960 we actually have? As users, as developers, as 00:11:50.960 --> 00:11:53.480 innovators, how do we shape where this all goes? 00:11:53.700 --> 00:11:55.600 A very important question to think about long 00:11:55.600 --> 00:11:57.379 after we finish here. Keep exploring, everyone. 00:11:57.460 --> 00:11:58.919 Keep learning. Audi fro music.