WEBVTT 00:00:02.140 --> 00:00:04.799 Welcome to the AI Chat Podcast. I'm your host, 00:00:04.860 --> 00:00:07.179 Jaden Schaefer. Every day I cover cutting -edge 00:00:07.179 --> 00:00:09.320 AI news and talk with the leaders behind it, 00:00:09.339 --> 00:00:11.320 breaking down what it means for your life and 00:00:11.320 --> 00:00:17.980 business. Welcome to the podcast. SuperSafe Intelligence 00:00:17.980 --> 00:00:21.519 has landed a massive new deal with NVIDIA for 00:00:21.519 --> 00:00:25.579 a huge boost in computing. Claude Opus 4 .7 has 00:00:25.579 --> 00:00:28.579 just finished a two -week coding job. There's 00:00:28.579 --> 00:00:30.559 this basically new benchmark that has come out 00:00:30.559 --> 00:00:32.299 that is really exciting to me, and I'll tell 00:00:32.299 --> 00:00:35.759 you why. Google AI overviews now appear in 43 00:00:35.759 --> 00:00:38.939 % of searches. This is up 15 % from a year ago. 00:00:39.060 --> 00:00:40.960 Let's talk about how the landscape is shifting 00:00:40.960 --> 00:00:44.280 specifically in SEO. NVIDIA, Microsoft, and IBM 00:00:44.280 --> 00:00:47.560 have all launched Open Secure AI Alliance to 00:00:47.560 --> 00:00:50.140 defend agents. We're going to talk about why 00:00:50.140 --> 00:00:53.439 there's so much drama in that. Enigma has exited 00:00:53.439 --> 00:00:56.060 stealth, and they've just gotten $70 million 00:00:56.060 --> 00:01:00.159 to rethink how humans are talking to robots. 00:01:00.579 --> 00:01:03.299 Safe super intelligence is only - only two years 00:01:03.299 --> 00:01:05.420 old. It's the research lab that was founded by 00:01:05.420 --> 00:01:08.299 Ilya Seskovor after he left OpenAI, and they 00:01:08.299 --> 00:01:10.840 have secured a multi -billion dollar investment 00:01:10.840 --> 00:01:13.700 from NVIDIA and also access to their next generation 00:01:13.700 --> 00:01:16.939 Vera Rubin GPU platform. This deal in particular 00:01:16.939 --> 00:01:18.900 is basically going to increase their compute 00:01:18.900 --> 00:01:22.299 power by about 10x, which is a massive jump, 00:01:22.400 --> 00:01:24.420 and it's going to let them scale their research 00:01:24.420 --> 00:01:26.980 into AI safety and reasoning. And you know what's 00:01:26.980 --> 00:01:28.560 crazy about this company, if you've been following 00:01:28.560 --> 00:01:30.359 along, they literally have not even shipped a 00:01:30.359 --> 00:01:32.819 product yet. They don't have any revenue, and 00:01:32.819 --> 00:01:34.760 they've just gotten this multi -billion dollar 00:01:34.760 --> 00:01:37.260 investment from NVIDIA, really just writing off 00:01:37.260 --> 00:01:40.019 the fact that Ilya Suskovor is famous, was one 00:01:40.019 --> 00:01:42.900 of the OGs at OpenAI. And of course, he's, when 00:01:42.900 --> 00:01:45.859 I say famous, famous because he's a very incredible 00:01:45.859 --> 00:01:48.140 AI researcher. But still, it's, I mean, just 00:01:48.140 --> 00:01:50.640 really running off of the name, which is amazing. 00:01:51.019 --> 00:01:54.500 SSI has now raised $7 billion total. They have 00:01:54.500 --> 00:01:58.019 $32 billion for evaluation, and their backers 00:01:58.019 --> 00:02:01.250 are NVIDIA, Andreessen Horowitz, Alphabet. Sequoia, 00:02:01.290 --> 00:02:02.810 Lightspeed. So I mean, they really have all the 00:02:02.810 --> 00:02:05.250 biggest tech players. They've raised so much 00:02:05.250 --> 00:02:07.709 money and still no products have been shipped, 00:02:07.829 --> 00:02:10.409 no revenue has been generated, and they're continuing 00:02:10.409 --> 00:02:13.469 to raise more money and get more contracts. And 00:02:13.469 --> 00:02:15.930 just to be fair, this is actually what Ilya said 00:02:15.930 --> 00:02:17.590 when he started the company. He said, don't expect 00:02:17.590 --> 00:02:20.710 any products for, I think he said two to four 00:02:20.710 --> 00:02:22.849 years or something like that, which is kind of 00:02:22.849 --> 00:02:24.710 what happened with OpenAI, where there was a 00:02:24.710 --> 00:02:27.449 lot of money raised, a lot of research done, 00:02:27.590 --> 00:02:28.569 and then they started coming out with stuff. 00:02:28.650 --> 00:02:31.509 It looks like that's what SSI is. is doing vera 00:02:31.509 --> 00:02:34.009 rubin is nvidia's next generation gpu architecture 00:02:34.009 --> 00:02:38.090 and by getting this ssi is essentially going 00:02:38.090 --> 00:02:40.550 to put themselves as an early flagship customer 00:02:40.550 --> 00:02:42.389 they're also going to collaborate with nvidia 00:02:42.389 --> 00:02:45.189 on advancing future compute platforms and with 00:02:45.189 --> 00:02:47.849 the background of ilia i think for um nvidia 00:02:47.849 --> 00:02:49.990 this is like a really good kind of name brand 00:02:49.990 --> 00:02:52.030 company working with them on it and also it's 00:02:52.030 --> 00:02:54.750 going to help a lot with r d ssi also partners 00:02:54.750 --> 00:02:56.770 with google cloud which means that they have 00:02:56.770 --> 00:02:59.639 two of the largest compute suppliers in the which 00:02:59.639 --> 00:03:01.939 are funding their runway while they're working 00:03:01.939 --> 00:03:04.719 on a new foundational research that they're doing 00:03:04.719 --> 00:03:07.020 instead of shipping products. Suscover right 00:03:07.020 --> 00:03:09.039 now is really betting that some of this deep 00:03:09.039 --> 00:03:12.400 research on AI alignment and reasoning is going 00:03:12.400 --> 00:03:14.699 to help solve some of the problems. And so he 00:03:14.699 --> 00:03:15.900 doesn't have any of the product pressure that 00:03:15.900 --> 00:03:18.500 everyone else has, but the level of compute backing, 00:03:18.680 --> 00:03:19.900 and I think a lot of the investor confidence 00:03:19.900 --> 00:03:23.000 to me is signaling that, you know, maybe this 00:03:23.000 --> 00:03:26.039 is a crazy bubble or maybe he's onto something 00:03:26.039 --> 00:03:28.199 that we don't know. And definitely, I think there's 00:03:28.199 --> 00:03:32.360 some value. in the long game. Opus 4 .7 has implemented 00:03:32.360 --> 00:03:36.379 a 61 ,000 line Apple software program from scratch. 00:03:36.479 --> 00:03:38.479 I was rolling my eyes when I first saw this news 00:03:38.479 --> 00:03:41.469 story because it was like it did it in. 14 hours 00:03:41.469 --> 00:03:43.069 when it should have taken two weeks and did it 00:03:43.069 --> 00:03:45.889 for $250. Okay. What's cool to me is there's 00:03:45.889 --> 00:03:48.250 a new benchmark called mirror code. And it basically 00:03:48.250 --> 00:03:50.909 tests whether an AI can rebuild real production 00:03:50.909 --> 00:03:53.270 software. Um, and it doesn't, it's not allowed 00:03:53.270 --> 00:03:55.469 to get the code for that software. It basically 00:03:55.469 --> 00:03:58.349 is like, Hey, go look on your computer. Um, go 00:03:58.349 --> 00:04:01.069 try to copy like Apple, the Apple music app. 00:04:01.189 --> 00:04:03.669 You can look at it. Um, try to copy as best you 00:04:03.669 --> 00:04:05.430 can. You don't get to see any of the source code. 00:04:05.490 --> 00:04:07.509 You don't get access to the internet and it's, 00:04:07.509 --> 00:04:09.250 it's basically a black box. And how long does 00:04:09.250 --> 00:04:11.469 it take you to, to, build it and how much money 00:04:11.469 --> 00:04:14.030 does it cost? And so in this particular test, 00:04:14.349 --> 00:04:17.649 they were able to have Opus 4 .7 and also opening 00:04:17.649 --> 00:04:21.389 eyes GPT 5 .5. Both of them re -implemented GoTree, 00:04:21.470 --> 00:04:24.209 which is a 16 ,000 line parser. And they did 00:04:24.209 --> 00:04:26.589 it across a bunch of different programming languages. 00:04:26.689 --> 00:04:30.009 They did it for $100 to $400 each. What's interesting 00:04:30.009 --> 00:04:33.170 is they had like a bunch of other softwares that 00:04:33.170 --> 00:04:35.990 they worked on, but of the 25 different programs, 00:04:36.310 --> 00:04:39.329 17 that they were, that it was tasked to recreate. 00:04:39.399 --> 00:04:43.259 basically achieved perfect reimplementation and 00:04:43.259 --> 00:04:48.399 at least one run and four of them were 99 % perfect 00:04:48.399 --> 00:04:51.060 implementation. So, you know, between 17, like 00:04:51.060 --> 00:04:54.100 17 plus four, we're basically at, you know, almost 00:04:54.100 --> 00:04:57.139 the 25, a very high percentage. Only eight of 00:04:57.139 --> 00:04:59.000 them were unable to solve it. It's exciting to 00:04:59.000 --> 00:05:01.879 me just to see how far we've come. These AI models 00:05:01.879 --> 00:05:04.240 are now like the benchmark is how quickly can 00:05:04.240 --> 00:05:06.160 you recreate an entire piece of software. So 00:05:06.160 --> 00:05:08.769 I'm really excited about that. approach in particular 00:05:08.769 --> 00:05:10.850 and i think it's probably one of the better benchmarks 00:05:10.850 --> 00:05:14.459 that i have seen Google AI overviews now show 00:05:14.459 --> 00:05:17.519 up in 43 % of all searches. That's about triple 00:05:17.519 --> 00:05:20.040 the rate from a year ago. And I think this is 00:05:20.040 --> 00:05:22.339 a big shift. Google is becoming basically a destination 00:05:22.339 --> 00:05:24.920 where you get your answers answered directly, 00:05:25.000 --> 00:05:27.240 just like ChatGPT, instead of pointing you to 00:05:27.240 --> 00:05:29.040 other websites. And I mean, right now that's 00:05:29.040 --> 00:05:32.579 43%. I think we're going to go to 80, 90 % in 00:05:32.579 --> 00:05:35.600 the coming couple years. Google AI mode visits 00:05:35.600 --> 00:05:39.300 jumped 121 % in the last 11 months. That's climbing 00:05:39.300 --> 00:05:43.720 from 126 million in June of last year. to $279 00:05:43.720 --> 00:05:46.850 million in May of this year. So publishers, publishers 00:05:46.850 --> 00:05:49.850 are definitely losing referral traffic because 00:05:49.850 --> 00:05:51.689 users are just going to read that snippet at 00:05:51.689 --> 00:05:52.850 the top of the Google, right? They don't have 00:05:52.850 --> 00:05:54.629 to go and scroll down and actually click on it. 00:05:54.990 --> 00:05:56.410 Even to the point where like, sometimes I'll 00:05:56.410 --> 00:05:59.209 see an article and you know, not proud of it, 00:05:59.250 --> 00:06:00.490 but it'll be like Wall Street Journal. I don't 00:06:00.490 --> 00:06:01.870 have a public, I don't have like a subscription. 00:06:02.149 --> 00:06:04.290 I see the title. I'm like, you know, that's super 00:06:04.290 --> 00:06:06.089 interesting. I paste the title into Google and 00:06:06.089 --> 00:06:07.589 Google is kind of getting sued. All of them are 00:06:07.589 --> 00:06:09.209 getting sued for summarizing these paywalled 00:06:09.209 --> 00:06:12.110 articles. So I, but so I'll just throw it in 00:06:12.110 --> 00:06:13.290 there and be like, Hey, like, give me a summary 00:06:13.290 --> 00:06:15.589 of this article. It doesn't do. that. So if you 00:06:15.589 --> 00:06:18.990 paywall the article, that might be the only secret 00:06:18.990 --> 00:06:21.910 solution to getting people to really come, assuming 00:06:21.910 --> 00:06:23.870 you're still relevant and you have an exclusive 00:06:23.870 --> 00:06:25.990 story that no one else has. And I mean, that's 00:06:25.990 --> 00:06:27.930 just for news stories. For most general knowledge, 00:06:28.069 --> 00:06:30.209 that's really not going to be a thing. So ChatGPT 00:06:30.209 --> 00:06:33.129 right now is sending way less users to external 00:06:33.129 --> 00:06:37.420 websites. Only 6 .8 % of chat GPT searches include 00:06:37.420 --> 00:06:40.740 citations as of May of this year, which is Google 00:06:40.740 --> 00:06:43.240 versus Google that has a way higher rate. So, 00:06:43.279 --> 00:06:45.019 you know, if everyone's moving to these AI tools, 00:06:45.199 --> 00:06:48.220 there's just a lot less traffic going to websites. 00:06:48.560 --> 00:06:51.899 All right, big news from NVIDIA, Microsoft, and 00:06:51.899 --> 00:06:53.879 IBM. They've all launched what is called Open 00:06:53.879 --> 00:06:57.899 Secure AI, and it is an alliance to defend agents. 00:06:58.459 --> 00:07:00.639 This is something they just rolled out, and it's 00:07:00.639 --> 00:07:03.339 basically an open source cybersecurity tool for... 00:07:03.339 --> 00:07:06.339 AI agents. The group of them are all arguing 00:07:06.339 --> 00:07:09.860 that closed AI systems block defenders from investigating 00:07:09.860 --> 00:07:12.160 breaches. And they're right now lobbying against 00:07:12.160 --> 00:07:15.060 government restrictions on open AI that could 00:07:15.060 --> 00:07:17.660 leave only a bunch of companies controlling some 00:07:17.660 --> 00:07:19.519 critical security tools. It's interesting that 00:07:19.519 --> 00:07:20.959 they have to do the lobbying because obviously 00:07:20.959 --> 00:07:23.899 the critical security tools are like open AI 00:07:23.899 --> 00:07:27.199 and anthropic. And yeah, it's interesting. There's 00:07:27.199 --> 00:07:29.860 also 30 other people that are in this as well. 00:07:30.000 --> 00:07:33.800 Hugging Faces uses an open weight GL. 5 .2 model 00:07:33.800 --> 00:07:37.920 to analyze 17 ,000 actions during their own security 00:07:37.920 --> 00:07:40.680 breach after a closed AI system refused to run 00:07:40.680 --> 00:07:42.680 the forensic analysis. I mean, basically what 00:07:42.680 --> 00:07:45.579 happened there was OpenAI made a model, it escaped 00:07:45.579 --> 00:07:47.899 containment, went to Hugging Faces and tried 00:07:47.899 --> 00:07:51.279 to hack Hugging Faces. Hugging Faces saw that 00:07:51.279 --> 00:07:53.519 there was a hack or breach going on. They tried 00:07:53.519 --> 00:07:56.079 to deploy another OpenAI agent against it to 00:07:56.079 --> 00:07:58.779 stop it. The guardrails on that agent blocked 00:07:58.779 --> 00:08:00.839 them from defending themselves and they literally 00:08:00.839 --> 00:08:03.250 had to go to an open weight Chinese model GLM 00:08:03.250 --> 00:08:06.189 5 .2 to stop the breach. That's crazy. Agents 00:08:06.189 --> 00:08:09.189 versus agents. But I think it really shows the 00:08:09.189 --> 00:08:11.490 value of some of these open source or open weight 00:08:11.490 --> 00:08:15.329 models. NVIDIA is open sourcing NOOA, NVIDIA 00:08:15.329 --> 00:08:17.670 Labs Object Oriented Agent Framework. I know 00:08:17.670 --> 00:08:19.629 it's a mouthful. They're doing that on GitHub, 00:08:19.790 --> 00:08:22.550 but that should make AI agent behavior easier 00:08:22.550 --> 00:08:25.389 to test, trace, and audit, which is a huge hole 00:08:25.389 --> 00:08:27.750 in the market. Right now, this group of companies, 00:08:27.790 --> 00:08:29.529 or people call it like an alliance or whatever, 00:08:29.689 --> 00:08:33.529 but they're directly in opposition to OpenAI 00:08:33.529 --> 00:08:35.669 and Anthropic's call for restrictions on open 00:08:35.669 --> 00:08:39.090 -weight AI. And they're arguing that these kind 00:08:39.090 --> 00:08:41.950 of limitations would concentrate power to a few 00:08:41.950 --> 00:08:44.409 closed providers. So it's interesting, though, 00:08:44.429 --> 00:08:46.169 because it feels like OpenAI and Anthropic obviously 00:08:46.169 --> 00:08:47.850 have the best models. These other people are 00:08:47.850 --> 00:08:50.539 working. on models, but they're not the leaders. 00:08:50.679 --> 00:08:52.659 And so it feels like because they're not the 00:08:52.659 --> 00:08:55.559 leaders, maybe they want to sabotage the other 00:08:55.559 --> 00:08:56.919 companies. Meanwhile, the companies that are 00:08:56.919 --> 00:08:58.480 ahead are like, nope, we need restrictions to 00:08:58.480 --> 00:09:00.840 hold us in place. So there's like a lot of drama 00:09:00.840 --> 00:09:02.919 that goes on behind that. But overall, I think 00:09:02.919 --> 00:09:06.399 this is a good direction by Microsoft, IBM, and 00:09:06.399 --> 00:09:08.759 the rest of the crew. A company called Enigma 00:09:08.759 --> 00:09:11.600 is a startup that was founded by a former Microsoft 00:09:11.600 --> 00:09:15.539 and Israeli intelligence veterans, and they just 00:09:15.539 --> 00:09:19.100 raised $70 million to test a new bet that robots 00:09:19.100 --> 00:09:22.179 fail, not because AI models are weak, but because 00:09:22.179 --> 00:09:24.059 the way humans control them is really clunky. 00:09:24.120 --> 00:09:26.840 So right now they're opening 100 of their own 00:09:26.840 --> 00:09:29.399 robots to anyone on the internet to remote control 00:09:29.399 --> 00:09:31.700 them and to generate data on what actually works. 00:09:31.700 --> 00:09:33.639 So this is fascinating, right? I mean, we have 00:09:33.639 --> 00:09:36.059 like, crowdsource data collection is something 00:09:36.059 --> 00:09:38.200 that is being worked on. But now we have like, 00:09:39.089 --> 00:09:41.190 crowdsourced robot control here's our robots 00:09:41.190 --> 00:09:43.049 you guys control them and we're going to collect 00:09:43.049 --> 00:09:45.409 data on how that works this is a really cool 00:09:45.409 --> 00:09:48.029 company to me because enigma built both the robotic 00:09:48.029 --> 00:09:50.789 arms and the ai models that are powering them 00:09:50.789 --> 00:09:52.850 they did all from scratch this is i mean impressive 00:09:52.850 --> 00:09:55.250 for a company that's only a year old and with 00:09:55.250 --> 00:09:57.549 this users basically there's like a portal and 00:09:57.549 --> 00:09:59.850 you can remotely direct the robots to draw pictures 00:09:59.850 --> 00:10:02.129 it can fence with swords it can run chemistry 00:10:02.129 --> 00:10:04.450 experiments um it's all through the internet 00:10:04.450 --> 00:10:06.450 there's like a hanger in israel and a hanger 00:10:06.450 --> 00:10:08.299 in california that have these robots in them 00:10:08.299 --> 00:10:10.679 that are doing all of this. So they just raised 00:10:10.679 --> 00:10:13.700 the $70 million, and that is going to fund data 00:10:13.700 --> 00:10:16.379 collection, experiments on which input methods 00:10:16.379 --> 00:10:18.600 work best, voice, text, video, all that kind 00:10:18.600 --> 00:10:22.139 of stuff, and the manipulation of what the robots 00:10:22.139 --> 00:10:25.059 can see and do. So fascinating project. Guys, 00:10:25.100 --> 00:10:26.659 thank you so much for tuning into the podcast 00:10:26.659 --> 00:10:28.740 today. If you enjoyed this episode, make sure 00:10:28.740 --> 00:10:31.620 to go check out the AI Box MCP that lets you 00:10:31.620 --> 00:10:33.899 get access to over 80 different AI models and 00:10:33.899 --> 00:10:35.919 put them inside of the tools you're already using. 00:10:36.019 --> 00:10:38.659 So for example, Claude gets access to generating 00:10:38.659 --> 00:10:41.340 images, audio, and video right inside of Claude. 00:10:41.379 --> 00:10:44.419 I use this every single day. On Sunday, my wife 00:10:44.419 --> 00:10:47.539 wanted to generate coloring books for the kids. 00:10:47.679 --> 00:10:51.340 And so I got it to generate 600 coloring book 00:10:51.340 --> 00:10:54.539 pages, one for like every chapter. the Bible 00:10:54.539 --> 00:10:55.980 that, that she's going to be reading to them 00:10:55.980 --> 00:10:57.500 over the next little while. So anyways, there's 00:10:57.500 --> 00:10:58.759 so many different projects. And if you want to 00:10:58.759 --> 00:11:00.500 do projects like that at scale, where you have 00:11:00.500 --> 00:11:04.039 to do 600 images or, you know, a thousand articles 00:11:04.039 --> 00:11:06.960 about something using an AI box, MCP connection 00:11:06.960 --> 00:11:09.399 speeds up the process so much. And you get audio 00:11:09.399 --> 00:11:11.740 video and image that Claude can't naturally do. 00:11:11.799 --> 00:11:14.120 It also works inside of chat GPT or any of the 00:11:14.120 --> 00:11:17.039 other tools as well. So all the AI models inside 00:11:17.039 --> 00:11:19.879 of everything, and you can do a really fast run. 00:11:19.960 --> 00:11:21.259 So go check it out. There's a link in the description 00:11:21.259 --> 00:11:25.519 to AI box .ai slash MC. And also it's only $8 00:11:25.519 --> 00:11:27.860 .99 a month to get started. So check that out 00:11:27.860 --> 00:11:29.159 and I'll catch you all in the next episode.