WEBVTT 00:00:02.239 --> 00:00:04.780 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 --> 00:00:11.300 breaking down what it means for your life and 00:00:11.300 --> 00:00:22.690 business. OpenAI has just bought back $7 billion 00:00:22.690 --> 00:00:26.989 worth of employee shares at about an $852 billion 00:00:26.989 --> 00:00:31.030 valuation. Anthropic's unreleased model has advanced 00:00:31.030 --> 00:00:33.909 the Riemann hypothesis. They spent 31 million 00:00:33.909 --> 00:00:37.369 tokens to do this. OpenAI's Astra model has solved 00:00:37.369 --> 00:00:41.509 a 10 long open math problem for $2 ,000 worth 00:00:41.509 --> 00:00:43.530 of tokens. So it wasn't cheap, but it was able 00:00:43.530 --> 00:00:46.869 to get it done. Alibaba released Quen 3 .8 Max, 00:00:47.090 --> 00:00:50.390 which is a 3 .4 trillion parameter model. It's 00:00:50.390 --> 00:00:53.109 basically competitive with Claude. Researchers 00:00:53.109 --> 00:00:55.990 right now are extracting hidden reasoning from 00:00:55.990 --> 00:00:58.409 Claude, ChatGPT, and Gemini. They're using a 00:00:58.409 --> 00:01:00.710 trick with the API, and they're comparing the 00:01:00.710 --> 00:01:03.350 reasoning of different models with the open source 00:01:03.350 --> 00:01:05.650 or open weight model. that are Chinese versions 00:01:05.650 --> 00:01:08.409 to see if they ripped off those models. Almost 00:01:08.409 --> 00:01:11.090 everyone I know is paying $20 a month for some 00:01:11.090 --> 00:01:13.829 AI tool, whether that's ChatGPT, Cloud, Gemini, 00:01:13.909 --> 00:01:16.329 Canva, there's so many out there. And the one 00:01:16.329 --> 00:01:18.370 thing that I will say to all of them is I would 00:01:18.370 --> 00:01:21.000 love for them to check out... AI box .ai, which 00:01:21.000 --> 00:01:23.540 is my own startup that lets you access 80 different 00:01:23.540 --> 00:01:26.379 AI models all on one platform. Our cheapest tier 00:01:26.379 --> 00:01:29.219 is $8 and 99 cents a month. It is incredibly 00:01:29.219 --> 00:01:31.620 cheap. It's the price of a coffee and you get 00:01:31.620 --> 00:01:33.819 access to over 80 different AI models. There's 00:01:33.819 --> 00:01:36.680 image models, there's audio models, there's video 00:01:36.680 --> 00:01:39.060 generation models like Google VO three. And of 00:01:39.060 --> 00:01:40.659 course you have all of the text and reasoning 00:01:40.659 --> 00:01:42.920 models like cloud chat, GPD and Gemini. If you 00:01:42.920 --> 00:01:44.760 want to get access to all of that in one place 00:01:44.760 --> 00:01:47.900 and also have an MCP, which basically is a link 00:01:47.900 --> 00:01:50.159 that lets If that's your main model that you 00:01:50.159 --> 00:01:52.659 use, generate images or let's cloud generate 00:01:52.659 --> 00:01:55.239 audio or let's cloud generate video. So you can 00:01:55.239 --> 00:01:57.340 pull any model into any model. If you want to 00:01:57.340 --> 00:01:59.939 check that out, it is AI box .ai. I'll leave 00:01:59.939 --> 00:02:02.900 a link in the description. OpenAI has just completed 00:02:02.900 --> 00:02:06.819 a $7 billion buyback of employee shares. They 00:02:06.819 --> 00:02:08.639 did this, which I think a lot of people are kind 00:02:08.639 --> 00:02:11.699 of shocked by, at an $852 billion valuation. 00:02:12.039 --> 00:02:14.319 Now, if you remember, this is the exact same 00:02:14.319 --> 00:02:16.599 valuation that they had back in March. We are 00:02:16.599 --> 00:02:18.840 now many months later. And if you look at a company 00:02:18.840 --> 00:02:20.979 like Anthropic, which month over month is adding 00:02:20.979 --> 00:02:24.379 like $100 billion to their valuation every single 00:02:24.379 --> 00:02:27.379 month, basically, this is not, I mean, you could 00:02:27.379 --> 00:02:29.240 look at this a couple different ways, but like 00:02:29.240 --> 00:02:31.610 technically that's not great. Their valuation 00:02:31.610 --> 00:02:35.250 hasn't grown since March. I think this is basically 00:02:35.250 --> 00:02:37.590 they're trying to keep the staff from leaving 00:02:37.590 --> 00:02:40.479 and going to Anthropic or going to Gemini. But 00:02:40.479 --> 00:02:42.659 also, I think this is basically showing they're 00:02:42.659 --> 00:02:45.020 not in a big rush to go public anytime soon. 00:02:45.199 --> 00:02:48.500 This $7 billion tender offer is one of the largest 00:02:48.500 --> 00:02:51.039 employee buybacks a private company has ever 00:02:51.039 --> 00:02:53.740 done. I think that's no shocker. But this is 00:02:53.740 --> 00:02:56.139 very consistent with what OpenAI has done. They 00:02:56.139 --> 00:02:58.860 have a pattern of doing really regular secondary 00:02:58.860 --> 00:03:01.659 sales. They've done this for the last three years 00:03:01.659 --> 00:03:03.840 where they'll let the employees sell at a regular 00:03:03.840 --> 00:03:07.400 basis. Sam Altman told the staff last month that 00:03:07.400 --> 00:03:10.120 OpenAI missed a bunch of internal financial targets 00:03:10.120 --> 00:03:12.560 over the last year, but he expects that the next 00:03:12.560 --> 00:03:14.439 year is going to be stronger. I mean, there's 00:03:14.439 --> 00:03:16.599 the elephant in the room and the no surprise 00:03:16.599 --> 00:03:19.219 is that Anthropic just absolutely mopped the 00:03:19.219 --> 00:03:21.539 floor with their growth. And OpenAI, I don't 00:03:21.539 --> 00:03:23.280 think has actually shrunk. They probably just 00:03:23.280 --> 00:03:25.840 went stagnant or they kind of plateaued for a 00:03:25.840 --> 00:03:28.479 minute. And the reason why is just because Anthropic 00:03:28.479 --> 00:03:30.939 grew so fast, they just took a lot of the oxygen 00:03:30.939 --> 00:03:32.580 out of the room. Is it going to stay that way 00:03:32.580 --> 00:03:35.719 forever? I personally don't. think so. I'm heavily 00:03:35.719 --> 00:03:38.620 testing both Claude and chat GPT. And right now, 00:03:38.639 --> 00:03:41.319 my main tool that I'm using is chat GPT work, 00:03:41.379 --> 00:03:44.139 I switched from Claude co work about two weeks 00:03:44.139 --> 00:03:46.919 ago. And I've been really impressed Claude co 00:03:46.919 --> 00:03:49.379 work or chat GPT work can do a lot of things 00:03:49.379 --> 00:03:51.560 that Claude co work didn't, it feels like opening, 00:03:51.639 --> 00:03:53.180 I felt pretty threatened, and they put a lot 00:03:53.180 --> 00:03:55.039 of time and energy into getting this right. So 00:03:55.039 --> 00:03:57.219 I love a good healthy rivalry, I think it was 00:03:57.219 --> 00:03:59.539 going to go back and forth. I'm excited if we 00:03:59.539 --> 00:04:01.560 have other competitors, if it feels like Google 00:04:01.560 --> 00:04:03.520 gets a little bit more in this space. You have 00:04:03.520 --> 00:04:06.479 perplexity in there. But overall, I think OpenAI 00:04:06.479 --> 00:04:09.639 has a superior product right now. And so I think 00:04:09.639 --> 00:04:11.419 they might start getting back a little bit of 00:04:11.419 --> 00:04:13.740 that enterprise that they've lost. Anthropic 00:04:13.740 --> 00:04:16.720 was reportedly profitable earlier this year. 00:04:16.819 --> 00:04:18.959 I think that's definitely going to threaten OpenAI 00:04:18.959 --> 00:04:22.339 as far as if Anthropic can reach the stock market 00:04:22.339 --> 00:04:24.040 first. If they can do their IPO first, they're 00:04:24.040 --> 00:04:27.170 going to get a lot of... a lot of buyers that 00:04:27.170 --> 00:04:30.230 would possibly have invested in OpenAI if not. 00:04:30.370 --> 00:04:33.230 So I think definitely there's a timing issue 00:04:33.230 --> 00:04:37.290 or there's a motivation to get to an IPO first 00:04:37.290 --> 00:04:40.250 for OpenAI and Anthropic. But right now, if they're 00:04:40.250 --> 00:04:42.389 holding their valuation flat and they're doing 00:04:42.389 --> 00:04:44.750 their tender offer for all their employees now 00:04:44.750 --> 00:04:47.589 rather than waiting for the IPO, OpenAI is basically 00:04:47.589 --> 00:04:50.089 buying time to prove that their business model 00:04:50.089 --> 00:04:52.209 is going to work at scale, especially for enterprise 00:04:52.209 --> 00:04:55.550 and their API products that have real... recurring 00:04:55.550 --> 00:04:58.149 revenue also i mean if you want to be like sort 00:04:58.149 --> 00:05:00.730 of pessimistic maybe you could say look, perhaps 00:05:00.730 --> 00:05:03.110 Sam Altman believes that the company is worth 00:05:03.110 --> 00:05:05.610 more and it's growing, but he just wants to buy 00:05:05.610 --> 00:05:07.509 back all the shares at a really good deal so 00:05:07.509 --> 00:05:10.430 he can go buy $7 billion worth of shares from 00:05:10.430 --> 00:05:13.290 the employees. And when they IPO, they can bump 00:05:13.290 --> 00:05:17.029 up the IPO price 50 % and now those $7 billion 00:05:17.029 --> 00:05:19.730 of shares will be worth $14 billion. The company 00:05:19.730 --> 00:05:22.509 that's worth over $850 billion, I don't know 00:05:22.509 --> 00:05:24.730 if that's really the main incentive here, but 00:05:24.730 --> 00:05:27.029 there's definitely a case to be made that perhaps 00:05:27.029 --> 00:05:29.189 those shares are at a discount. I think more 00:05:29.189 --> 00:05:30.910 like - Likely he's just trying to keep his employees 00:05:30.910 --> 00:05:34.089 happy and keep them around. But, you know, it's 00:05:34.089 --> 00:05:36.129 got to make you think that an IPO is not just 00:05:36.129 --> 00:05:38.930 weeks away if they're able to or if they're going 00:05:38.930 --> 00:05:40.750 to be buying back these shares from their employees. 00:05:40.870 --> 00:05:42.589 You know, if an IPO is around the corner, they 00:05:42.589 --> 00:05:44.430 probably just let them all sell on the open market. 00:05:45.259 --> 00:05:47.860 But if it's going to be months away, they probably 00:05:47.860 --> 00:05:50.939 will just buy them and wrap that up now. There's 00:05:50.939 --> 00:05:53.199 an unreleased anthropic model that is advancing 00:05:53.199 --> 00:05:55.180 the Riemann hypothesis. So basically this is 00:05:55.180 --> 00:05:58.079 a 150 -year -old math problem. There is a $1 00:05:58.079 --> 00:06:00.620 million bounty on this math problem, by the way, 00:06:00.639 --> 00:06:03.899 which is kind of cool that this exists. But by 00:06:03.899 --> 00:06:07.529 extending the range of numbers for... which it's 00:06:07.529 --> 00:06:09.769 been verified. This new model from Anthropic 00:06:09.769 --> 00:06:12.129 ran itself for a day and a half. It organized 00:06:12.129 --> 00:06:17.290 60 subagents to test 650 different ideas and 00:06:17.290 --> 00:06:20.230 formalize the results in a way that mathematicians 00:06:20.230 --> 00:06:21.970 could check because not only do you have to solve 00:06:21.970 --> 00:06:23.550 these problems, but you have to explain how you 00:06:23.550 --> 00:06:25.779 got it. So this is a big part of it. The model 00:06:25.779 --> 00:06:28.279 orchestrated its own workflow without a human 00:06:28.279 --> 00:06:30.759 mathematician directing it. It was assigning 00:06:30.759 --> 00:06:33.879 tasks like idea generation, validation, paper 00:06:33.879 --> 00:06:36.639 drafting to different subagents. It used about 00:06:36.639 --> 00:06:40.240 31 million tokens to produce this. Two anthropic 00:06:40.240 --> 00:06:42.759 mathematicians confirmed the results, which was 00:06:42.759 --> 00:06:45.379 then formalized in Lean, which is an open source 00:06:45.379 --> 00:06:47.839 proof assistant that lets other researchers verify 00:06:47.839 --> 00:06:51.500 it step by step. OpenAI's internal Astra model 00:06:51.980 --> 00:06:55.300 proved 10 major results this year and a separate 00:06:55.300 --> 00:06:58.120 anthropic effort disproved the Jacobian conjecture, 00:06:58.240 --> 00:07:02.660 which is a 1939 problem. So I think we're seeing 00:07:02.660 --> 00:07:06.459 a big shift in how AI is tackling big problems 00:07:06.459 --> 00:07:09.879 in math right now. I think AI is producing verifiable 00:07:09.879 --> 00:07:12.800 mathematical progress on a bunch of unsolved 00:07:12.800 --> 00:07:15.040 problems, but there's a lot of people in the 00:07:15.040 --> 00:07:17.120 math field that are debating what it means for 00:07:17.120 --> 00:07:20.600 credit and accountability when a machine is coordinating 00:07:20.600 --> 00:07:22.230 all this work. instead of a person, but also 00:07:22.230 --> 00:07:23.709 you can imagine, well, what if a person just 00:07:23.709 --> 00:07:26.709 used? You know, something like open AI, like 00:07:26.709 --> 00:07:28.470 right now, this is, you know, being done by the 00:07:28.470 --> 00:07:30.569 actual researchers are anthropic. But what would 00:07:30.569 --> 00:07:32.389 happen if, you know, they put out this unreleased 00:07:32.389 --> 00:07:34.790 model that's good at math and a regular person 00:07:34.790 --> 00:07:37.810 or a regular mathematician uses it to solve one 00:07:37.810 --> 00:07:39.889 of these problems? Who gets the credit for it? 00:07:39.949 --> 00:07:41.889 And maybe they say like, oh, no, I did it all 00:07:41.889 --> 00:07:44.709 by myself and I never even used AI. Would they 00:07:44.709 --> 00:07:47.529 get credit for it for solving these really hard 00:07:47.529 --> 00:07:49.730 kinds of problems? This is what everyone in the 00:07:49.730 --> 00:07:52.589 math field is currently debating. Now, I mentioned 00:07:52.589 --> 00:07:55.819 that. problem or those 10 problems that Astra 00:07:55.819 --> 00:07:59.800 solved. It basically the way this worked is this 00:07:59.800 --> 00:08:01.839 is also an unreleased model, by the way. So OpenAI 00:08:01.839 --> 00:08:03.920 and Anthropic, it's so funny. It feels like they 00:08:03.920 --> 00:08:06.899 do all their PR stunts like in at the same time. 00:08:07.019 --> 00:08:09.680 So it's like OpenAI goes and hacks Hugging Face 00:08:09.680 --> 00:08:11.360 and Anthropic's like, we hacked people. And then 00:08:11.360 --> 00:08:14.000 Meta's like, we hacked people too. And then Quinn's 00:08:14.000 --> 00:08:17.139 KimmyK3 is like, we hacked people or not Quinn, 00:08:17.180 --> 00:08:18.839 Moonshot's KimmyK3 is like, we hacked people 00:08:18.839 --> 00:08:21.079 as well with our unreleased model. And now we 00:08:21.079 --> 00:08:22.939 have like, I don't know, all of these different 00:08:22.939 --> 00:08:25.480 models are like, we solved, you know, these unsolvable 00:08:25.480 --> 00:08:27.480 hundreds. year old math problems. And they're 00:08:27.480 --> 00:08:30.040 like, Oh, we did too with our with our unpublished 00:08:30.040 --> 00:08:32.419 model. Anyways, I find it funny. But I mean, 00:08:32.440 --> 00:08:34.639 this one's pretty cool. So in this case, though, 00:08:34.679 --> 00:08:37.159 the Astra model, so not released, it solved 10 00:08:37.159 --> 00:08:39.559 long standing math problems that mathematicians 00:08:39.559 --> 00:08:42.279 had worked on for decades. It published a 250 00:08:42.279 --> 00:08:47.519 page proofs that verified with formal proof checking 00:08:47.519 --> 00:08:50.139 software, the solution like to actually get the 00:08:50.139 --> 00:08:52.480 solution to this cost about $2 ,000 in tokens 00:08:52.480 --> 00:08:55.100 to generate it. And on the one hand, I'm sure 00:08:55.100 --> 00:08:56.409 people are like, well, that's a lot of money 00:08:56.409 --> 00:08:58.649 to solve, you know, solve these math problems. 00:08:58.710 --> 00:09:01.269 I was like, if these are like famous math problems 00:09:01.269 --> 00:09:02.570 that have never been solved, they've been around 00:09:02.570 --> 00:09:05.230 for like hundreds of years, I mean, 2000 bucks, 00:09:05.250 --> 00:09:07.330 that seems pretty cheap, especially for a PR 00:09:07.330 --> 00:09:08.830 stunt like this. I mean, this is worth a lot 00:09:08.830 --> 00:09:11.289 more than $2 ,000 to open AI, this is probably 00:09:11.289 --> 00:09:15.179 worth, you know. uh, like 50 or a hundred million 00:09:15.179 --> 00:09:17.620 dollars worth of good PR for them. So it makes 00:09:17.620 --> 00:09:19.840 sense why they would do this. Um, but yeah, these 00:09:19.840 --> 00:09:23.100 are, the math is getting cracked by, um, these 00:09:23.100 --> 00:09:25.159 AI models. And this is stuff that's been, you 00:09:25.159 --> 00:09:27.779 know, pretty tricky for a very long time. Alibaba 00:09:27.779 --> 00:09:31.220 has just released Quen 3 .8 Max on Monday. This 00:09:31.220 --> 00:09:34.659 is a huge AI model. It has 2 .4 trillion parameters. 00:09:34.940 --> 00:09:38.799 It is on a bunch of benchmarks. It is right behind 00:09:38.799 --> 00:09:42.220 Anthropix Claude Fable 5 on the public leaderboard. 00:09:42.360 --> 00:09:44.080 So it's, I mean, it's right up there as number 00:09:44.080 --> 00:09:46.000 two. The company said that they're going to release 00:09:46.000 --> 00:09:48.799 the model's code and weights next week, making 00:09:48.799 --> 00:09:50.899 it free for developers. So any developer can 00:09:50.899 --> 00:09:53.340 go and download it. They can go modify it. I 00:09:53.340 --> 00:09:54.980 mean, the thing that's cool here, it's not like 00:09:54.980 --> 00:09:56.919 they're like, oh, we released a number two. model, 00:09:57.039 --> 00:09:58.980 that's great meta just released like a number 00:09:58.980 --> 00:10:01.700 four model yesterday or whatever. The thing that's 00:10:01.700 --> 00:10:04.019 amazing about it is that it's a number two model, 00:10:04.100 --> 00:10:06.379 and they're going to release the codes, open 00:10:06.379 --> 00:10:08.820 weights and like give this to anyone to go and 00:10:08.820 --> 00:10:12.539 download and use. on arena .ai's text leaderboard. 00:10:12.740 --> 00:10:17.139 It is behind Fable 5 and 3 Cloud Opus models. 00:10:18.039 --> 00:10:20.700 It's ahead of a bunch of other ones. Alibaba 00:10:20.700 --> 00:10:23.379 is coming to the openweight distribution. They 00:10:23.379 --> 00:10:25.840 had a little bit of a proprietary pivot earlier 00:10:25.840 --> 00:10:28.620 this year, which was kind of showing how a lot 00:10:28.620 --> 00:10:31.480 of Chinese labs like Moonshot and ByteDance were 00:10:31.480 --> 00:10:34.480 ever, there's like kind of this weird, and it 00:10:34.480 --> 00:10:35.659 feels like it happens in sync, but there's this 00:10:35.659 --> 00:10:38.259 weird moment for a lot of these. companies that 00:10:38.259 --> 00:10:39.879 are making these open weight models it feels 00:10:39.879 --> 00:10:41.379 like if they fall a little bit too far behind 00:10:41.379 --> 00:10:43.960 then they're like they go proprietary they fall 00:10:43.960 --> 00:10:45.580 a little bit behind then they're like okay now 00:10:45.580 --> 00:10:47.240 that we're getting caught up we got our mojo 00:10:47.240 --> 00:10:48.679 back we're going to release an open weight model 00:10:48.679 --> 00:10:53.659 even meta did that moonshots AI's Kimi K3 released 00:10:53.659 --> 00:10:56.940 last week had a 2 .8 trillion parameter model. 00:10:57.159 --> 00:10:59.340 So, I mean, this is 2 .4 trillion. It's very 00:10:59.340 --> 00:11:02.179 similar, but it's definitely, you know, Kimi 00:11:02.179 --> 00:11:06.279 K3 is a little bit more than Quen 3 .8 max. But 00:11:06.279 --> 00:11:08.759 the parameters don't 100 % matter, especially 00:11:08.759 --> 00:11:11.080 because this is currently ranking above it or 00:11:11.080 --> 00:11:13.600 on a bunch of different benchmarks. Researchers 00:11:13.600 --> 00:11:16.179 have found a way to extract hidden reasoning 00:11:16.179 --> 00:11:19.940 from Claude, GPT, and Gemini. This is a really 00:11:19.940 --> 00:11:22.179 interesting one for me. What they're doing is 00:11:22.179 --> 00:11:25.940 they go and trick smaller AI model variants into 00:11:25.940 --> 00:11:29.559 decrypting what larger models keep secret. So 00:11:29.559 --> 00:11:32.580 inside of any time that you ask an AI model like 00:11:32.580 --> 00:11:34.720 ChatGP or Cloud a question, it's not just running 00:11:34.720 --> 00:11:37.940 it to one model and spitting back the results. 00:11:38.120 --> 00:11:39.759 It takes your question and it gives it to like 00:11:39.759 --> 00:11:43.899 16 or a whole bunch of different sub models that 00:11:43.899 --> 00:11:45.639 all review it and they all come up with their 00:11:45.639 --> 00:11:48.779 own response. And then it sends that to another 00:11:48.779 --> 00:11:50.460 model that looks at all the responses and it 00:11:50.460 --> 00:11:53.639 can solve Anyways, there's like these orchestrations 00:11:53.639 --> 00:11:55.899 of models in the background looking at it and 00:11:55.899 --> 00:11:58.039 reasoning through it and trying to come up with 00:11:58.039 --> 00:12:01.389 the best response possible for you. But apparently 00:12:01.389 --> 00:12:05.490 that is the vulnerability. So the same flaw has 00:12:05.490 --> 00:12:09.289 also leaked API keys and passwords before the 00:12:09.289 --> 00:12:11.710 companies patched it last month. So essentially 00:12:11.710 --> 00:12:14.629 the main model was kind of better protected, 00:12:14.629 --> 00:12:16.490 but these little sub models where they could 00:12:16.490 --> 00:12:18.669 actually go and exploit them. This particular 00:12:18.669 --> 00:12:21.990 attack exploits shared decryption keys between 00:12:21.990 --> 00:12:24.710 large and small model versions. So you're feeding 00:12:24.710 --> 00:12:28.090 encrypted reasoning to a less restricted. sibling 00:12:28.090 --> 00:12:30.590 model and then you're making it decrypt and expose 00:12:30.590 --> 00:12:33.029 the hidden chain of thought from the bigger and 00:12:33.029 --> 00:12:35.110 better model that is you know more encrypted 00:12:35.110 --> 00:12:37.549 or you know higher security but the little models 00:12:37.549 --> 00:12:39.210 like you're using their own models against them 00:12:39.210 --> 00:12:42.110 kind of moonshots open weight model kimmy k3 00:12:42.110 --> 00:12:44.950 produced reasoning outputs that were very similar 00:12:44.950 --> 00:12:48.950 apparently to claude opus 4 .8 and gpt 4 .0 their 00:12:48.950 --> 00:12:52.049 reasoning traced across 90 different test questions 00:12:52.049 --> 00:12:56.309 so basically the idea here is that kimmy k3 probably 00:12:56.309 --> 00:12:58.710 use model distillation. I mean, this isn't a 00:12:58.710 --> 00:13:00.769 shocker. Why wouldn't they? They obviously use 00:13:00.769 --> 00:13:04.610 model distillation on Opus 4 .8 and GPT -4 .0. 00:13:05.120 --> 00:13:07.139 OpenAI, Anthropic, and Google all patched that 00:13:07.139 --> 00:13:09.580 particular vulnerability after they were notified. 00:13:10.159 --> 00:13:12.980 Researcher Alexander Panfulov says that some 00:13:12.980 --> 00:13:15.720 reasoning content can still be reconstructed 00:13:15.720 --> 00:13:18.700 and closing the hole entirely would require rebuilding 00:13:18.700 --> 00:13:21.659 how APIs handle offloaded computation. So in 00:13:21.659 --> 00:13:24.179 some of these cases, it's really, really hard 00:13:24.179 --> 00:13:27.259 to patch this, this kind of exploit. And, you 00:13:27.259 --> 00:13:28.500 know, they'd have to really rebuild a lot of 00:13:28.500 --> 00:13:31.399 their infrastructure and rebuild how APIs handle 00:13:31.399 --> 00:13:33.419 stuff altogether, which, you know, obviously 00:13:33.419 --> 00:13:36.840 that's a mess. problem. I think this exposes 00:13:36.840 --> 00:13:40.139 a really big structural weakness in how frontier 00:13:40.139 --> 00:13:42.419 models offload reasoning. The companies share 00:13:42.419 --> 00:13:45.399 infrastructure for cost reasons, but that shared 00:13:45.399 --> 00:13:48.019 infrastructure becomes a backdoor when alignment 00:13:48.019 --> 00:13:50.659 training is different. So the distillation angle 00:13:50.659 --> 00:13:52.799 matters most for policy. I think when we're talking 00:13:52.799 --> 00:13:54.779 about distillation, if U .S. models are being 00:13:54.779 --> 00:13:56.799 systematically copied, you know, by all these 00:13:56.799 --> 00:13:58.539 open weight competitors, which some of them are 00:13:58.539 --> 00:14:00.940 American, right? We have Meta doing that with 00:14:00.940 --> 00:14:03.919 Meta's Muse Spark. And then, of course, we have 00:14:03.919 --> 00:14:08.379 K3. But basically, this is a technique that is 00:14:08.379 --> 00:14:11.909 this particular. method is a very plausible avenue. 00:14:12.110 --> 00:14:13.870 A couple other interesting things that happened 00:14:13.870 --> 00:14:17.330 this week. Brad Lightcap, OpenAI's longtime COO, 00:14:17.389 --> 00:14:19.590 is leaving after eight years to start something 00:14:19.590 --> 00:14:22.850 new. And Anthropic is going to start watermarking 00:14:22.850 --> 00:14:24.750 claw -generated text. They have to do this to 00:14:24.750 --> 00:14:27.990 comply with the EU AI Act. So there's a lot going 00:14:27.990 --> 00:14:30.049 on. Thank you so much for tuning into the podcast. 00:14:30.230 --> 00:14:32.570 If you enjoyed this episode and you haven't left 00:14:32.570 --> 00:14:35.009 a review on the show yet, it helps the show out 00:14:35.009 --> 00:14:37.250 so much. I'm sure you hear me say this all the 00:14:37.250 --> 00:14:38.970 time, but honestly, if you're one of the people 00:14:38.970 --> 00:14:41.309 that hasn't left a review yet. It really helps 00:14:41.309 --> 00:14:44.090 me show up in the algorithm. I read them all. 00:14:44.149 --> 00:14:46.169 So I just appreciate hearing from you guys. If 00:14:46.169 --> 00:14:47.909 there's any topics you want me to cover or any 00:14:47.909 --> 00:14:50.129 things that you enjoy, drop it in, drop it in 00:14:50.129 --> 00:14:52.070 a review, leave a comment. I will read it. And 00:14:52.070 --> 00:14:54.009 I really appreciate it. Also make sure to check 00:14:54.009 --> 00:14:56.669 out AI box .ai. If you want to get access to 00:14:56.669 --> 00:14:59.330 over 80 different image, audio, video, and text 00:14:59.330 --> 00:15:01.850 models, all in one platform for $8 and 99 cents 00:15:01.850 --> 00:15:04.129 a month. We also have 20 % off annual subscriptions. 00:15:04.250 --> 00:15:05.509 And if you want to get more tokens, there's a 00:15:05.509 --> 00:15:07.029 bunch of different tiers. You could do some really, 00:15:07.190 --> 00:15:10.039 truly incredible. things with AI box dot AI, 00:15:10.100 --> 00:15:13.179 especially our MCP that lets you connect AI box 00:15:13.179 --> 00:15:15.860 to Claude and Claude can generate images, video 00:15:15.860 --> 00:15:18.440 and audio all inside of Claude, which is what 00:15:18.440 --> 00:15:20.440 I use it for a ton. All right, catch you guys 00:15:20.440 --> 00:15:21.519 all in the next episode.