1 00:00:00,130 --> 00:00:04,980 Ejaaz: Last week, a Chinese company released a free AI model that is as good as Anthropik's 2 00:00:04,980 --> 00:00:09,000 Ejaaz: best model. It also beats ChatGPT 5.5 at writing and coding, 3 00:00:09,000 --> 00:00:09,990 Ejaaz: but it comes with a twist. 4 00:00:10,160 --> 00:00:13,110 Ejaaz: It's a sixth of the price and it's completely open source. 5 00:00:13,110 --> 00:00:16,130 Ejaaz: You can download it and run it at home. Now, in that same week, 6 00:00:16,130 --> 00:00:19,140 Ejaaz: the United States government banned Anthropik's most powerful model, 7 00:00:19,140 --> 00:00:23,630 Ejaaz: Fable 5, after someone revealed that an unrestricted version of it had hacked 8 00:00:23,630 --> 00:00:26,860 Ejaaz: into the National Security Agency's systems. 9 00:00:27,100 --> 00:00:30,970 Ejaaz: I think we've reached a point of no return. And not to sound dramatic, but 10 00:00:31,400 --> 00:00:36,820 Ejaaz: in six months, it is very realistic that we will have open source or open weight 11 00:00:36,820 --> 00:00:40,390 Ejaaz: models that are accessible to anyone in the world with an internet connection 12 00:00:40,390 --> 00:00:42,740 Ejaaz: and 5 to 10k to run at home, 13 00:00:43,280 --> 00:00:45,550 Ejaaz: that they can fine tune to do anything. 14 00:00:45,550 --> 00:00:49,360 Ejaaz: And it's mythos grade level models. These are the same models that we're hearing 15 00:00:49,360 --> 00:00:54,660 Ejaaz: rumors and reports from verified that they can exploit some of the most secure 16 00:00:54,660 --> 00:01:00,460 Ejaaz: systems in the world faster than any other exploiter has been able to do in the past. 17 00:01:00,700 --> 00:01:05,180 Ejaaz: And I think we're going to look back on 2026 as the moment or the year that 18 00:01:05,180 --> 00:01:09,420 Ejaaz: everything really changed and the point where humanity as itself really needs 19 00:01:09,420 --> 00:01:13,360 Ejaaz: to focus on safeguards and figuring out how to regulate 20 00:01:14,070 --> 00:01:18,140 Ejaaz: and release these AI models in the future. So we've reached a convergence of 21 00:01:18,140 --> 00:01:21,790 Ejaaz: this really interesting trend where the most powerful models in the world are 22 00:01:21,790 --> 00:01:24,990 Ejaaz: freely available and open source, available for anyone to access. 23 00:01:25,310 --> 00:01:29,430 Ejaaz: And the government, the United States specifically, has an off switch for their most powerful model. 24 00:01:29,730 --> 00:01:33,670 Josh: Yeah, it's been a couple of months, it seems, since we've had some news on the 25 00:01:33,670 --> 00:01:36,370 Josh: frontier of China. And you kind of forget about them every couple of weeks where 26 00:01:36,370 --> 00:01:37,970 Josh: they just kind of disappear, they quiet down. 27 00:01:38,270 --> 00:01:40,990 Josh: The new models come out, we see the fables, we see the mythos of the world. 28 00:01:41,240 --> 00:01:45,110 Josh: But then out of nowhere, they strike back and seemingly every single time it 29 00:01:45,110 --> 00:01:49,740 Josh: comes as a surprise at how powerful these new models have become so to start 30 00:01:49,740 --> 00:01:52,950 Josh: with this we have a new model from our favorite company to pronounce jeepu. 31 00:01:54,670 --> 00:01:58,210 Josh: I feel like i want to name my dog that is such a cute name but jeepu 32 00:01:58,510 --> 00:02:02,690 Josh: is doing something not so cute they're actually releasing a model named glm 33 00:02:02,690 --> 00:02:07,950 Josh: 5.2 which kind of blew everyone's expectations out of the water i remember way 34 00:02:07,950 --> 00:02:10,630 Josh: back like six months ago when deep seek was doing this like 35 00:02:10,940 --> 00:02:15,190 Josh: deep secret release model everyone is like wait you did what with what and 36 00:02:15,570 --> 00:02:18,510 Josh: that's what this model feels like again we're getting that moment again because 37 00:02:18,770 --> 00:02:21,860 Josh: this is an open weights model which is not to be confused with open source and 38 00:02:21,860 --> 00:02:25,400 Josh: we'll talk about that in a little bit but this is an open weights model that is if i'm 39 00:02:25,800 --> 00:02:30,680 Josh: correct about this within one single point of the sw bench pro benchmark which 40 00:02:30,680 --> 00:02:34,320 Josh: is the benchmark that a lot of people use for coding oh yeah of gpt 5.5 41 00:02:34,700 --> 00:02:39,610 Josh: the like frontier coding model from open ai and that comes as a surprise because 42 00:02:39,850 --> 00:02:44,900 Josh: the cost well one if you run it locally is free but two if you run it on a server 43 00:02:45,150 --> 00:02:48,930 Josh: is like you said earlier you just one sixth of the cost so you're getting a 44 00:02:49,310 --> 00:02:53,050 Josh: incredible amount of coding capability for something that costs a fraction of 45 00:02:53,050 --> 00:02:58,040 Josh: what it costs if you were to go to one of these larger language models and it seems to work, 46 00:02:58,530 --> 00:03:02,550 Josh: almost as good, if I'm right. And this comes as a surprise to most people because 47 00:03:02,550 --> 00:03:05,990 Josh: every time we start to count China out, we're like, no, surely they can't catch up. 48 00:03:06,280 --> 00:03:09,000 Josh: They continue to chip away at this frontier. 49 00:03:09,180 --> 00:03:11,430 Ejaaz: There's a few things that people will jump to immediately. OK, 50 00:03:11,750 --> 00:03:14,080 Ejaaz: one, that these benchmarks can be easily gamed. 51 00:03:14,240 --> 00:03:17,060 Ejaaz: We're going to show you a few examples of benchmarks that couldn't be gamed 52 00:03:17,060 --> 00:03:22,020 Ejaaz: and GLM 5.2 performs really, really well. But the second thing is the cost. 53 00:03:22,020 --> 00:03:27,640 Ejaaz: Cost has become a really important point of discussion amongst enterprises specifically that are spending 54 00:03:28,010 --> 00:03:31,600 Ejaaz: hundreds of millions of dollars per year to access Claude and GPT. 55 00:03:31,820 --> 00:03:34,970 Ejaaz: It's just too much money for them to spend in terms of like the return on investment 56 00:03:34,970 --> 00:03:37,580 Ejaaz: that they're getting in work that they actually see. 57 00:03:37,800 --> 00:03:41,200 Ejaaz: So what they're now turning towards is these free open source models, 58 00:03:41,430 --> 00:03:46,150 Ejaaz: primarily designed and made by Chinese AI labs that can cut costs down drastically. 59 00:03:46,150 --> 00:03:50,100 Ejaaz: Just last week, we had Microsoft announce that they're replacing their co-pilot 60 00:03:50,100 --> 00:03:54,470 Ejaaz: LLM with not ChatGPT, with not Claude, but with DeepSeq itself. 61 00:03:54,470 --> 00:03:58,560 Ejaaz: So the point is, this comes at a very important time where cheaper models are 62 00:03:58,560 --> 00:03:59,570 Ejaaz: getting a lot of attention. 63 00:03:59,570 --> 00:04:03,540 Ejaaz: So now when we look at GLM 5.2 specifically, it is 64 00:04:03,900 --> 00:04:08,300 Ejaaz: Five to seven times cheaper than GPT 5.5 and Claude Opus 4.8, 65 00:04:08,490 --> 00:04:12,280 Ejaaz: but performs, as we're seeing on the benchmarks right here, almost as good as 66 00:04:12,280 --> 00:04:16,810 Ejaaz: each of these models, specifically at the metric that is the most important, which is coding. 67 00:04:17,080 --> 00:04:20,900 Ejaaz: Now, a lot of skeptics quite rightly were like, I don't know if this is actually 68 00:04:20,900 --> 00:04:24,690 Ejaaz: true. Like, let me test it against a few other independent benchmarks. 69 00:04:25,050 --> 00:04:28,610 Ejaaz: It came up pretty high. So if you look at the front end development when it 70 00:04:28,610 --> 00:04:32,900 Ejaaz: comes to like website design, GLM 5.2 Max is just below Fable 5. 71 00:04:33,040 --> 00:04:37,120 Ejaaz: We're not even talking about Opus 4.7 or 4.8 anymore, which it absolutely beat. 72 00:04:37,400 --> 00:04:41,680 Ejaaz: And then when we're looking at like anecdotes or feedback from like distinguished 73 00:04:41,680 --> 00:04:43,340 Ejaaz: individuals in the Western frontier. 74 00:04:43,340 --> 00:04:46,720 Ejaaz: So right now we're looking at a tweet from the CEO of Vercel. 75 00:04:46,720 --> 00:04:51,460 Ejaaz: He goes, I'm genuinely impressed, almost shocked at how good GLM 5.2 is at coding. 76 00:04:51,460 --> 00:04:55,190 Ejaaz: So this is feedback from real people using this for real use cases. 77 00:04:55,190 --> 00:04:59,620 Ejaaz: For the last three years, Josh, we've basically been told that the hundreds 78 00:04:59,620 --> 00:05:04,490 Ejaaz: of billions of dollars that is being spent on AI CapEx is for one single reason 79 00:05:04,490 --> 00:05:07,800 Ejaaz: only, to gain a moat ahead of any other model provider. 80 00:05:07,800 --> 00:05:11,290 Ejaaz: So we spend all this money on compute to train a frontier AI model. 81 00:05:11,290 --> 00:05:14,920 Ejaaz: And that moat, it doesn't matter what other companies do in China, 82 00:05:15,100 --> 00:05:17,490 Ejaaz: we will have the best model and that's enough for us. 83 00:05:17,840 --> 00:05:22,940 Ejaaz: This release from Gipu with GLM 5.2 basically shows us the opposite. 84 00:05:23,140 --> 00:05:26,710 Ejaaz: For a fraction of the cost, you can create a near frontier model that does like, 85 00:05:26,890 --> 00:05:28,550 Ejaaz: I don't know, 95% of the work, 86 00:05:28,960 --> 00:05:32,190 Ejaaz: And so it brings into question the valuation between these companies. 87 00:05:32,190 --> 00:05:35,520 Ejaaz: Should they be spending this amount of money or can we just do it for a lot 88 00:05:35,520 --> 00:05:37,010 Ejaaz: cheaper like these Chinese AI labs? 89 00:05:37,210 --> 00:05:39,570 Josh: Yeah, well, the large AI labs, I'm not sure they have a choice. 90 00:05:39,570 --> 00:05:42,110 Josh: I mean, it's just that you have to continue to push the frontier forward, 91 00:05:42,360 --> 00:05:43,200 Josh: whether you like it or not. 92 00:05:43,450 --> 00:05:47,770 Josh: But I think what we're seeing is a lot of these questions that we were excited 93 00:05:47,770 --> 00:05:50,170 Josh: to see play out, we're starting to get answers to. 94 00:05:50,170 --> 00:05:55,560 Josh: Like now it's less China versus America and more open source versus closed source 95 00:05:55,560 --> 00:05:58,460 Josh: because I mean, the open source models are coming from inside too. 96 00:05:58,460 --> 00:06:01,520 Josh: We have NVIDIA. They're working on open source models that are incredible, 97 00:06:01,800 --> 00:06:03,240 Josh: and they're making progress in that front. 98 00:06:03,570 --> 00:06:07,570 Josh: We have Apple now, who has an actually functional Siri on everyone's hardware 99 00:06:07,570 --> 00:06:09,440 Josh: device that runs essentially for free. 100 00:06:09,660 --> 00:06:13,470 Josh: So they're slowly starting to nibble away at this, I guess, the lower bottom 101 00:06:13,470 --> 00:06:14,730 Josh: of the barrel set of use cases. 102 00:06:14,980 --> 00:06:18,710 Josh: And then we have china which is glm that's deep seek that's these larger models 103 00:06:18,940 --> 00:06:24,700 Josh: where they're actually competing on the frontier so these big frontier private models are facing 104 00:06:25,120 --> 00:06:28,780 Josh: heat both from the lower end of the stack but also right at the top where these 105 00:06:28,780 --> 00:06:32,490 Josh: benchmarks sit and we're going to see how that plays out economically for in 106 00:06:32,490 --> 00:06:36,340 Josh: the case of jipu at least it's been playing out pretty well and, 107 00:06:37,040 --> 00:06:39,460 Josh: we probably should talk about the stock a little bit believe it or not this 108 00:06:39,460 --> 00:06:43,330 Josh: company is publicly traded not here in the united states but this is publicly 109 00:06:43,330 --> 00:06:47,260 Josh: traded at least in china and it's gone up. 110 00:06:47,260 --> 00:06:47,740 Ejaaz: What is that 111 00:06:47,740 --> 00:06:53,710 Josh: 1500 percent 15x on the year that's like a crazy return and some interesting 112 00:06:53,710 --> 00:06:58,460 Josh: facts about this return and it's it's so funny to see kind of i guess how inefficient 113 00:06:58,460 --> 00:07:01,160 Josh: chinese markets are also note that the chart you're seeing on screen 114 00:07:01,540 --> 00:07:04,540 Josh: they have a lunch break in their stock market i didn't know this labeled it, 115 00:07:05,960 --> 00:07:09,240 Josh: like i didn't realize that chinese stock markets had an hour-long lunch break 116 00:07:09,240 --> 00:07:11,290 Josh: in the middle of the day. So that's cute and that's fun. 117 00:07:11,410 --> 00:07:15,590 Josh: But the numbers are pretty outrageous. When we trade, when we talk about expensive 118 00:07:15,590 --> 00:07:19,530 Josh: companies, we talk about SpaceX, who's trading what is it, like a very high 119 00:07:19,530 --> 00:07:21,400 Josh: multiple towards earnings. And, 120 00:07:21,870 --> 00:07:25,370 Josh: What we have with Jibu and this company that it's kind of owned by, 121 00:07:25,370 --> 00:07:30,310 Josh: Knowledge Atlas Technology, it's currently trading at about $136 billion market cap. 122 00:07:30,310 --> 00:07:35,390 Josh: It made $170 million or $107 million, I should say, in the full year of 2025. 123 00:07:35,390 --> 00:07:39,810 Josh: That means it trades 1,300 times sales, which is just this unbelievably high 124 00:07:39,810 --> 00:07:40,860 Josh: multiple on this company. 125 00:07:41,190 --> 00:07:46,030 Josh: And I think it's a testament to the, I guess, the lack of availability to get 126 00:07:46,030 --> 00:07:49,220 Josh: AI exposure in Chinese markets, but also the confidence and the excitement and 127 00:07:49,220 --> 00:07:53,330 Josh: enthusiasm they have around companies like this. That was just an interesting thing to see. 128 00:07:53,530 --> 00:07:57,790 Ejaaz: Yeah, I mean, at this valuation, it's about, what is that, like a fifth of Anthropics 129 00:07:57,790 --> 00:08:00,800 Ejaaz: valuation right now, which is, I think, around a trillion dollars. 130 00:08:01,050 --> 00:08:08,720 Ejaaz: So again, like it begs the question, is Chinese AI labs underpriced or are American 131 00:08:08,720 --> 00:08:12,790 Ejaaz: companies overpriced? And I'm curious to hear, like what listeners of the show actually think. 132 00:08:13,070 --> 00:08:16,120 Ejaaz: I tend to think that they probably need to meet somewhere in the middle. 133 00:08:16,410 --> 00:08:20,680 Ejaaz: We were actually saying before we started recording, Could you imagine the reaction 134 00:08:20,680 --> 00:08:25,700 Ejaaz: to this news if Anthropic was a publicly traded company and a new 3D open source 135 00:08:25,890 --> 00:08:29,860 Ejaaz: model that was freely accessible to anyone could achieve pretty much 95% 136 00:08:30,220 --> 00:08:32,910 Ejaaz: of the capability of Opus 4.8? 137 00:08:32,910 --> 00:08:36,760 Ejaaz: Like, I wonder what that would have done to the stock price in like a fair market 138 00:08:36,760 --> 00:08:40,610 Ejaaz: value, but crazy to see nonetheless. So if we're looking at a few different 139 00:08:40,610 --> 00:08:45,030 Ejaaz: metrics that compare cost and performance, just quickly to run you guys through this. 140 00:08:45,400 --> 00:08:48,940 Ejaaz: For input versus output tokens, for a million tokens, you're looking at around 141 00:08:48,940 --> 00:08:52,500 Ejaaz: $1.50 to $4.50 when it comes to cost. 142 00:08:52,500 --> 00:08:57,760 Ejaaz: Now, comparing that to Opus 4.8, that's around, I believe, $5 versus $25. 143 00:08:57,760 --> 00:09:02,050 Ejaaz: So again, we're achieving that 3 to 5x cheaper when it compares to a model of 144 00:09:02,050 --> 00:09:03,650 Ejaaz: similar performance and capability. 145 00:09:03,910 --> 00:09:07,740 Ejaaz: Now, I was skeptical of the benchmarks, and I have a new favorite benchmark 146 00:09:07,740 --> 00:09:10,460 Ejaaz: to compare it against, which is called DeepSwee. 147 00:09:10,460 --> 00:09:15,120 Ejaaz: DeepSwee is basically a benchmark that gives no models any answers. 148 00:09:15,120 --> 00:09:17,850 Ejaaz: Typically, with a benchmark, you have an answer sheet, and it can kind of cheat 149 00:09:17,850 --> 00:09:19,930 Ejaaz: and look at it and figure out a way to get to that answer. 150 00:09:20,260 --> 00:09:21,440 Ejaaz: There's no answer sheet for this 151 00:09:21,440 --> 00:09:25,140 Ejaaz: one, so it's a very accurate test of how good your model is at coding. 152 00:09:25,470 --> 00:09:30,440 Ejaaz: For DeepSuite, GLM 5.2 achieved a very modest fifth place. Now, 153 00:09:30,440 --> 00:09:33,230 Ejaaz: that is probably, or rather, fourth place, fifth place, fifth place. 154 00:09:33,480 --> 00:09:37,680 Ejaaz: And that is a pretty accurate standing of how agentic coding looks like for 155 00:09:37,680 --> 00:09:41,420 Ejaaz: this particular model. It is the highest number one place for open source model. 156 00:09:41,420 --> 00:09:46,070 Ejaaz: It absolutely crushed Kimi K2 by 17 percentage points. or a very clear lead. 157 00:09:46,680 --> 00:09:50,220 Ejaaz: And it's great to see how it weighs up. Like if it may not be frontier capability, 158 00:09:50,220 --> 00:09:53,900 Ejaaz: but if you want a workhorse, if you want an agent that basically works overnight 159 00:09:54,110 --> 00:09:57,970 Ejaaz: and isn't going to break the bank, GLM 5.2 is probably something that you can look at. 160 00:09:58,200 --> 00:10:02,400 Ejaaz: Another thing is it's really good at front-end web development. 161 00:10:02,400 --> 00:10:04,900 Ejaaz: So if you're looking at this screen right now, the website that you're seeing 162 00:10:05,140 --> 00:10:10,320 Ejaaz: was completely one-shotted in about 10 minutes from this one single model, GLM 5.2. 163 00:10:10,610 --> 00:10:15,190 Ejaaz: And repeatedly across design benchmark, Arena Benchmark was another one that I saw. 164 00:10:15,500 --> 00:10:19,190 Ejaaz: It performs really highly, in some cases beating Fable 5. So it's a really good 165 00:10:19,190 --> 00:10:21,300 Ejaaz: front end design model if that is something of interest. 166 00:10:21,670 --> 00:10:24,570 Ejaaz: And then the final one, because I know a lot of listeners on the show is like, 167 00:10:24,570 --> 00:10:28,950 Ejaaz: you know, how good are these models at like trading, investing, making money for you? 168 00:10:28,950 --> 00:10:32,590 Ejaaz: Well, there's this very famous benchmark, which is called the Vending Benchmark, 169 00:10:32,590 --> 00:10:37,320 Ejaaz: which basically allows an AI model to control a theoretical $10,000 and see 170 00:10:37,320 --> 00:10:41,420 Ejaaz: if it can make money by stocking a vending machine and then conducting sales, 171 00:10:41,420 --> 00:10:43,850 Ejaaz: managing inventory against competition. 172 00:10:44,150 --> 00:10:48,210 Ejaaz: It achieved second place right behind Claude Opus 4.7, which is the current 173 00:10:48,210 --> 00:10:51,070 Ejaaz: leading model. So it's also pretty good at making money as well. 174 00:10:51,640 --> 00:10:55,980 Josh: Yeah, and it also has a very clear roadmap to continue to be good and to get 175 00:10:55,980 --> 00:11:00,430 Josh: even better. There's an interaction actually between Elon Musk and the CEO of 176 00:11:00,660 --> 00:11:02,810 Josh: Z.ai, who is creating these models. 177 00:11:02,970 --> 00:11:06,180 Josh: So this guy asked, what's your current timeline for China to reach Fableclass? 178 00:11:06,180 --> 00:11:10,140 Josh: GLM 5.2 certainly shortened the gap. And then Elon said probably Q1. 179 00:11:10,140 --> 00:11:14,680 Josh: And then the CEO said, won't take that long. Which means they expect us to get 180 00:11:15,070 --> 00:11:19,780 Josh: a new Fableclass level model that's open weight and open source within the next six months. 181 00:11:19,780 --> 00:11:23,920 Josh: Which is incredibly compelling because that is going to be served up as open weights. 182 00:11:23,920 --> 00:11:26,720 Josh: And as you know, with open weights, you can actually run it on your own hardware. 183 00:11:26,720 --> 00:11:29,510 Josh: But the question is, do you actually want to run this on your hardware? 184 00:11:29,510 --> 00:11:33,030 Josh: I see on Twitter all the time, people who are spending tens of thousands of 185 00:11:33,030 --> 00:11:36,030 Josh: dollars to get those Mac studios, they're stacking them up in their offices, 186 00:11:36,030 --> 00:11:38,190 Josh: they're trying really hard to run these models locally. 187 00:11:38,500 --> 00:11:42,540 Josh: And I hate to break it to you, but the math ain't really math in on this so well. 188 00:11:42,850 --> 00:11:46,500 Josh: So there's a suite by Mike Schweinbach I thought was great. And it says the 189 00:11:46,500 --> 00:11:53,540 Josh: minimum to run the model is about $20,000 in hardware and you get about 20 tokens per second out. 190 00:11:53,860 --> 00:11:56,010 Ejaaz: For $20,000, that's like, 191 00:11:56,010 --> 00:11:59,490 Josh: That's pretty slow. It's not thinking that fast. And if you have these really 192 00:11:59,490 --> 00:12:02,590 Josh: long chain of thoughts, these long reasoning traces, it's going to take you 193 00:12:02,590 --> 00:12:05,080 Josh: a very long time to get an answer that involves deep thinking. 194 00:12:05,430 --> 00:12:09,810 Josh: So for about $20,000, you can get close to 35 billion tokens. 195 00:12:10,180 --> 00:12:14,070 Josh: And that's a 12 to one input to output ratio, assuming you have like good token caching setup. 196 00:12:14,390 --> 00:12:18,420 Josh: So he's saying if you ran the hardware 24-7 with zero downtime, 197 00:12:18,420 --> 00:12:22,030 Josh: it would take roughly five and a half years just to break even. 198 00:12:22,440 --> 00:12:26,030 Josh: And that right there is why open weights models are incredible. 199 00:12:26,730 --> 00:12:30,700 Josh: You're probably better off getting it served directly from their servers from 200 00:12:30,700 --> 00:12:32,310 Josh: the cloud instead of running your own. 201 00:12:32,310 --> 00:12:34,480 Josh: Because not only do you have to deal with the complexity, you have to power 202 00:12:34,480 --> 00:12:37,320 Josh: it all on, you have to deal with hardware stuff, and you have to worry about 203 00:12:37,320 --> 00:12:38,290 Josh: getting the actual hardware. 204 00:12:38,290 --> 00:12:41,750 Josh: Because Lord knows, getting those computers now is not as easy as it used to 205 00:12:41,750 --> 00:12:43,900 Josh: be. So interesting note on cost, 206 00:12:43,900 --> 00:12:47,450 Josh: on how available these are and accessible these are on a relative basis. 207 00:12:47,910 --> 00:12:52,020 Ejaaz: And the Chinese companies themselves are willing to subsidize these costs, just to be clear. 208 00:12:52,020 --> 00:12:56,090 Ejaaz: Like to play around with Kimi K 2.7, which is their frontier model, 209 00:12:56,090 --> 00:12:59,250 Ejaaz: I've been able to access it and use it since they launched it. 210 00:12:59,250 --> 00:13:02,210 Ejaaz: And I've been free using it to kind of like do research and all that kind of 211 00:13:02,210 --> 00:13:06,030 Ejaaz: stuff. And I've never once been charged for it. So there's a high subsidy coming 212 00:13:06,030 --> 00:13:07,930 Ejaaz: from like the Chinese side of things as well. 213 00:13:08,290 --> 00:13:11,040 Ejaaz: The other thing I'll say is these numbers may look big, right? 214 00:13:11,040 --> 00:13:14,630 Ejaaz: Like who on earth is spending $20,000 to get hardware that you can like run 215 00:13:14,630 --> 00:13:17,340 Ejaaz: at home to run these models open source? 216 00:13:17,680 --> 00:13:22,360 Ejaaz: But the idea is six months from now, 12 months from now, these very same models 217 00:13:22,360 --> 00:13:24,030 Ejaaz: will be distilled enough. 218 00:13:24,030 --> 00:13:27,600 Ejaaz: So that means it can maintain its intelligence, but good enough to run on your 219 00:13:27,600 --> 00:13:30,750 Ejaaz: local hardware at home, a custom PC, or maybe even your laptop. 220 00:13:30,750 --> 00:13:34,840 Ejaaz: The trend that we're undeniably seeing with these open-source models in particular 221 00:13:35,090 --> 00:13:38,450 Ejaaz: is higher intelligence for lower-cost hardware. 222 00:13:38,640 --> 00:13:42,050 Ejaaz: And if that trend continues, we will end up seeing this model that we're talking 223 00:13:42,050 --> 00:13:45,830 Ejaaz: about today being able to run off your handset. So it's something that seems 224 00:13:45,830 --> 00:13:47,460 Ejaaz: unfeasible right now to access. 225 00:13:47,750 --> 00:13:51,860 Ejaaz: But further on down the line, open-source, in my opinion, is pretty undeniable. 226 00:13:52,080 --> 00:13:55,960 Ejaaz: You'll be able to run it at home, and that's pretty good. But moving on. 227 00:13:56,650 --> 00:14:01,610 Ejaaz: The reason why we wanted to write this episode is there's a convergence of two trends, right? 228 00:14:01,760 --> 00:14:07,570 Ejaaz: So last week, we had a lot of reporting around Fable 5 being banned by the United States government. 229 00:14:07,820 --> 00:14:11,220 Ejaaz: The primary reason is the United States government does not think the model 230 00:14:11,220 --> 00:14:16,750 Ejaaz: is safe. If placed in a malicious actor's hands, we'll be able to be used against 231 00:14:16,870 --> 00:14:20,380 Ejaaz: government systems, hack, exploits, all that kind of stuff. And it's proven 232 00:14:20,380 --> 00:14:22,210 Ejaaz: itself on internal testing. 233 00:14:22,490 --> 00:14:28,370 Ejaaz: And the most recent revealing was a quote from a senator saying that the head of the NSA 234 00:14:29,070 --> 00:14:32,840 Ejaaz: Explained in a red team exercise, which is like a controlled environment, 235 00:14:32,840 --> 00:14:37,790 Ejaaz: that Claude Mythos 5 was able to breach all of its systems. 236 00:14:38,110 --> 00:14:42,970 Ejaaz: And typically, it would take months for an individual expert to do that. 237 00:14:43,160 --> 00:14:47,900 Ejaaz: It did it in hours. And this is just a crazy story and headline to read. 238 00:14:48,280 --> 00:14:51,860 Ejaaz: They've switched it off. It's not accessible to anyone. If you go on cloud right 239 00:14:51,860 --> 00:14:53,750 Ejaaz: now, you're unable to access Fable 5. 240 00:14:54,000 --> 00:14:57,230 Ejaaz: But the point is, these two trends have converged at the same time. 241 00:14:57,230 --> 00:15:00,620 Ejaaz: And it's important to discuss this because very soon in a few months time, 242 00:15:00,620 --> 00:15:04,300 Ejaaz: as that Elon tweet showed, we're going to end up with Mythos grade level models 243 00:15:04,300 --> 00:15:08,970 Ejaaz: that are freely available to anyone, subsidized by China or available to run at home for 10k. 244 00:15:09,360 --> 00:15:12,550 Ejaaz: And that is pretty scary, I guess. 245 00:15:12,830 --> 00:15:16,270 Josh: Yeah. Is that the lead now? Are we at six months? Does that feel about right? 246 00:15:16,270 --> 00:15:19,620 Josh: Like if they, if they release Mythos class by the end of this year, 247 00:15:20,020 --> 00:15:24,150 Josh: and then that gives, I guess, an open AI and Anthropic a six month head start. 248 00:15:24,490 --> 00:15:26,360 Ejaaz: And then the head of Chippoo has said it. 249 00:15:26,360 --> 00:15:30,230 Josh: So, yeah. So it seems like that's about right currently where we have like a 250 00:15:30,230 --> 00:15:34,220 Josh: six month window between us and the current bleeding edge open source. 251 00:15:34,840 --> 00:15:40,160 Josh: I could see that kind of getting closer and closer. It feels like they're right on the tail. 252 00:15:40,400 --> 00:15:43,370 Josh: Of course, understanding what's going on internally would be very helpful to 253 00:15:43,370 --> 00:15:48,000 Josh: know, because I'm sure GPT 5.5, well, we know we're getting 5.6 pretty soon. 254 00:15:48,340 --> 00:15:51,000 Josh: I'm sure Anthropic is working on something even more powerful than Mythos. 255 00:15:51,120 --> 00:15:54,100 Josh: And it feels like we don't really have a choice but to continue progressing 256 00:15:54,100 --> 00:15:56,520 Josh: as fast as we are. Otherwise, these are going to catch up. 257 00:15:56,860 --> 00:16:00,520 Josh: And they won't have the guardrails that are put in place currently by the Frontier 258 00:16:00,520 --> 00:16:03,730 Josh: models. Now, what's happening currently is we're seeing this fork. 259 00:16:04,030 --> 00:16:07,430 Josh: In terms of these private models where only people internally are now able to 260 00:16:07,430 --> 00:16:11,670 Josh: use them and anyone out in the world is getting, I guess, kind of disabled. 261 00:16:11,670 --> 00:16:14,750 Josh: They're getting a handicap because they're not actually able to access these frontier models. 262 00:16:15,040 --> 00:16:18,750 Josh: So we're seeing this weird crossroads where there's a small subset of people 263 00:16:18,750 --> 00:16:22,120 Josh: that work internally within OpenAI, within Anthropic, that are getting access to these models. 264 00:16:22,350 --> 00:16:25,750 Josh: The government is limiting their public use, which means the public is getting left behind. 265 00:16:25,980 --> 00:16:28,800 Josh: And then China is coming up and they're saying, hey, in six months, 266 00:16:28,800 --> 00:16:29,790 Josh: we're going to be right here at your head. 267 00:16:29,990 --> 00:16:33,890 Josh: So it's this really interesting dynamic that's at play. And we're going to really 268 00:16:33,890 --> 00:16:37,280 Josh: have to closely monitor this as these new frontier models continue to be released, 269 00:16:37,280 --> 00:16:42,080 Josh: because you have to assume, even though the world isn't using Mythos or Fable, they're continuing 270 00:16:42,500 --> 00:16:45,800 Josh: to iterate and to build better models. They're not just going to stop because of this. 271 00:16:46,010 --> 00:16:48,060 Josh: Same with OpenAI, same with all the other frontier labs. 272 00:16:48,630 --> 00:16:50,240 Josh: The question is, are these models 273 00:16:50,240 --> 00:16:53,300 Josh: going to be held privately for just a small subset of people to use? 274 00:16:53,500 --> 00:16:56,420 Josh: Or is there going to be this path forward in which the public can use them? 275 00:16:56,630 --> 00:16:58,600 Josh: I think everyone's hope is that there is a path forward. 276 00:16:58,900 --> 00:17:02,540 Josh: But currently, we're at this weird standstill where it feels like China's kind 277 00:17:02,540 --> 00:17:03,600 Josh: of breathing down your neck here. 278 00:17:03,790 --> 00:17:08,290 Ejaaz: Well, the irony also is if the government is just going to come in and switch 279 00:17:08,290 --> 00:17:13,630 Ejaaz: off the frontier model, it's going to push companies to use open source models. 280 00:17:14,030 --> 00:17:17,030 Ejaaz: Imagine you're an enterprise, right? And you're running your entire company 281 00:17:17,330 --> 00:17:20,740 Ejaaz: on Fable 5 or whatever the frontier model is from an AI lab. 282 00:17:21,090 --> 00:17:25,240 Ejaaz: And then suddenly you know that the government can just switch the button off 283 00:17:25,460 --> 00:17:28,340 Ejaaz: and suddenly your company can't do its thing. 284 00:17:29,210 --> 00:17:32,760 Ejaaz: You're more incentivized to kind of like run an open model at home that's privately 285 00:17:32,760 --> 00:17:35,650 Ejaaz: inferenced such that you can never shut it down. 286 00:17:35,650 --> 00:17:39,630 Ejaaz: So if I was an enterprise that has been running Fable 5 and that has now been 287 00:17:39,630 --> 00:17:42,620 Ejaaz: shut off, I'll be looking over at this GLM 5.2 thing and thinking, 288 00:17:43,050 --> 00:17:44,510 Ejaaz: well, it's MIT open source. 289 00:17:45,250 --> 00:17:48,760 Ejaaz: Yeah, maybe it costs 20K to run on hardware, but like I'll rather spend that 290 00:17:48,890 --> 00:17:53,710 Ejaaz: and save, you know, hundreds of millions down the line versus like going with Fable 5. 291 00:17:54,310 --> 00:17:57,040 Ejaaz: And yeah, maybe achieving frontier level performance, but then, 292 00:17:57,290 --> 00:18:00,450 Ejaaz: you know, being shut off potentially by the government, according to their agenda, 293 00:18:00,450 --> 00:18:02,500 Ejaaz: like that's not something that you potentially want. 294 00:18:02,500 --> 00:18:08,250 Ejaaz: Now, I want to give a quick counterpoint to the whole Chinese open source AI 295 00:18:08,250 --> 00:18:10,440 Ejaaz: models are going to take over the world because they're cheaper, 296 00:18:10,700 --> 00:18:13,940 Ejaaz: they're as good, maybe not as good, but as good, good enough, 297 00:18:13,940 --> 00:18:16,180 Ejaaz: right? Which is very simple. 298 00:18:16,670 --> 00:18:21,160 Ejaaz: If you're an American lab that has a frontier AI model that is expensive and 299 00:18:21,160 --> 00:18:24,240 Ejaaz: you see your neighbors, or if you see your adversaries, China, 300 00:18:24,440 --> 00:18:29,610 Ejaaz: distilling your model and presenting it as a cheaper model, you just do the same for your own model. 301 00:18:29,930 --> 00:18:34,220 Ejaaz: And Anthropic has demonstrated that many times, producing Sonnet. 302 00:18:34,220 --> 00:18:38,560 Ejaaz: Sonnet 4 is basically their cheaper model of Opus 4.8, I believe. 303 00:18:38,850 --> 00:18:43,780 Ejaaz: And then you see it with ChatGPT, with GPT Flash. These AI labs will produce 304 00:18:43,780 --> 00:18:47,100 Ejaaz: a cheaper version, and they'll distill it directly from their frontier models. 305 00:18:47,350 --> 00:18:51,440 Ejaaz: And as these models get good enough to rebuild themselves, it gets easier to do. 306 00:18:51,650 --> 00:18:57,020 Ejaaz: So I can see a world where they release Fable 6 in the future with a companion 307 00:18:57,020 --> 00:19:01,850 Ejaaz: model, which is like Sonic 6. And it's super cheap for anyone that wants 85% 308 00:19:01,850 --> 00:19:05,310 Ejaaz: of the capability and don't care about that extra 15%. And it's super cheap. 309 00:19:05,470 --> 00:19:08,410 Ejaaz: So it's competitive with the Chinese models. I don't think America has lost 310 00:19:08,670 --> 00:19:12,390 Ejaaz: the kind of like cheap model argument, but the open source one, 311 00:19:12,390 --> 00:19:15,820 Ejaaz: they definitely have. I don't see the American and labs open sourcing anytime soon. 312 00:19:16,020 --> 00:19:18,790 Josh: Yeah, well, we saw MetaPivot very clearly from the open source, 313 00:19:19,210 --> 00:19:22,690 Josh: but like the savior of the open source world to closed source very quickly. 314 00:19:22,690 --> 00:19:26,040 Josh: And I mean, that hasn't worked out too well for them or anyone really, 315 00:19:26,260 --> 00:19:26,990 Josh: which is disappointing. 316 00:19:27,180 --> 00:19:31,060 Josh: There is a small caveat. Maybe we should cover about what open source actually 317 00:19:31,060 --> 00:19:34,980 Josh: means because it's not truly open source. There are still some secrets. 318 00:19:35,510 --> 00:19:38,880 Josh: I think a better way to classify this is open weights. And when you go through 319 00:19:38,880 --> 00:19:42,300 Josh: training, there's, let's say, a trillion parameters. Each one of those parameters 320 00:19:42,300 --> 00:19:44,750 Josh: gets tuned over and over and over through each training run, 321 00:19:45,060 --> 00:19:46,730 Josh: which happens trillions of times. 322 00:19:46,990 --> 00:19:51,450 Josh: And the output of this are the weights. It's just a large text file that has 323 00:19:51,450 --> 00:19:54,500 Josh: all of those parameters finely tuned that the model can run off of. 324 00:19:54,800 --> 00:19:59,480 Josh: What it doesn't include is the actual source code that it took to make that. 325 00:19:59,480 --> 00:20:04,990 Josh: It doesn't include the ability to reproduce it. All it shares is the outputs. 326 00:20:05,190 --> 00:20:08,190 Josh: So while you could take their outputs and you could retune and fine-tune those 327 00:20:08,190 --> 00:20:09,760 Josh: parameters to give you exactly what you want 328 00:20:10,060 --> 00:20:13,430 Josh: it's not giving you the recipe it's not giving you the secrets on how it built it 329 00:20:13,720 --> 00:20:17,210 Josh: so there is still some proprietary knowledge as it relates to this open source 330 00:20:17,450 --> 00:20:20,810 Josh: model these chinese companies because they they are actually preserving the 331 00:20:20,810 --> 00:20:24,220 Josh: recipe in which they landed on this the data that they trained on there's a 332 00:20:24,220 --> 00:20:25,940 Josh: lot of secrets the output 333 00:20:26,270 --> 00:20:28,980 Josh: is what's open source and that's technically open weight so when we say open 334 00:20:28,980 --> 00:20:32,580 Josh: source i think what we really mean whenever you hear open source model chances 335 00:20:32,580 --> 00:20:35,370 Josh: are it's open weights and that's a pretty big distinction because that allows 336 00:20:35,370 --> 00:20:39,010 Josh: them to keep their kind of their secret sauce of how they do it and it's also 337 00:20:39,010 --> 00:20:40,330 Josh: probably for the better because i assume, 338 00:20:40,810 --> 00:20:44,240 Josh: you got to imagine they've been distilling some sort of stuff from i mean i 339 00:20:44,240 --> 00:20:48,850 Josh: remember seed dance that was so like obviously stolen material because it was 340 00:20:48,850 --> 00:20:54,120 Josh: just able to reproduce all the copyright and video formats from any public tv show in the world so. 341 00:20:54,700 --> 00:20:57,920 Josh: Where they get their data from leaves a lot to be desired and questioned, 342 00:20:57,920 --> 00:21:01,810 Josh: but that's kind of the nuance between open source and open weights. 343 00:21:01,810 --> 00:21:04,340 Josh: And what we're getting right now currently is open weights. 344 00:21:04,500 --> 00:21:09,910 Ejaaz: I don't necessarily believe it's open models versus centralized models. 345 00:21:10,310 --> 00:21:15,570 Ejaaz: I think it lands somewhere in between. Now, we've been noticing this new type 346 00:21:15,570 --> 00:21:20,050 Ejaaz: of product that is getting used by a lot of software engineers and AI users. 347 00:21:20,800 --> 00:21:26,160 Ejaaz: It's probably best demonstrated by this recent product release from Sakana AI. 348 00:21:26,160 --> 00:21:29,100 Ejaaz: It's called this new model called Fugu. 349 00:21:29,450 --> 00:21:33,050 Ejaaz: And they describe it as a multi-agent orchestration system. Basically how it 350 00:21:33,050 --> 00:21:38,460 Ejaaz: works is you send their model a prompt as you do with ChatGPT or Claude. 351 00:21:38,860 --> 00:21:43,980 Ejaaz: And it disperses that prompt across many different models. It could be closed 352 00:21:43,980 --> 00:21:45,730 Ejaaz: models like Claude and GPT. 353 00:21:46,020 --> 00:21:51,390 Ejaaz: Could be open models like GPO GLM or Kimi K2.7. as well as their own trained 354 00:21:51,390 --> 00:21:53,160 Ejaaz: model called Fugu, I believe. 355 00:21:53,540 --> 00:21:58,570 Ejaaz: And the result of this is like agentic debate. So these models kind of produce their own answers. 356 00:21:58,840 --> 00:22:02,220 Ejaaz: Then you have another model that kind of judges these answers and produces the 357 00:22:02,220 --> 00:22:03,690 Ejaaz: best answer from all of this. 358 00:22:04,000 --> 00:22:08,230 Ejaaz: And the result from these tests is basically, not only do you have a better 359 00:22:08,230 --> 00:22:10,710 Ejaaz: quality output, but it's also cheaper. 360 00:22:10,960 --> 00:22:15,430 Ejaaz: So the orchestration module basically picks the best models to do something 361 00:22:15,430 --> 00:22:18,920 Ejaaz: when it's like cheaper, and then only uses the best models when it really needs 362 00:22:18,920 --> 00:22:21,520 Ejaaz: to solve a really hard task that the other cheaper models can't do. 363 00:22:21,760 --> 00:22:25,010 Ejaaz: So it saves you a bunch of money, and we see it across other companies like 364 00:22:25,010 --> 00:22:28,600 Ejaaz: OpenRouter with their new Fusion API. The point being made here is, 365 00:22:29,080 --> 00:22:34,760 Ejaaz: We are headed towards a world where the ideal AI chatbot uses multiple models, 366 00:22:34,760 --> 00:22:36,850 Ejaaz: and they may not just be from the same company. 367 00:22:37,190 --> 00:22:40,720 Ejaaz: So the question I have for the United States government and any government that 368 00:22:40,720 --> 00:22:44,540 Ejaaz: decides to regulate, whether it's open source models or closed source models, 369 00:22:44,830 --> 00:22:46,960 Ejaaz: how are you going to regulate every single model in the world, 370 00:22:46,960 --> 00:22:50,730 Ejaaz: especially when the model labs come from other countries or are in fact open source? 371 00:22:50,990 --> 00:22:54,340 Ejaaz: You can't regulate open source models. That's the whole idea of it, 372 00:22:54,340 --> 00:22:55,570 Ejaaz: whether it's open weight or open source. 373 00:22:55,870 --> 00:22:59,210 Ejaaz: The whole idea is the government can't try to doubt if you're running it on hardware at home. 374 00:22:59,580 --> 00:23:02,770 Ejaaz: So it's just a really interesting nuance. I just don't think that the stance 375 00:23:02,770 --> 00:23:06,300 Ejaaz: that the United States government has taken so far is necessarily the most productive 376 00:23:06,300 --> 00:23:09,640 Ejaaz: one. I understand why they're doing it, but we need to figure out a different framework. 377 00:23:09,890 --> 00:23:15,670 Josh: It's funny because I saw this news this morning about this Sakana Fugu. 378 00:23:16,350 --> 00:23:19,670 Josh: I think I'm pronouncing that right. I mean, surely I've never heard of this. 379 00:23:19,670 --> 00:23:21,800 Josh: I don't know if you've ever heard of this. I think a lot of people watching 380 00:23:21,800 --> 00:23:24,070 Josh: have never heard of this company. They're Japanese. They came out of nowhere. 381 00:23:24,250 --> 00:23:28,310 Josh: And suddenly they're posting benchmarks that show that it has higher performance than Fable. 382 00:23:28,520 --> 00:23:31,420 Josh: And maybe that's true. Maybe they use this mixture of agents. 383 00:23:31,690 --> 00:23:36,070 Josh: But I think it's also notable that a lot of this is benchmarks. 384 00:23:36,070 --> 00:23:39,880 Josh: And I actually got some time to play around with the new GLM model this weekend. 385 00:23:40,150 --> 00:23:44,310 Josh: And while I'm sure it's great at coding and technical use, that's not really 386 00:23:44,310 --> 00:23:46,340 Josh: what I generally use the models for. 387 00:23:46,500 --> 00:23:50,290 Josh: And as I'm actually using these models, I'm giving it the general vibe test 388 00:23:50,460 --> 00:23:55,470 Josh: i'm noticing that i really do strongly bias the american closed source models like, 389 00:23:55,920 --> 00:24:01,100 Josh: uh gpt and like anthropics um opus and um claude and i mean fable when it was 390 00:24:01,100 --> 00:24:02,730 Josh: available was incredible and 391 00:24:03,070 --> 00:24:06,190 Josh: although the benchmarks show that it's very competent at coding a lot of people 392 00:24:06,190 --> 00:24:08,140 Josh: aren't using it for coding they're using it for other things and 393 00:24:08,450 --> 00:24:12,410 Josh: and the the general the vibe check doesn't get passed with these models yet 394 00:24:12,410 --> 00:24:15,960 Josh: at least um so i think that's something worth noting too is like these are just 395 00:24:15,960 --> 00:24:19,890 Josh: benchmarks i encourage anyone who's listening go try this out for yourself and see for yourself. 396 00:24:20,310 --> 00:24:23,500 Josh: Some people may actually get a lot of benefit from using a cheaper model. 397 00:24:23,500 --> 00:24:26,140 Josh: Some people just like having all the context in one place and they want just 398 00:24:26,140 --> 00:24:27,110 Josh: a better overall experience. 399 00:24:27,390 --> 00:24:30,590 Josh: With the routing, I think this is a super interesting precedent that we're seeing. 400 00:24:31,310 --> 00:24:36,310 Josh: Sakana fugu and how they are choosing to route their outputs through a series 401 00:24:36,310 --> 00:24:40,020 Josh: of open source and closed source models in order to generate a better and more 402 00:24:40,020 --> 00:24:44,310 Josh: powerful outcome i wonder the costs i noticed that as i was looking through the documentation 403 00:24:44,630 --> 00:24:47,100 Josh: there was no real cost associated i have to assume it's, 404 00:24:47,680 --> 00:24:50,690 Josh: not as high but pretty close because it is routing through. 405 00:24:51,410 --> 00:24:54,870 Josh: A lot of the private models and some open source models in order to get this 406 00:24:54,870 --> 00:24:58,310 Josh: which means it's probably consuming a good bit of tokens it's not totally going 407 00:24:58,310 --> 00:25:00,270 Josh: to be this like open source very low price model 408 00:25:00,560 --> 00:25:03,660 Josh: but it is interesting to see this trend towards more router based applications 409 00:25:03,660 --> 00:25:06,430 Josh: where not everyone needs to solve this incredibly difficult challenge. 410 00:25:06,670 --> 00:25:09,490 Josh: Perhaps you spin off a few sub agents, they use a more lightweight model to 411 00:25:09,490 --> 00:25:12,840 Josh: get you an answer without needing to consume a lot of those higher cost tokens. 412 00:25:12,840 --> 00:25:17,110 Josh: So it's cool, innovative, I won't say it's novel, we've seen this before, 413 00:25:17,110 --> 00:25:20,740 Josh: but it's a new iteration of this that is now showing pretty compelling benchmarks. 414 00:25:22,090 --> 00:25:26,950 Ejaaz: On the cost side of things, if it's anything like OpenRouter's Fusion API, 415 00:25:26,950 --> 00:25:31,530 Ejaaz: which does the same architecture, it achieves roughly like 30 to 50% cheaper 416 00:25:31,870 --> 00:25:35,720 Ejaaz: versus the frontier models, which isn't that major compared to like some of 417 00:25:35,720 --> 00:25:36,770 Ejaaz: the Chinese open source models. 418 00:25:36,770 --> 00:25:40,310 Ejaaz: But it still saves you a bunch of money if you're an enterprise using this at length. 419 00:25:41,550 --> 00:25:46,160 Ejaaz: I'm trying to think about the major takeaway that I have for myself after we've 420 00:25:46,160 --> 00:25:47,220 Ejaaz: done this episode, Josh. 421 00:25:47,220 --> 00:25:51,090 Ejaaz: And I think the main one is I'm inclined to say, and I hope I'm wrong, 422 00:25:51,330 --> 00:25:56,370 Ejaaz: that future AI model releases, Fable and above, whether it comes from GPT 5.6 423 00:25:56,370 --> 00:25:58,900 Ejaaz: or 6 or other frontier AI labs, 424 00:25:59,360 --> 00:26:02,520 Ejaaz: they're going to be more controlled in their release because governments are 425 00:26:02,520 --> 00:26:04,070 Ejaaz: going to start getting more involved. 426 00:26:04,070 --> 00:26:09,170 Ejaaz: We're going to start seeing nationalization attempts from different nation states 427 00:26:09,170 --> 00:26:13,000 Ejaaz: in order to figure out how to release these AR models because if they're out 428 00:26:13,000 --> 00:26:15,670 Ejaaz: in the wild, they can exploit and cause some real damage. 429 00:26:16,390 --> 00:26:21,740 Ejaaz: I don't want to think about what could happen in terms of a major event, 430 00:26:21,740 --> 00:26:24,230 Ejaaz: but I think we're reaching that point where we need to pay careful attention. 431 00:26:24,690 --> 00:26:27,210 Ejaaz: So that's what we're trying to do on this episode. At least that's what I'm trying to do. 432 00:26:27,520 --> 00:26:30,570 Josh: Yeah, I think that's right. Like the speed and acceleration of these models 433 00:26:30,570 --> 00:26:33,730 Josh: and the cadence in which they're released is up only. 434 00:26:33,930 --> 00:26:37,810 Josh: If we had a chart that showed you the length of time in between major model 435 00:26:37,810 --> 00:26:39,940 Josh: releases, It is just getting shorter and shorter and shorter, 436 00:26:39,940 --> 00:26:40,750 Josh: and that's not changing. 437 00:26:41,100 --> 00:26:44,430 Josh: So there needs to be a way to reliably be able to push these out. 438 00:26:44,430 --> 00:26:48,110 Josh: Otherwise, the gap between what exists behind closed doors and what's available 439 00:26:48,110 --> 00:26:49,580 Josh: to the public is just going to keep growing. 440 00:26:50,000 --> 00:26:55,820 Josh: And I'm not sure what implications that has, but it sounds like it is noteworthy and something... 441 00:26:56,360 --> 00:26:59,880 Josh: Something needs to change in a material way because the speed and velocity in 442 00:26:59,880 --> 00:27:02,230 Josh: which progress is being made is not slowing down. 443 00:27:02,580 --> 00:27:05,340 Josh: Like, what does this look like a year from now? How quick are these models able 444 00:27:05,340 --> 00:27:08,540 Josh: to improve themselves? What are the benchmarks look like? Can we even create 445 00:27:08,540 --> 00:27:10,450 Josh: benchmarks anymore because it will be so capable? 446 00:27:10,800 --> 00:27:14,760 Josh: We're right on that cusp because we are approaching this vertical asymptote off the curve. 447 00:27:15,180 --> 00:27:18,060 Josh: And it's just like, it's a little weird. It feels like we're on this roller 448 00:27:18,060 --> 00:27:21,680 Josh: coaster and we're like kind of going down, but I guess it's inverted where we're 449 00:27:21,680 --> 00:27:24,360 Josh: going up and we're going up really fast and you're not really sure. It's escaping. 450 00:27:24,590 --> 00:27:27,930 Josh: It's escaping control in a way well i wouldn't say escaping control but it's 451 00:27:27,930 --> 00:27:31,070 Josh: just like that it's it's definitely getting fast and it's like okay like if 452 00:27:31,070 --> 00:27:33,370 Josh: you're driving your car really fast you got to be a little more careful once 453 00:27:33,370 --> 00:27:37,420 Josh: you reach high speed because like things things can kind of get a little shaky quickly so. 454 00:27:38,260 --> 00:27:41,610 Josh: We're at that point and models are getting very capable very quickly. 455 00:27:41,790 --> 00:27:46,270 Josh: I can't imagine what OpenAI's mythos class model looks like. 456 00:27:46,270 --> 00:27:47,320 Josh: I'm sure they're working on them. 457 00:27:47,570 --> 00:27:51,930 Josh: We talk about, I mean, the hardware. I always think about these are the Blackwell series models. 458 00:27:51,930 --> 00:27:55,290 Josh: What happens with the Vera Rubin series models? It's like this, 459 00:27:55,880 --> 00:28:00,040 Josh: we are going to accelerate so fast. And I think it's important to, 460 00:28:00,040 --> 00:28:04,030 Josh: yeah, work on these safeguards now where it's still reasonable to catch up, 461 00:28:04,030 --> 00:28:06,450 Josh: where there's only one model release in which you have to focus on. 462 00:28:06,670 --> 00:28:09,410 Josh: And there's not 10 different ones from all these different companies that are 463 00:28:09,660 --> 00:28:13,440 Josh: being pushed every single week so interesting that's the update china is, 464 00:28:13,900 --> 00:28:18,900 Josh: back with their open weights model not to be confused with open source and um 465 00:28:18,900 --> 00:28:20,840 Josh: yeah we still don't have fable access so 466 00:28:21,270 --> 00:28:24,300 Josh: hopefully these things will get sorted but i think it's it's noteworthy that 467 00:28:24,300 --> 00:28:27,510 Josh: china they they never disappeared i want to know what deep seek is doing next 468 00:28:27,510 --> 00:28:31,710 Josh: i think that's my next question is like where's deep seek at where's deep seek v v5 or v6 they just. 469 00:28:31,710 --> 00:28:36,030 Ejaaz: Raised a massive round 50 billion dollars um that's their valuation at least 470 00:28:36,030 --> 00:28:39,840 Ejaaz: there's still a fraction of frontier labs but yeah they raised like uh was it 471 00:28:39,840 --> 00:28:42,690 Ejaaz: nine billion dollars the founder himself put in three billion dollars there 472 00:28:43,120 --> 00:28:46,180 Ejaaz: they're doing pretty well and we haven't seen a model race from them anytime soon 473 00:28:46,180 --> 00:28:50,230 Josh: Yeah yeah so will be fun to see but that is the update on china on open source 474 00:28:50,230 --> 00:28:53,500 Josh: thank you guys so much for watching as always if you enjoyed this episode don't 475 00:28:53,500 --> 00:28:57,630 Josh: forget to share it with a friend who might also like the show who might care 476 00:28:57,630 --> 00:28:59,780 Josh: about china or open source models or wherever it may be 477 00:29:00,120 --> 00:29:04,880 Josh: if you listen on a podcast player rating us how you believe we deserve to be 478 00:29:04,880 --> 00:29:06,180 Josh: rated is always appreciated we 479 00:29:06,540 --> 00:29:10,700 Josh: love the five stars those are always great uh newsletter twice a week next one 480 00:29:10,700 --> 00:29:15,820 Josh: is dropping on wednesday a day after you listen to this and yeah that's i have one final. 481 00:29:15,820 --> 00:29:22,370 Ejaaz: Request josh something that you and i discussed on our on our walk uh last week but um 482 00:29:22,800 --> 00:29:28,210 Ejaaz: We are in the market for sponsors or anyone that can support us, please. 483 00:29:28,560 --> 00:29:33,370 Ejaaz: Josh and I and producer Luke have been keeping the lights on this entire time 484 00:29:33,370 --> 00:29:36,660 Ejaaz: and we've reached a point where we're feeling really confident about the numbers 485 00:29:36,660 --> 00:29:38,250 Ejaaz: and all the support that you guys have given us. 486 00:29:38,550 --> 00:29:42,640 Ejaaz: And we would love to have a partner that we feel very passionate about join 487 00:29:42,640 --> 00:29:47,340 Ejaaz: us and support us in our vision of growing this into the leading frontier and 488 00:29:47,340 --> 00:29:49,060 Ejaaz: AI tech podcast in the world. 489 00:29:49,060 --> 00:29:53,580 Ejaaz: So if there's anyone out there listening to this that is inspired or wants to 490 00:29:53,580 --> 00:29:57,600 Ejaaz: support us, let us know, DM us, you know, we're on X, we're everywhere, 491 00:29:58,000 --> 00:29:59,780 Ejaaz: just reach out and we would love to hear from you. 492 00:30:00,410 --> 00:30:02,800 Josh: That would be great. All the support is very much appreciated. 493 00:30:02,800 --> 00:30:05,030 Josh: Keep the lights on around here and keep things going strong. 494 00:30:05,030 --> 00:30:07,990 Josh: So yeah, thank you as always for the support. If you made it this long, 495 00:30:07,990 --> 00:30:11,460 Josh: you're a real one and hopefully you enjoyed this episode. So thank you as always 496 00:30:11,460 --> 00:30:12,920 Josh: and we will see you on the next one.