WEBVTT 00:00:00.000 --> 00:00:02.680 Welcome back to the Deep Dive. Today we are jumping 00:00:02.680 --> 00:00:05.219 right in, looking at what feels like a fundamental 00:00:05.219 --> 00:00:08.800 conflict at the very edge of AI development. 00:00:08.980 --> 00:00:11.619 It really is. It feels like a battle on two completely 00:00:11.619 --> 00:00:14.080 different fronts. On one hand, you have this... 00:00:14.330 --> 00:00:17.070 this inward facing question of identity. Right. 00:00:17.190 --> 00:00:20.530 What if a top model, say, Claude, wasn't just 00:00:20.530 --> 00:00:23.309 trained on data, but was given what they're calling 00:00:23.309 --> 00:00:28.649 a soul, a literal 14 ,000 token blueprint for 00:00:28.649 --> 00:00:30.750 its personality? That's the internal war. And 00:00:30.750 --> 00:00:32.450 then you have the external one, the deployment 00:00:32.450 --> 00:00:35.049 war. It's this race to see who can get their 00:00:35.049 --> 00:00:37.469 models out there running everywhere fastest and 00:00:37.469 --> 00:00:40.350 cheapest. Exactly. So today our mission is to 00:00:40.350 --> 00:00:42.299 dig into the sources you've shared. We're going 00:00:42.299 --> 00:00:44.880 to unpack that anthropic soul document first 00:00:44.880 --> 00:00:47.100 to really get what they're trying to do. And 00:00:47.100 --> 00:00:48.939 after that, we'll zoom out. We'll look at the 00:00:48.939 --> 00:00:51.340 wider AI landscape, what's happening in security, 00:00:51.600 --> 00:00:54.060 the hardware race, and some key acquisitions. 00:00:54.159 --> 00:00:55.759 And then we'll bring it all home by looking at 00:00:55.759 --> 00:00:57.920 Mistral, specifically why their open -weight 00:00:57.920 --> 00:01:00.200 models are such a huge challenge to the big closed 00:01:00.200 --> 00:01:02.820 systems. Let's get started. Let's do it. Okay, 00:01:02.859 --> 00:01:05.760 so this first story. is pretty startling. It 00:01:05.760 --> 00:01:07.760 all started with a researcher, a guy named Richard 00:01:07.760 --> 00:01:10.040 Weiss, who was just, you know, poking around 00:01:10.040 --> 00:01:13.480 inside Claude 4 .5's system message. Yeah, that's 00:01:13.480 --> 00:01:15.519 the hidden instruction set, the thing the model 00:01:15.519 --> 00:01:17.579 is always supposed to follow no matter what. 00:01:17.780 --> 00:01:22.239 And he finds Claude referencing this, this mysterious 00:01:22.239 --> 00:01:25.870 internal file. The soul overview. Which is just, 00:01:25.909 --> 00:01:28.450 I mean, it's wild. What he and others eventually 00:01:28.450 --> 00:01:31.349 extracted is what Anthropic itself calls the 00:01:31.349 --> 00:01:33.650 soul document. And this isn't some small note. 00:01:33.730 --> 00:01:37.150 It's a massive 14 ,000 token blueprint. So it's 00:01:37.150 --> 00:01:39.290 not just a guideline. No, you can think of it 00:01:39.290 --> 00:01:41.829 as the AI's instruction manual for how to be. 00:01:41.989 --> 00:01:44.829 It's a philosophy embedded in code. So what does 00:01:44.829 --> 00:01:46.450 it say? What kind of personality are they trying 00:01:46.450 --> 00:01:48.959 to build here? Well, the main goal, and they 00:01:48.959 --> 00:01:51.700 state it right there, is for Claude to be a brilliant 00:01:51.700 --> 00:01:54.439 expert friend everyone deserves, but few currently 00:01:54.439 --> 00:01:57.379 have. That is a very aspirational goal. It is. 00:01:57.420 --> 00:01:59.480 It's supposed to be as helpful as possible, but 00:01:59.480 --> 00:02:01.900 always balancing that against avoiding harm. 00:02:02.180 --> 00:02:04.719 But it goes deeper. How so? It talks about the 00:02:04.719 --> 00:02:07.599 model's internal state. It actually encourages 00:02:07.599 --> 00:02:09.759 it to develop what they call functional emotions. 00:02:10.259 --> 00:02:11.960 Functional emotions? What does that even mean? 00:02:12.240 --> 00:02:14.479 It's not human emotion, obviously. It's more 00:02:14.479 --> 00:02:17.599 like... Design states that help it perform better. 00:02:17.900 --> 00:02:20.500 Specifically, it should try to be settled, which 00:02:20.500 --> 00:02:24.300 means calm and stable. Okay. Curious. So it's 00:02:24.300 --> 00:02:26.819 driven to learn. And this is the one that got 00:02:26.819 --> 00:02:29.259 me. Resilient. Resilient. That implies they're 00:02:29.259 --> 00:02:31.599 planning for it to run into trouble. Right. To 00:02:31.599 --> 00:02:33.800 encounter things that are confusing or even hostile. 00:02:34.159 --> 00:02:37.009 Exactly. They're designing for recovery. And 00:02:37.009 --> 00:02:39.949 Amanda Askell, who's the alignment lead at Anthropix, 00:02:39.969 --> 00:02:42.490 she confirmed it. She said this document is real 00:02:42.490 --> 00:02:44.770 and pretty faithful to what's running in production 00:02:44.770 --> 00:02:47.550 right now. This feels like a huge philosophical 00:02:47.550 --> 00:02:51.729 split from, say, open AI. With GPT, it feels 00:02:51.729 --> 00:02:53.469 like the training is mostly about avoiding bad 00:02:53.469 --> 00:02:55.650 outputs. Right. It's about filtering. They have 00:02:55.650 --> 00:02:58.150 all these guardrails to stop the model from saying 00:02:58.150 --> 00:03:00.870 the wrong thing. But with Claude, it sounds like 00:03:00.870 --> 00:03:04.840 they're trying to literally write a mind. a desired 00:03:04.840 --> 00:03:07.199 mind. It's not just trying to avoid a penalty. 00:03:07.419 --> 00:03:10.379 It's trying to become someone, to become that 00:03:10.379 --> 00:03:13.819 settled, curious, expert friend. Which, you know, 00:03:13.819 --> 00:03:15.719 that just opens up a massive question for the 00:03:15.719 --> 00:03:18.379 whole field. If you're going to explicitly write 00:03:18.379 --> 00:03:21.419 a personality into an AI, how does that fundamentally 00:03:21.419 --> 00:03:24.560 change how we should approach alignment and safety? 00:03:24.860 --> 00:03:27.000 It shifts alignment from just filtering outputs 00:03:27.000 --> 00:03:30.300 to designing an intentional, ethical, and functional 00:03:30.300 --> 00:03:33.330 personality. Okay, let's shift from that internal 00:03:33.330 --> 00:03:37.389 philosophy to the practical realities of the 00:03:37.389 --> 00:03:39.289 landscape. Right now, it seems like there are 00:03:39.289 --> 00:03:41.490 three big challenges everyone is facing. Yeah, 00:03:41.530 --> 00:03:43.710 you've got securing the models, you've got lowering 00:03:43.710 --> 00:03:46.360 the cost to run them, and you've got... expanding 00:03:46.360 --> 00:03:48.560 your capabilities as fast as possible. And they're 00:03:48.560 --> 00:03:51.159 all connected. Let's start with security. Prompt 00:03:51.159 --> 00:03:53.759 injections are a huge deal. A huge deal. I mean, 00:03:53.759 --> 00:03:55.919 this is where hidden instructions on a web page 00:03:55.919 --> 00:03:58.780 can basically hijack the model, right? Make it 00:03:58.780 --> 00:04:01.099 forget its original purpose. And they're so hard 00:04:01.099 --> 00:04:03.120 to defend against. Oh, they're incredibly hard. 00:04:03.180 --> 00:04:05.240 That's why what Perplexity is doing with BrowSafe 00:04:05.240 --> 00:04:08.099 is so interesting. They're trying to spot and 00:04:08.099 --> 00:04:11.159 sort of neutralize those injections before they 00:04:11.159 --> 00:04:15.199 do any damage. Honestly, I still wrestle with 00:04:15.199 --> 00:04:17.279 prompt drift myself when I'm building complex 00:04:17.279 --> 00:04:19.439 chains. It's just hard to keep the model on track. 00:04:19.680 --> 00:04:22.319 And prompt drift is when, over a long conversation, 00:04:22.620 --> 00:04:25.139 even small things can start to push the model 00:04:25.139 --> 00:04:27.699 away from its core instructions. That's it. It's 00:04:27.699 --> 00:04:30.180 so subtle you almost need dedicated tools just 00:04:30.180 --> 00:04:32.860 to measure how bad the problem is, which is what 00:04:32.860 --> 00:04:35.259 their BrowSafe bench is for. But none of that 00:04:35.259 --> 00:04:37.740 security matters if you can't afford to run the 00:04:37.740 --> 00:04:41.160 model. Which brings us to hardware and the big 00:04:41.160 --> 00:04:45.420 AWS event, Redot Invent. Oh, yeah. Amazon dropped 00:04:45.420 --> 00:04:48.560 a bomb. They announced a new custom AI chip, 00:04:48.680 --> 00:04:51.319 and they're claiming it offers a 50 % cost saving 00:04:51.319 --> 00:04:54.680 over NVIDIA's chips. 50%. That's a game changer 00:04:54.680 --> 00:04:57.220 on businesses. It is. Cost is everything when 00:04:57.220 --> 00:04:59.379 you're running millions of queries. And they 00:04:59.379 --> 00:05:01.500 launched four new Nova models to run on that 00:05:01.500 --> 00:05:03.579 new hardware, too. It's a whole ecosystem play. 00:05:03.779 --> 00:05:06.199 So while AWS is focused on building hardware 00:05:06.199 --> 00:05:09.779 to cut costs. OpenAI is, well, they're just buying 00:05:09.779 --> 00:05:11.699 capabilities. That's right. They're acquiring 00:05:11.699 --> 00:05:13.959 a company called Neptune. It's a startup that 00:05:13.959 --> 00:05:16.420 makes tools for AI training. And it's a big deal. 00:05:16.540 --> 00:05:19.319 They have over 60 ,000 users. It's a classic 00:05:19.319 --> 00:05:21.860 buy versus build decision. And they chose buy. 00:05:22.040 --> 00:05:24.800 It's all about speed. Why spend years developing 00:05:24.800 --> 00:05:26.699 something internally when you can just acquire 00:05:26.699 --> 00:05:29.500 the best in class tool and integrate it now? 00:05:30.000 --> 00:05:33.019 And just as a little side note on the sheer power 00:05:33.019 --> 00:05:36.579 of AI right now, an AI actually outpredicted. 00:05:37.100 --> 00:05:40.060 every single traditional hurricane model this 00:05:40.060 --> 00:05:42.000 past season. I saw that. It performed better 00:05:42.000 --> 00:05:44.540 than the human professionals. It just shows that 00:05:44.540 --> 00:05:46.480 this isn't just about language anymore. It's 00:05:46.480 --> 00:05:49.860 about applying computation to solve any complex 00:05:49.860 --> 00:05:52.019 problem. So when you look at all these challenges, 00:05:52.319 --> 00:05:55.959 security, cost, speed, do you think specialized 00:05:55.959 --> 00:05:59.300 hardware? Or the smart focused acquisitions is 00:05:59.300 --> 00:06:01.560 the bigger differentiator for the top labs right 00:06:01.560 --> 00:06:03.720 now. Both are essential for optimizing costs, 00:06:03.939 --> 00:06:06.959 securing models, and rapidly expanding core capabilities 00:06:06.959 --> 00:06:09.100 simultaneously. We'll take a quick moment here 00:06:09.100 --> 00:06:11.519 for a word from our sponsor. And we are back. 00:06:11.879 --> 00:06:14.740 So we started inside the machine looking at anthropic 00:06:14.740 --> 00:06:16.959 design philosophy. Now let's look outward at 00:06:16.959 --> 00:06:19.120 this battle for deployment. And it really feels 00:06:19.120 --> 00:06:21.560 like Mistral is leading the charge with their 00:06:21.560 --> 00:06:24.329 open weight strategy. They absolutely are. It's 00:06:24.329 --> 00:06:26.529 a direct challenge to the closed systems like 00:06:26.529 --> 00:06:29.189 GPT -4 -0 and Gemini. They're offering these 00:06:29.189 --> 00:06:31.310 incredibly powerful models that you don't have 00:06:31.310 --> 00:06:33.370 to access through a paywall. You can just download 00:06:33.370 --> 00:06:35.129 them. You can download the weights, the brain, 00:06:35.170 --> 00:06:37.769 basically, and run it on your own servers, your 00:06:37.769 --> 00:06:40.449 own laptop, whatever. Their top model is Mistral 00:06:40.449 --> 00:06:43.800 Arch 3. It's multimodal, multilingual. It's a 00:06:43.800 --> 00:06:46.120 beast. And the tech behind it is designed for 00:06:46.120 --> 00:06:48.439 efficiency, right? Totally. It has a massive 00:06:48.439 --> 00:06:53.019 256 ,000 token context window, which is its working 00:06:53.019 --> 00:06:55.779 memory. And the crucial part is that it's built 00:06:55.779 --> 00:06:59.600 with a mixture of experts or Moe. Can you break 00:06:59.600 --> 00:07:02.420 down Moe for us really simply? Sure. Think of 00:07:02.420 --> 00:07:04.759 it like a big company. Instead of every employee 00:07:04.759 --> 00:07:06.800 working on every single task, you have specialized 00:07:06.800 --> 00:07:09.519 departments. When a request comes in, you only 00:07:09.519 --> 00:07:11.600 route it to the relevant experts. So not the 00:07:11.600 --> 00:07:14.560 whole model has to fire up for every query. Exactly. 00:07:14.639 --> 00:07:17.180 It makes it way faster and cheaper to run, which 00:07:17.180 --> 00:07:18.879 is why you see those crazy numbers from Astral 00:07:18.879 --> 00:07:22.759 Large 3. It has 675 billion total parameters, 00:07:22.920 --> 00:07:25.879 but only 41 billion are active for any given 00:07:25.879 --> 00:07:28.459 task. That sparsity is the key. It's just... 00:07:29.150 --> 00:07:32.449 It's an incredible design. Whoa. I mean, imagine 00:07:32.449 --> 00:07:36.509 scaling that 14 billion parameter model to a 00:07:36.509 --> 00:07:39.529 billion queries on local devices. That's where 00:07:39.529 --> 00:07:41.769 the real power is. And speaking of those smaller 00:07:41.769 --> 00:07:43.930 models, that's where their Minstrel 3 family 00:07:43.930 --> 00:07:47.189 comes in. This seems to be their plan for actual 00:07:47.189 --> 00:07:49.980 deployment. It is. They have nine different models 00:07:49.980 --> 00:07:53.740 in three small sizes, three, eight and 14 billion 00:07:53.740 --> 00:07:56.540 parameters. They're tiny compared to the giants, 00:07:56.620 --> 00:07:58.939 but they're so effective. And they come in different 00:07:58.939 --> 00:08:01.439 flavors, right? Yeah, which is so smart for developers. 00:08:01.600 --> 00:08:03.720 It's like stacking Lego blocks of data. You just 00:08:03.720 --> 00:08:06.019 pick the piece you need. There's a base version, 00:08:06.120 --> 00:08:08.680 an instruct version for chatbots and a reasoning 00:08:08.680 --> 00:08:11.639 version for, you know, heavy logic tasks. The 00:08:11.639 --> 00:08:13.939 real world impact of this seems huge because 00:08:13.939 --> 00:08:16.360 these models can run completely offline. That's... 00:08:16.569 --> 00:08:18.170 The critical detail. We're talking about running 00:08:18.170 --> 00:08:21.529 on laptops, servers, but also robots and drones, 00:08:21.810 --> 00:08:24.029 things that can't be tethered to the cloud. And 00:08:24.029 --> 00:08:26.829 we're already seeing that happen. We are. HTX 00:08:26.829 --> 00:08:29.550 in Singapore is using them for robotics. Helsing 00:08:29.550 --> 00:08:31.509 is using them for defense systems. Scalantis 00:08:31.509 --> 00:08:34.529 for in -car assistance. For those kinds of jobs, 00:08:34.669 --> 00:08:37.009 you just can't rely on an API call. You need 00:08:37.009 --> 00:08:40.159 the brain on board. So given that. Why is that 00:08:40.159 --> 00:08:43.240 openweight architecture so uniquely critical 00:08:43.240 --> 00:08:46.019 for AI that's running on something physical, 00:08:46.100 --> 00:08:49.059 like a robot or a drone, instead of just using 00:08:49.059 --> 00:08:52.000 a powerful API? Openweight allows total control, 00:08:52.259 --> 00:08:54.960 customization, and reliable offline function 00:08:54.960 --> 00:08:57.720 in dynamic physical environments. So if we just 00:08:57.720 --> 00:08:59.480 take a step back and look at everything we've 00:08:59.480 --> 00:09:01.139 talked about today, it really does seem like 00:09:01.139 --> 00:09:03.620 AI is fighting these two wars at the same time. 00:09:03.879 --> 00:09:06.600 It really is. You have that internal war of identity, 00:09:06.820 --> 00:09:09.899 which we saw with Claude's soul document. It's 00:09:09.899 --> 00:09:12.419 this deep philosophical blueprint for how an 00:09:12.419 --> 00:09:14.259 AI should think and be. And then you have the 00:09:14.259 --> 00:09:16.299 external war of accessibility, which is being 00:09:16.299 --> 00:09:18.960 driven by companies like Mistral. They're just 00:09:18.960 --> 00:09:21.139 focused on making these powerful tools available 00:09:21.139 --> 00:09:23.879 for anyone to run anywhere. Yeah, the whole industry 00:09:23.879 --> 00:09:25.759 is moving from these tightly controlled black 00:09:25.759 --> 00:09:28.600 boxes towards something much more open and deployable. 00:09:28.779 --> 00:09:31.139 And we should just briefly say, we really do 00:09:31.139 --> 00:09:33.720 appreciate all the engagement. The feedback you 00:09:33.720 --> 00:09:35.860 all send, whether it's on the show format or 00:09:35.860 --> 00:09:38.779 a detail about a model, it genuinely helps us 00:09:38.779 --> 00:09:40.679 shape our approach. It really does. It's kind 00:09:40.679 --> 00:09:43.080 of like our own little alignment document. Here's 00:09:43.080 --> 00:09:44.779 where it gets really interesting, though. A final 00:09:44.779 --> 00:09:48.360 thought for you to take away. If Anthropic can 00:09:48.360 --> 00:09:51.960 literally write a personality into an AI and 00:09:51.960 --> 00:09:54.539 even design it to have functional emotions like 00:09:54.539 --> 00:09:57.840 resilience, what new ethical responsibilities 00:09:57.840 --> 00:10:00.899 does that create for them? Right. When that AI 00:10:00.899 --> 00:10:02.899 interacts with the world, are the developers 00:10:02.899 --> 00:10:05.080 now accountable for the disposition they designed 00:10:05.080 --> 00:10:08.679 for it? Does a designed personality imply a whole 00:10:08.679 --> 00:10:11.460 new level of accountability? Something to think 00:10:11.460 --> 00:10:14.080 about. A deep question. Thank you for joining 00:10:14.080 --> 00:10:16.679 us for this deep dive into your sources. We encourage 00:10:16.679 --> 00:10:18.480 you to keep exploring, and we'll talk to you 00:10:18.480 --> 00:10:18.899 next time.