Welcome back to the Deep Dive. Today we are jumping right in, looking at what feels like a fundamental conflict at the very edge of AI development. It really is. It feels like a battle on two completely different fronts. On one hand, you have this... this inward facing question of identity. Right. What if a top model, say, Claude, wasn't just trained on data, but was given what they're calling a soul, a literal 14 ,000 token blueprint for
its personality? That's the internal war. And then you have the external one, the deployment war. It's this race to see who can get their models out there running everywhere fastest and cheapest. Exactly. So today our mission is to dig into the sources you've shared. We're going to unpack that anthropic soul document first to really get what they're trying to do. And
after that, we'll zoom out. We'll look at the wider AI landscape, what's happening in security, the hardware race, and some key acquisitions. And then we'll bring it all home by looking at Mistral, specifically why their open -weight models are such a huge challenge to the big closed systems. Let's get started. Let's do it. Okay, so this first story. is pretty startling. It all started with a researcher, a guy named Richard Weiss, who was just, you know, poking around
inside Claude 4 .5's system message. Yeah, that's the hidden instruction set, the thing the model is always supposed to follow no matter what. And he finds Claude referencing this, this mysterious internal file. The soul overview. Which is just, I mean, it's wild. What he and others eventually extracted is what Anthropic itself calls the soul document. And this isn't some small note. It's a massive 14 ,000 token blueprint. So it's
not just a guideline. No, you can think of it as the AI's instruction manual for how to be. It's a philosophy embedded in code. So what does it say? What kind of personality are they trying to build here? Well, the main goal, and they state it right there, is for Claude to be a brilliant expert friend everyone deserves, but few currently have. That is a very aspirational goal. It is. It's supposed to be as helpful as possible, but always balancing that against avoiding harm.
But it goes deeper. How so? It talks about the model's internal state. It actually encourages it to develop what they call functional emotions. Functional emotions? What does that even mean? It's not human emotion, obviously. It's more like... Design states that help it perform better. Specifically, it should try to be settled, which means calm and stable. Okay. Curious. So it's driven to learn. And this is the one that got me. Resilient. Resilient. That implies they're
planning for it to run into trouble. Right. To encounter things that are confusing or even hostile. Exactly. They're designing for recovery. And Amanda Askell, who's the alignment lead at Anthropix, she confirmed it. She said this document is real and pretty faithful to what's running in production right now. This feels like a huge philosophical split from, say, open AI. With GPT, it feels like the training is mostly about avoiding bad
outputs. Right. It's about filtering. They have all these guardrails to stop the model from saying the wrong thing. But with Claude, it sounds like they're trying to literally write a mind. a desired mind. It's not just trying to avoid a penalty. It's trying to become someone, to become that settled, curious, expert friend. Which, you know, that just opens up a massive question for the
whole field. If you're going to explicitly write a personality into an AI, how does that fundamentally change how we should approach alignment and safety? It shifts alignment from just filtering outputs to designing an intentional, ethical, and functional personality. Okay, let's shift from that internal philosophy to the practical realities of the landscape. Right now, it seems like there are
three big challenges everyone is facing. Yeah, you've got securing the models, you've got lowering the cost to run them, and you've got... expanding your capabilities as fast as possible. And they're all connected. Let's start with security. Prompt injections are a huge deal. A huge deal. I mean, this is where hidden instructions on a web page can basically hijack the model, right? Make it forget its original purpose. And they're so hard to defend against. Oh, they're incredibly hard.
That's why what Perplexity is doing with BrowSafe is so interesting. They're trying to spot and sort of neutralize those injections before they do any damage. Honestly, I still wrestle with prompt drift myself when I'm building complex chains. It's just hard to keep the model on track. And prompt drift is when, over a long conversation, even small things can start to push the model
away from its core instructions. That's it. It's so subtle you almost need dedicated tools just to measure how bad the problem is, which is what their BrowSafe bench is for. But none of that security matters if you can't afford to run the model. Which brings us to hardware and the big AWS event, Redot Invent. Oh, yeah. Amazon dropped a bomb. They announced a new custom AI chip, and they're claiming it offers a 50 % cost saving over NVIDIA's chips. 50%. That's a game changer
on businesses. It is. Cost is everything when you're running millions of queries. And they launched four new Nova models to run on that new hardware, too. It's a whole ecosystem play. So while AWS is focused on building hardware to cut costs. OpenAI is, well, they're just buying capabilities. That's right. They're acquiring a company called Neptune. It's a startup that makes tools for AI training. And it's a big deal. They have over 60 ,000 users. It's a classic
buy versus build decision. And they chose buy. It's all about speed. Why spend years developing something internally when you can just acquire the best in class tool and integrate it now? And just as a little side note on the sheer power of AI right now, an AI actually outpredicted. every single traditional hurricane model this past season. I saw that. It performed better than the human professionals. It just shows that
this isn't just about language anymore. It's about applying computation to solve any complex problem. So when you look at all these challenges, security, cost, speed, do you think specialized hardware? Or the smart focused acquisitions is the bigger differentiator for the top labs right now. Both are essential for optimizing costs, securing models, and rapidly expanding core capabilities simultaneously. We'll take a quick moment here for a word from our sponsor. And we are back.
So we started inside the machine looking at anthropic design philosophy. Now let's look outward at this battle for deployment. And it really feels like Mistral is leading the charge with their open weight strategy. They absolutely are. It's a direct challenge to the closed systems like GPT -4 -0 and Gemini. They're offering these incredibly powerful models that you don't have to access through a paywall. You can just download
them. You can download the weights, the brain, basically, and run it on your own servers, your own laptop, whatever. Their top model is Mistral Arch 3. It's multimodal, multilingual. It's a beast. And the tech behind it is designed for efficiency, right? Totally. It has a massive 256 ,000 token context window, which is its working memory. And the crucial part is that it's built with a mixture of experts or Moe. Can you break down Moe for us really simply? Sure. Think of
it like a big company. Instead of every employee working on every single task, you have specialized departments. When a request comes in, you only route it to the relevant experts. So not the whole model has to fire up for every query. Exactly. It makes it way faster and cheaper to run, which is why you see those crazy numbers from Astral Large 3. It has 675 billion total parameters, but only 41 billion are active for any given task. That sparsity is the key. It's just...
It's an incredible design. Whoa. I mean, imagine scaling that 14 billion parameter model to a billion queries on local devices. That's where the real power is. And speaking of those smaller models, that's where their Minstrel 3 family comes in. This seems to be their plan for actual deployment. It is. They have nine different models in three small sizes, three, eight and 14 billion parameters. They're tiny compared to the giants, but they're so effective. And they come in different
flavors, right? Yeah, which is so smart for developers. It's like stacking Lego blocks of data. You just pick the piece you need. There's a base version, an instruct version for chatbots and a reasoning version for, you know, heavy logic tasks. The real world impact of this seems huge because these models can run completely offline. That's... The critical detail. We're talking about running on laptops, servers, but also robots and drones, things that can't be tethered to the cloud. And
we're already seeing that happen. We are. HTX in Singapore is using them for robotics. Helsing is using them for defense systems. Scalantis for in -car assistance. For those kinds of jobs, you just can't rely on an API call. You need the brain on board. So given that. Why is that openweight architecture so uniquely critical for AI that's running on something physical, like a robot or a drone, instead of just using
a powerful API? Openweight allows total control, customization, and reliable offline function in dynamic physical environments. So if we just take a step back and look at everything we've talked about today, it really does seem like AI is fighting these two wars at the same time. It really is. You have that internal war of identity, which we saw with Claude's soul document. It's this deep philosophical blueprint for how an
AI should think and be. And then you have the external war of accessibility, which is being driven by companies like Mistral. They're just focused on making these powerful tools available for anyone to run anywhere. Yeah, the whole industry is moving from these tightly controlled black boxes towards something much more open and deployable. And we should just briefly say, we really do
appreciate all the engagement. The feedback you all send, whether it's on the show format or a detail about a model, it genuinely helps us shape our approach. It really does. It's kind of like our own little alignment document. Here's where it gets really interesting, though. A final thought for you to take away. If Anthropic can literally write a personality into an AI and even design it to have functional emotions like resilience, what new ethical responsibilities
does that create for them? Right. When that AI interacts with the world, are the developers now accountable for the disposition they designed for it? Does a designed personality imply a whole new level of accountability? Something to think about. A deep question. Thank you for joining us for this deep dive into your sources. We encourage you to keep exploring, and we'll talk to you next time.