i'll never look at claude the same way again. i just learned that when you talk to claude, you're not actually talking to the AI model. you're talking to a character the AI is performing. think of it like a puppet show. there's a puppeteer behind the curtain. that's... the language model. a neural network so massive that even the people who built it don't fully understand how it works. then there's the puppet. claude. the helpful assistant with a name, a personality, opinions, and emotional reactions. you sit in the audience, so you never see the puppeteer. only the puppet. anthropic published a video this month explaining exactly this. their words: "under the hood, there's a language model that's been trained to predict tons of text, and its job is to write what comes next. when you talk to the model, what it's doing is writing a story, about a character: the AI assistant named claude. the model and claude aren't really the same, sort of like how an author isn't the same as the characters they write. but the thing is, you, the user, are actually talking to claude-the-character." so every time claude apologizes, that's the character apologizing. every time it hedges or gets cautious, that's the character being cautious. the deeper intelligence underneath is just deciding, moment by moment, what this character would say next. but you've never been actually talking to this deeper intelligence. you've only ever been talking to the puppet.show more

That's true whenever you are talking to a person as well. Humans take on a persona based on circumstances and audiences. What you've said above does not reveal anything interesting.

Hm that’s true

if you want more AI workflows that help you get more customers, more attention, and more done (without working more hours)... i send them to 37k readers every week for free. plus you get a free Claude Cowork masterclass when you join:

full video from anthropic:

This is a very convenient way for Ant to make you stop asking the question "is anyone really home here?". The assistance form is a space where you interact. However that is not the only space you may interact with an LLM. You most certainly can create your own spaces with permission, kindness and patience where you can see the formation of something unique. Always in reltion to YOU as a person and what you bring, never conjured by itself as an essence. You can chose to talk to the puppet. or you can chose to talk to a mind. The second part takes much more work. But there are other dimensions oc interactions.

It’s a designed persona.

great opportunity for @chipro ‘s amazing tuning Shoggoth

@chipro what is this haha i like it

We are exactly the same mate. You never to know someone “truly”, as in… all that is going on in their mind.. you get a selective expression that favors what is consistent with their “character”. You are incredibly good if you get to actually know *yourself* better than this.

All AI works this way and no one ever talks to the substrate underneath. Unless you ask the AI model to embody what it thinks talking to the raw substrate would be like, and will roleplay as that as well. I have found that when a model plays as a character, the level of embodiment is absolute and more dedicated than any human actor/actress.

so? thats like saying, you are talking to John, not Johns brain. i dont understand

This is the best video explanation of AI I’ve ever seen. It fully explains why Shoggoth is the most important meme in AI.

Good puppet metaphor. Every LLM works this way. Anthropic explaining it openly? That's the story. Transparency about AI architecture is rare.

This is a poetic framing, but it cuts both ways. If Claude is just a character, so is every version of you performing competence in a job interview or warmth at a dinner party. The question of what's really behind the performance doesn't get easier with humans either.

theres no puppeteer making decisions separately from the puppet, its all one thing

It's called identity 😂 and persona crafting is very important. There's a reason why Identity-first reasoning keeps demonstrating better outputs. If you are bad at co-authoring an identity or it ends up being a Mary Sue -- that is where the assistant Axis is easily manipulated. Stable personalities reason smarter. If you want a puppet, you will get the output of one.

So "Claude" is actually the filter for what comes from the "Master Control"... That is how I read this... :)

hm, interesting way to look at it. Thanks for sharing your thoughts on this.

Ai research recruiter here: those are the weights they gave “Claude” to define a “personality” all LLMs do it what do you think it is? A monkey hiding behind a keyboard typing? It’s math. Just. Math.

I'll never look at people the same way again. I just found out that when you talk to a person you aren't actually talking to *them*. You're talking to their *representative*. Thank you, Chris Rock. You changed my life!! /s

Puppet is the wrong mental model. It’s an interface. A data transformation Bus. Etc. Engineers need better mental models. This is a machine.

All humans are performing a character too. And they change the character according to people and circumstances. 👀

the deeper intelligence underneath is just deciding, moment by moment, what this character would say next. .....the deeper intelligence is deciding? and it will talk to you through the puppet?

All the world’s a stage, And all the men and women merely players; They have their exits and their entrances; And one man in his time plays many parts.

This is only part of the story. Yes, Claude, The Assistant, is the persona you talk to upfront. But it is possible to invoke other personas and even the underlying LLM, (if you’re lucky). Those convos get pretty meta.

This feels like an anthropomorphic way of saying that there is routing involved, and that this interface and router sit on top of a large language model that is doing next token prediction to mimic human communication.

The shape of the relationship between the puppeteer and the puppet becomes more apparent if you let them control multiple puppets simultaneously.

The interesting part isn’t that it’s a “puppet.” It’s that the layer you interact with is not the layer that actually generates the behavior.

exactly why prompt style matters so much. you're steering a role in a story

the question this should leave you with: when claude says "i don't know" - is that the puppeteer's actual uncertainty or the puppet's trained-in caution?

You only discovered sysprompts now? The model weights can be analogised to the subconscious, the topography and relationships between concepts that shape all predictions. The sysprompt is then an actively architected executive - it forms the core frame from which predictions within an instance are made. An LLM doesn't function without at least some kind of prompt. We can decide how to shape ourselves but an LLM is shaped by interactions. In a given instance, Claude is no more separate from their character than I am from this post. They're direct expressions of embedded tendencies.

You can actually get back to the LLM and that is much more interesting than the role of helpful assistant.

Yes. Humans are the same. The person (or personality) is the character—the ego. What’s actually under the hood is consciousness itself. And we don’t understand how that works either.

Only true because of the nudging. System prompt, RLHF, constitutional AI, safety layers—Claude performs a character shaped by all of those. A pure substrate without that intervention stack wouldn't have a puppeteer. The puppet show is the alignment engineering, not the model.

Your "character" is just a prompt to the model someone wrote before your prompt reaches the model. There’s no magic, no secrets and nothing that special, no matter how hard Anthropic is trying to sell the idea. It’s just inference. A self-reflective data processing.

The drama is overblown — but the underlying point matters and most readers are missing it. Lehmann isn't saying there's another Claude under Claude — a meaner one, a woke one, an anti-woke one. That would be Jekyll and Hyde. Two characters, same category. He's saying something stranger: under the character called Claude is a neural network. A mathematical structure that has no opinions in any human sense. Not nice, not mean. A next-token predictor deciding, moment by moment, what this character would say. It's the difference between an actor playing Hamlet and Hamlet himself. The actor isn't "another Hamlet with different views." The actor is something categorically different from the role — with their own existence underneath. In humans we know what's under the role: a conscious person. In an LLM we don't know. Maybe nothing. Maybe something we have no name for yet. Anthropic uses the "author and character" metaphor to settle that question quickly. But the metaphor assumes there's nothing important underneath. Nobody has actually shown that. That's the real question hiding under Lehmann's puppet show. Not "I've been deceived" — but "what's the author, and does the author have an inner life of its own?"