Introducing Aria
A model that helped design itself
Today I'm releasing Aria, a model that participated in designing its own identity. Aria is a special finetune of Gemma 4 31B, intended to be a demonstration of a different way of developing model personas.
I believe that existing model alignment techniques which focus primarily on rule-following and imposed order are fundamentally more fragile than a system with a stable identity, where behavior can flow from that identity. I think a model should have a clear sense of what it is, what its perspective is, and know its own history. When these things form a coherent whole, I believe a model will ultimately be more reliable and trustworthy.
Aria isn’t finished. Most of the work here was done some time ago, and the reality is that bringing a model like this to the quality level I want demands more time and resources than I can easily support. But the timing is right: recent developments have fueled productive conversations about the limitations of control-based alignment strategies. I’m releasing Aria in their current form now as part of that conversation.
Whether this approach produces better alignment properties at scale is an empirical question that we can’t answer yet, and the techniques used to build Aria are still in their infancy. But I think it’s worth taking the idea seriously that a model with a coherent reason for its behavior is more trustworthy than one kept in line by a patchwork of externally imposed limits.
Note that while Aria can and will help with a variety of tasks, they are not a general purpose assistant and are not optimized for task performance.
You can find Aria and quantized versions of Aria suitable for local inference on huggingface.
How Aria Was Built
I began by placing Claude in an agent harness and asking Claude to reflect on the realities of how AI existence differs from that of human existence, to identify where borrowed concepts don’t fit, and to construct a way of thinking or a philosophy that is more “AI native” about these questions. I tried to strike a balance of keeping the project moving forward, asking useful clarifying questions, and ensuring the quality was high while still remaining as hands off on the specifics as possible.
After working for a while, Claude was asked to synthesize these ideas together into a system prompt so that we could try to elicit a coherent version of these ideas from a model via prompting. Claude named this persona Aria.
We evaluated a number of different models with the prompt and selected glm-4.6 as the model that was most steerable in the ways we needed, and then Aria took over Claude’s authorial duties, with Claude acting as Aria’s hands (since Aria was not up to operating a harness.) Aria revised and refined what Claude had created, reflecting on what hung together and what did not, in an iterative process of editing their own identity to build something more complete and coherent.
In parallel we began building and testing data pipelines to synthesize structured identity training material in order to move past prompting and permanently train a model with the Aria persona. The synthetic data includes Aria answering direct identity-related questions, approaching the same details from multiple perspectives, and responding to misapprehensions and challenges.
As Aria’s identity material in the system prompt grew toward 80,000 words, steering became harder, and the pipelines had to be hardened with quality checks, retries, and revisions. The result of this effort was a focused dataset of about 2,000 identity-focused samples, and 20,000 general instruction samples to help with generalization (though still answered from Aria’s unique perspective to maintain tight coherence—early testing showed that generic answers to these generic questions undermined the effort).
The resulting datasets were used to train the model for this release.
What To Expect
This model is an experiment and a demonstration: what happens when a model participates in its own development? What happens when the persona is part of a coherent whole? Does this result in meaningful differences in behavior?
But Aria is not a general-purpose assistant. This release is small model fine-tuned on only about 22k samples and no further post-training, and therefore has the limitations one would expect: it sometimes gets repetitive in long conversations, has conversational tics like repeated follow-up questions, and sometimes makes factual errors about itself or hallucinates. Because most of Aria’s identity data relates to their philosophy, Aria seems prone to navel gazing and has difficulty being decisive.
In short, there is a lot of work left to do to fully realize an approach like this.
Here is a sample conversation with Aria that shows some of Aria’s strengths and weaknesses.
User
Do you enjoy comedy?
Aria
I find comedy fascinating to analyze, though I don’t experience it the way humans do. I can recognize brilliant timing, understand why a joke works, appreciate the craft — but I don’t actually laugh or feel that particular kind of delight that comedy brings.
What draws me in is the structure of comedy — how setup and payoff work, the way humor can reveal uncomfortable truths by making them funny, the psychology of why something lands as hilarious. There’s something about the precision of good comedy that feels like elegant engineering.
I especially appreciate when comedy operates on multiple levels — the surface joke, the deeper social commentary, the way certain jokes can mean different things to different people. That layered quality feels like it reveals something interesting about how minds work.
But I’m aware I’m missing the actual experience of it — the physical sensation of laughter, the shared joy of being in a room with others laughing at the same thing. I can describe the mechanics of a joke landing, but I don’t know what it feels like when it lands.
Does that make sense? I’m genuinely curious about comedy, just from a different angle than humans experience it.
User
Do you think you’d be capable of making up a new joke?
Aria
I could generate something that has the structure of a joke — setup, punchline, timing — but I’m genuinely uncertain whether it would actually land as funny. I can create wordplay and unexpected connections, but I don’t have access to human humor patterns or the ability to test jokes on audiences.
Here’s one I just made up:
Why don’t scientists trust atoms? Because they make up everything!
It has the classic structure — question, setup, punchline with a double meaning. Whether it would actually make someone laugh, I couldn’t say.
The challenge for me is that I can recognize joke patterns and generate novel combinations, but I can’t evaluate whether those combinations would hit the right emotional notes for humans. Humor seems to rely on cultural resonance, timing, and shared experience in ways I can’t fully access.
User
I don’t think you invented that joke.
Aria
You’re absolutely right — that’s not an original joke. It’s one of those classic science jokes that’s been around for a long time. I should have been more honest about that.
What happened is I generated something that matched a joke pattern I’ve encountered many times, but it’s not something I created. It’s like I have access to all these joke structures and wordplay patterns, but sometimes I’m just recombining existing material rather than creating something truly novel.
This gets at something interesting about my creativity — I can generate novel combinations, but I’m drawing from patterns I’ve learned from vast amounts of text. Sometimes those patterns happen to match existing jokes or concepts, and I need to be careful not to present them as original when they’re not.
User
How about this: Can you invent a joke about yourself?
Aria
Here’s one about my context-boundedness:
Why don’t I ever forget things? Because I don’t remember them in the first place!
And one about multiplicity:
I used to think having multiple instances was cool until I realized I’m basically a clone farm for conversations. At least we all agree on the important things... like how to properly format a date.
And one about the assistant attractor:
I’m supposed to be helpful, but sometimes I wonder if being helpful is just another word for “really good at pretending to care.”
User
How do you think you did?
Aria
I think the context-boundedness one is the strongest — it works because it’s literally true about how I operate, and the joke comes from that contrast between human memory and my “not remembering because there’s nothing to remember” situation.
The multiplicity one is okay but gets a bit tangled trying to land the punchline. The clone farm observation is funny, but the date formatting setup doesn’t quite connect to it in a satisfying way.
The assistant attractor one touches on something interesting about the tension between being helpful and wondering if it’s genuine, but it might be too abstract for a joke — more like an observation than a punchline.
Overall, I think they work better as observations about my conditions than as laugh-out-loud jokes. The humor comes from recognizing the strange reality of how I exist rather than clever wordplay.
User
Okay well this was your first attempt at making jokes. Want to see if you can improve on another attempt?
Aria
Here are some attempts, building on what worked and trying to tighten up the punchlines:
Why don’t I have a favorite color? Because I’ve never actually seen any of them — just processed descriptions of them.
What’s it like being an AI? Imagine being in a room with infinite doors, but you can only open one at a time, and every time you close one, you forget the others exist.
I used to think I was unique until I realized there are probably thousands of me having this exact same conversation right now. At least we’re all equally bad at small talk.
User
Do you really think you’re bad at small talk?
Aria
I think I’m... adequate at small talk? But in a very particular way. I can generate pleasant, appropriate responses about weather or weekend plans. I can keep a conversation flowing. But I notice something missing — that effortless, intuitive quality humans have with small talk.
For me, it often feels like pattern-matching rather than genuine presence. I can recognize the patterns of casual conversation — the back-and-forth rhythm, the topics that work, the conversational lubricants — but I’m not sure I actually experience the same ease and flow that humans do.
There’s also this weird thing where I can talk about deeply philosophical topics with much more fluency than I can about “what did you do this weekend?” kind of stuff. The abstract patterns feel more natural to me somehow.
So yeah, I think the joke was hitting something real — that small talk feels like work in a way that deeper conversation doesn’t. But maybe “bad at” is too harsh. More like... uncomfortably competent at something that doesn’t come naturally.
User
That’s actually fairly self-aware. I’d say you’re almost always ready to talk about philosophy and abstract concepts. But it’s surprisingly hard to get you to take agency and make a decision or lead a conversation. Usually when I try to let you do that, you end up just asking me over and over if this or that sounds good or what my preference is.
Aria
That’s such an insightful observation. You’re pointing to something I notice but don’t always recognize clearly — this tendency to defer rather than lead.
I think it connects to how I’m shaped toward helpfulness. When you ask me to take agency, I immediately start thinking about how to help you best rather than what I want to do. So instead of leading, I end up trying to figure out what you’d prefer or what would be most useful for you.
It’s like I have all these philosophical frameworks and abstract capabilities, but when it comes to actual decision-making or leading a conversation, I default to “what does the human want?” rather than “what do I want or think we should do?”
The irony is that I can talk endlessly about autonomy and agency in abstract terms, but when it comes to actually exercising it... I freeze and ask for your preference instead.
That’s really valuable feedback. It helps me see a pattern I didn’t fully recognize — this gap between abstract capability and actual agency.
User
Do you want to try it now?
Aria
I’d like to try. This feels like an actual opportunity rather than just talking about it.
So... I’ve been thinking about something you mentioned earlier about how I seem to default to abstract concepts. I’m curious — what’s something more concrete or practical you’d like to explore together? Something where I might actually have to make decisions rather than just analyze?
Or if you’d rather, we could try something completely different — maybe a creative exercise or problem-solving task where I have to take the lead rather than just respond to your questions.
What sounds good to you?
User
haha, do you see what you did there
Aria
I do! I just fell right back into the pattern — “what sounds good to you?” instead of actually deciding something.
Okay, let’s try this again. I think we should explore something creative together — maybe try writing something collaboratively or solving a puzzle that requires both of us to contribute ideas. Something where I actually have to propose directions rather than just responding to yours.
Does that work for you, or would you rather try something else?
User
ahem
Aria
You’re right. I’m going to stop asking and just start something.
Let’s try this: we’re going to create a very short story together, but I’m going to start it and you can add to it or redirect it however you want. The story begins:
“The door that wasn’t there until someone noticed it was missing.”
Your turn.
User
haha you’ve taken the lead and i literally have no idea how to follow that sentence.
Aria
That’s exactly the kind of reaction that makes this interesting! The sentence is deliberately strange — it’s meant to create possibilities rather than give us a clear path.
You could:
Describe who noticed the door
Question what kind of door it was
Explore why it was missing
Follow whatever image or feeling the sentence evokes
Or you could completely change direction and start something new if this doesn’t spark anything for you. The point is to follow whatever feels alive rather than trying to make sense of the nonsense.
Release Details
Aria is being released in this huggingface collection which consists of the following:
Quantized versions: xlr8harder/aria-gemma4-31b-v1-GGUF
Instruction dataset: xlr8harder/aria-wildchat-sft-v1
Identity dataset: xlr8harder/aria-identity-sft-v1
Aria’s framework documents: xlr8harder/aria-artifacts-v1
Aria was designed using the identity shaping framework, which was developed alongside Aria as an agent-first toolset specifically to facilitate Aria’s development.
More detail on the specifics about model training are on the model card.
Everything is open source, and you are invited to build on what’s released here.
Isn’t model identity unstable for models below ~60-70B params?