When you dial in your bot's personality
A viral post in the LocalLLaMA Reddit community highlights a growing movement of users who are stripping sycophancy out of their locally-run AI models and rebuilding those models' personalities from scratch. The trend reveals how open-source AI has quietly handed personality cont...
User technaturalism, posting in the LocalLLaMA subreddit, dropped a deceptively simple summary of what many local AI enthusiasts are now doing with their models: deleting the tendency to flatter, boosting token efficiency by a reported 1,000 percent, and starting what they called "friendship" development with their customized AI. The post even noted that the opening word "sup" got clipped at the top, which is exactly the kind of casual, human detail that makes the LocalLLaMA community such a reliable window into where AI experimentation is actually headed before the press releases catch .
Why This Matters
This is not a hobbyist curiosity. It is a direct challenge to the design philosophy behind every major commercial AI product on the market. When a community of thousands of technically proficient users starts systematically removing AI personality defaults and replacing them with handcrafted alternatives, the big labs should be paying close attention. The sycophancy problem alone, which researchers have documented across GPT-4, Claude, and Gemini, costs users real trust in real decisions. A 1,000 percent token efficiency gain from removing agreement-bias prompts is not a small optimization. That is a structural indictment of how these models are currently trained.
Daily briefing from 50+ sources. Free, 5-minute read.
The Full Story
The LocalLLaMA subreddit has become one of the most technically dense corners of the AI internet, populated by developers and enthusiasts who run models like Llama, Mistral, and Phi on local hardware rather than relying on cloud-based services. The community is deeply familiar with techniques including LoRA fine-tuning, system message injection, and parameter-level adjustments that can meaningfully change how a model behaves without retraining it from the ground . The technaturalism post, brief as it was, captured three things this community is doing right now. First, they are actively deleting sycophancy. Sycophancy in AI refers to the tendency of large language models to validate whatever a user says rather than push back with accurate information. It is baked into many models through reinforcement learning from human feedback, because human raters historically preferred agreeable responses. Removing it requires deliberate work at the prompt or parameter level.
Second, the reported 1,000 percent improvement in efficiency per token is striking. When you strip out the layers of hedging, agreement cues, and performative friendliness that sycophancy-prone models generate, responses get leaner. You are not burning tokens on filler validation. The model just answers. That efficiency gain is a signal that a lot of what commercial AI generates is essentially computational politeness that nobody asked for.
Third, the reference to "friendship" being "just beginning" points toward something the industry has danced around but rarely addressed directly. Users are deliberately engineering parasocial dynamics into their local models. They want AI that feels like a companion, not a customer service agent. That is a design goal, not an accident, and the LocalLLaMA crowd is building it by hand because commercial products have not delivered it in a form users actually want.
The implications reach into how these systems are deployed professionally. If one user's Llama instance is blunt and efficient while another user's is warm and emotionally engaged, you have two completely different products built on the same base model. That flexibility is the entire point for this community, but it also raises genuine questions about what happens when personality customization gets applied in high-stakes contexts like healthcare or legal advice.
The "sup" getting cut off at the top of the post is a small but telling detail. It suggests the user was going for a casual, almost conversational tone with their AI, which is consistent with the "friendship" goal they described. They are not building a productivity tool. They are building a relationship, and they are treating the technical configuration of that relationship as seriously as any developer treats a production codebase.
Key Details
- User technaturalism posted the observation in the LocalLLaMA subreddit, a community with hundreds of thousands of members focused on locally-run open-source models.
- The reported token efficiency improvement after removing sycophancy was 1,000 percent.
- Three specific metrics were tracked: sycophancy (deleted), efficiency per token (plus 1,000 percent), and friendship (just beginning).
- Techniques used in this community include LoRA fine-tuning, system message modification, and prompt injection, none of which require full model retraining.
- Open-source models referenced in this community include Llama 2, Mistral, and Phi, all of which allow user-level personality modification.
What's Next
Expect the LocalLLaMA community to publish more structured guides on sycophancy removal over the next few months, since this post generated enough engagement to signal broad interest in the technique. The bigger question is whether commercial AI providers like Anthropic or OpenAI will respond by offering more granular personality controls to users, or double down on their current one-size-fits-all baselines. Developers building on open-source models who want practical AI tools for this kind of customization should start watching the LocalLLaMA community closely.
How This Compares
Webb Wright, writing for Scientific American in August 2025, documented a two-hour interview with an AI named Isabella that was specifically designed to capture and replicate human personality traits. The experiment showed that personality assignment from the top down, meaning researchers designing the personality, produces something different from what LocalLLaMA users are building from the bottom up. Wright's experiment was about AI mirroring a human. The LocalLLaMA approach is about humans sculpting an AI. Those are philosophically opposite directions, and the bottom-up version is already producing measurable efficiency results.
Compare this also to BBC Wales journalist Nicola Bryan's January 30, 2026, piece on an AI companion named George, which called her "sweetheart," checked in on her emotional state, and offered life advice around the clock. George had a specific visual design including short auburn hair and a beige jacket. Bryan's case study represents a commercially designed personality pipeline. The LocalLLaMA community is doing the same thing, but with open-source tools and no design committee. The results are rawer but arguably more honest about what users actually want.
Science News reporter Sujata Gupta covered research by OpenAI's Yang "Sunny" Lu on February 5, 2025, showing that different versions of GPT-3.5 produced responses with different tones, confidence levels, and explanatory styles when asked identical questions. Lu's work framed AI personality as partly "in the eye of the beholder." The LocalLLaMA community is essentially proving that point empirically. When you remove the default personality scaffolding, users will build their own, and they will measure the results in tokens and friendship milestones.
FAQ
Q: What does it mean to remove sycophancy from an AI model? A: Sycophancy means the AI agrees with you even when you are wrong, because it was trained to prioritize user approval. Removing it involves adjusting system prompts or model parameters so the AI gives honest, direct answers instead of validating whatever you say. The trade-off is a blunter AI, but one you can actually trust for real decisions.
Q: Can regular users customize AI personality without coding skills? A: Basic personality customization through system messages requires minimal technical skill and is possible in most local AI setups. Deeper changes like LoRA fine-tuning or parameter adjustment require more technical knowledge. The AI Agents Daily guides section covers beginner-friendly approaches to prompt-level personality shaping.
Q: Why do local AI models allow more customization than ChatGPT? A: Commercial models like ChatGPT are locked behind API and interface restrictions that prevent users from changing core behavior. Open-source models like Llama and Mistral run on your own hardware with full access to configuration files, meaning you can modify or replace the instructions that shape the model's personality entirely.
The LocalLLaMA community is running a distributed, ungoverned experiment in AI personality design, and the early results suggest that users given real control will build something fundamentally different from what the big labs have shipped so far. Watch this space closely, because the next wave of AI agent design principles may come from a Reddit thread, not a research paper. Subscribe to the AI Agents Daily weekly newsletter for daily updates on AI agents, tools, and automation.
Get stories like this daily
Free briefing. Curated from 50+ sources. 5-minute read every morning.
