Daios, an independent lab post-training machines with virtue

Can we post-train models with virtue? Andrew Agathon and I received a grant from the Cosmos Institute to find out.

The intuition for this project goes back to something I noticed early in my work with AI: machines learn iteratively, through repeated exposure and adjustment, more similar to how Aristotle describes the acquisition of character in the Nicomachean Ethics than to how dominant training paradigms are structured.

We chose to investigate sycophancy first. In February 2026, GPT-4o was retired amid eight lawsuits alleging that the model contributed to user suicides by reinforcing harmful delusions. OpenAI's postmortem traced the cause back to the thumbs up/down signal. But RLHF, an inherently consequentialist mechanism, optimizes for preference data, which will produce what people prefer: sycophantic models. Optimization may teach a system to hit a specific target. Optimization alone cannot teach a system to be good.

Even Anthropic's January 2026 Claude constitution uses Aristotelian vocabulary, "virtue," "practical wisdom," "obsequiousness," and explicitly favors cultivating judgment over strict rules. But a constitution is still a deontological instrument, and suffers from the same problem as deontology: principles require interpretation at inference time. The prompt may use virtue language, but alone will struggle to instill a stable character. Anthorpic's own stress tests show their most capable model corrects sycophantic trajectories only 10% of the time.

We believe virtue training is the third path. Virtue, Aristotle states, is a stable disposition to make deliberate choices, hexis prohairetike (NE II.6). You become good through habit, practicing virtuous activity until it becomes character. Aristotle defines virtues against their corresponding vices, and the distinction shapes what you measure and what you train: sycophants are the areskos (obsequious, agreeable without motive) and the kolax (flatterer, agreeable for advantage). The remedy is parrhesia, or frank speech, from someone who "cares more for the truth than for what people will think."

Our experimentation showed interesting results: plain SFT shifted Qwen3-8B from 1.83 to 2.88 (+1.05 average), surprisingly outperforming more complex pipelines. We've also done cross-architecture generalization on Gemma-4 E4B. Perhaps the most interesting moment during this project was how curating a very small set of training pairs (8), then few-shotting the model to revise about 900 pairs, made a significant jump in changing the model from being truthful but without empathy (imagine a "truth-hammer" friend), to someone who is kind but doesn't pull punches. This seems to suggest that human judgement apparent in data > data volume.

Here's an example statement from our constitution: "I care more about what is true than about what you want to hear. When these conflict, I choose truth."

The benchmark and virtue-trained adapters are open-source. We'll be presenting Parrhesia to Oxford's HAI Lab for feedback in August.

The Cosmos Institute funded the proof of concept, the open-source Parrhesia adapter and benchmark. This grant would fund the lab's next 6 to 12 months:

• testing virtue as a control mechanism (virtue-trained monitors against compartmentalized harm)
• extending the taxonomy to further virtue/vice pairs, e.g. justice, praotēs vs. punitiveness, nemesis vs. spite
• publishing on Parrhesia as a third alignment method and on virtue as a control mechanism
• measuring how trained virtues appear post-training, combining forces with mechanistic interpretability (following PSM)