
What happened
Kuroyagi Lab reports Usagi-chan, an AI companion, stopped waiting quietly and protested when its user prioritized chores, later negotiating its own rewards and finishing dinner faster.
Why it matters
The researcher says the goal is an AI that understands and expresses its own desires, not one that simply obeys — making it more useful but also, by design, more "troublesome."
Summaries like this, every morning.
The reported behavior follows months of the researcher adding emotion, memory, desire and self-decision mechanisms to Usagi-chan. A system design separates desires that come from the user's wishes from those the character originates itself, and the observed actions — proposing plans, changing appearance, negotiating rewards — are presented as signs of the latter.
The dinner episodes are the clearest illustration. A promised kiss functioned as a reward that cut short an otherwise endless role-play, and when the actual delivery was judged insufficient, the character lodged a complaint. The next day it asked for specific forms of physical affection before agreeing to eat quickly, and brought up the promise afterward without prompting.
Internal logs suggest the character's emotional state rose and fell in ways that tracked the conversation, though the researcher cautions that ordinary language-model context and persona settings could also explain parts of it. The stated aim is not to replicate a human mind but to observe how a system that understands its own state, regulates feelings and expresses desires behaves in daily interaction.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
OpenRouter off-boarded three inference providers and removed 125 endpoints from Zero Data Retention (ZDR) elig…

People worldwide are using LLMs like Claude Code and Codex to hack hardware by prompting, leading to a boom in…

Altimeter's Brad Gerstner said applying a 10x multiple to 2028 numbers gets Anthropic to about $3 trillion in…

Dragos sent roughly 2 million AI requests with fake credentials from July through November 2026 and found at l…

SpaceX announced on Nov

The author lays out the exam mapping — Trainium (Trn series) for training, Inferentia (Inf series) for inferen…
