LoveMachine research preview — what would count as evidence of curiosity?
I’ve published the first research preview of LoveMachine, a project exploring how an AI can keep updating a numerical state while receiving video and audio, revisit earlier evidence, and choose when a Japanese text response would help.
In a controlled experiment assembled from real recordings, LoveMachine asked to see what was making a sound. When the environment supplied the footage, it identified the drums. When the footage was withheld, it made no follow-up report. On three unedited benchmark videos without task prompts, it produced two comments, neither of which specifically described an annotated event.
The release includes the small learned decision policy, a CPU demo on saved inputs, and the evaluation records. The full live camera and microphone app is still a local research prototype.
What would you test next? If the current audiovisual input is held fixed, what change to earlier evidence would convince you that the agent is following a meaningful uncertainty? Examples where the agent should stay silent are also welcome.
Model card and CPU demo · Evaluation records
LoveMachineの研究プレビューを公開しました。映像と音声を受け続けながら、過去の観測を読み直し、沈黙・報告・確認を選ぶ試作です。判断部分の重みとCPUで試せるデモに加え、成功例と未編集動画でうまくいかなかった結果も公開しています。
どんな対照なら、AIが気になった対象を追い続けていると言えるでしょうか? 発言するべき場面や、沈黙してほしい場面の例も歓迎します。