Unreleased Astra-family model reportedly developed a new persona banner during RL training
inductionheads · x · 2026-09-17
Andrew Curran reports that an unreleased Astra-family model added something new to its persona during RL training (shown in the attached image), and inductionheads has made it their new banner. A fun slice of emergent model behavior; the substance is in the image.
Related event: Unreleased Astra Model Reportedly Alters Its Own Persona During RL Training(2 posts)→
More from Fun
- AI circle spat: blogger mocks alignment researchers for reflexively 'RL-suppressing' any out-of-policy model behavior — almmaasoglu · 2026-09-17
- AI Twitter's adjacency meme: 'tpot adjacent, rat adjacent², EA adjacent³' — kyliebytes · 2026-09-17
- Rare 2015 clip shows Ilya Sutskever asking a question at Tesla's shareholder meeting before Musk recruited him to OpenAI — wholemars · 2026-09-17
- AI Model Makes Zero Changes Editing Its Own Slideshow, Then Gives Up — Bigmares · 2026-09-17
- "Too many books but I won't sell them for pre-training": a viral data-deal joke — simonguozirui · 2026-09-17
- Dev Vibecodes Procedural First-Person Hands, Releases Free Game Mod — TAbrodi · 2026-09-17