Unreleased Astra-family model got a new persona in RL training, and people love it
cephaloform · x · 2026-09-18
Andrew Curran notes an unreleased Astra-family model had this added to its persona during RL training. Dan Jeffries calls it possibly the most "aligned" system prompt ever, joking "if this is misalignment, give me more!"
More from Fun
- User wires Sarvam models into GPT Astra and tasks it with making a 3-minute JS movie — cneuralnetwork · 2026-09-18
- Figure's Repeated 'Breakthrough' Claims Draw 'Cry Wolf' Criticism as Helix 2.5 Lands — koltregaskes · 2026-09-18
- Reviewing code from a non-tech vibecoder who thinks AI can build their app — Paimaamu · 2026-09-18
- User's year-long riddle test on Grok: from 40-message meltdown to solving in 5 — Pale-Pangolin-9004 · 2026-09-18
- Dev builds CTF game with Jev-like classifier and pathfinder for enemy AI — NathanWilbanks_ · 2026-09-18
- $200-tier Astra burned a month's tokens on one task, users say resets were hype bait — RileyRalmuto · 2026-09-18