BIGAI spinout Tongyan Lingzhi bets on personal LoRA and self-improving AI

Tongyan Lingzhi, incubated by BIGAI's industry platform, outlined a personal AI approach: a shared base model plus lightweight Personal LoRA adapters for each person, project or organization. Adapters learn via recursive self-improvement with evaluator filtering and rollback of regressions; TIP and Paged LoRA infrastructure aim to make thousands of adapters affordable.

• Architecture pairs a shared base model with per-user or per-team Personal LoRA adapters
• RSI filters candidate data through an evaluator and ships only improvements, rolling back regressions
• Paged LoRA on 8 H800 GPUs with a 5.79GB budget beat a flat layout at 1–30% LoRA traffic
• The fully automated learning loop is still in development amid forgetting and model drift