Tongyan Lingzhi, a subsidiary of the Beijing Institute of General Artificial Intelligence's (BIGAI) industrial platform, has introduced its approach to the concept of personal AI. This approach involves creating models that continuously learn based on the data of a specific individual, project, or company, rather than relying solely on queries and information retrieval.
Most existing corporate AI systems today adapt to external users. They use queries to establish rules, extract documents via search, store state in memory, and package reusable steps. However, the model weights themselves rarely change after interaction.
Tongyan Lingzhi asserts that stable habits, such as a team's engineering standards or an investment group's risk management discipline, can be integrated directly into the model parameters. The company's architecture provides for maintaining a general base model for processing common knowledge and reasoning, to which lightweight Personal LoRA adapters are added for each individual user, project, or organization. The company notes that these individual capabilities must be preserved even if the base model is replaced.
To maintain the training of these adapters, the team is developing a mechanism of recursive self-improvement, or RSI. Task traces, human corrections, and resulting outcomes become potential training data. A specialized evaluator, configured according to the standards of each user or company, selects signals worth studying before making an incremental LoRA update. The updated model is then run on the same tasks as the previous version, and only improvements are outputted to the working environment, while regressions are canceled.
CTO Chu Bingbin emphasized that one deviated instance is merely a signal, and only recurring patterns should become training data. Furthermore, the company is working on building infrastructure to make thousands of such adapters available. The TIP system performs training and inference in parallel on the same GPUs, and Paged LoRA manages adapter memory in pages. In one test phase using Qwen2.5-72B-Instruct on eight Nvidia H800 GPUs and an adapter memory budget of 5.79 GB, Paged LoRA demonstrated better latency and tail behavior compared to flat mapping when LoRA queries accounted for 1% to 30% of traffic, although performance decreased at 50%.
Tongyan Lingzhi acknowledges that the work is in its early stages. Although the technical path of RSI and the product architecture are defined, and modules are in the design phase, the fully automated training loop is still under development. Open challenges include catastrophic forgetting, feedback contamination, and model drift.
