In the past fortnight, the artificial intelligence (AI) landscape has been abuzz with discussions around Recursive Self-Improvement (RSI) technology. Both OpenAI and Anthropic have unveiled their intentions to accelerate R&D in this area, signaling a future where AI will play an increasingly active role in its own evolution. On the domestic front, Unisound has made waves with the introduction of its U2-Flash model, which integrates RSI mechanisms into its post-training phase, enabling continuous capability enhancement through autonomous closed-loop evolution.

Practical tests have demonstrated the U2-Flash model's remarkable versatility. It can adeptly identify and rectify concurrent code bugs, craft PowerPoint presentations for Xinjiang travel guides, compute the decade-long costs associated with homeownership versus renting, and even construct 3D office environments. The model leverages a Sparse Mixture of Experts architecture, boasting a total of approximately 266 billion parameters. However, only around 10 billion are activated at any given time, optimizing efficiency.

In various evaluations, encompassing coding, office automation tasks, and knowledge-intensive mathematical reasoning, the U2-Flash model's performance stands toe-to-toe with that of flagship trillion-parameter models. Furthermore, the model offers adjustable reasoning intensity, rapid first-character response times, a 35% reduction in Agent task completion duration, and cost-effectiveness.

The U2-Flash model's leap in capabilities can be attributed to its innovative post-training strategy. The model autonomously pinpoints its weaknesses and generates tailored training data, dynamically adjusting task difficulty to suit its learning curve. It employs asynchronous Agent Reinforcement Learning (RL) to bolster training efficiency and sidesteps the seesaw effect of ability fluctuations through multi-teacher online policy distillation. Additionally, an optimized supporting engineering infrastructure enables Agents to monitor and rectify training failures, ensuring seamless operation.

Currently, the U2-Flash model is undergoing trials within Unisound's internal daily operations, supporting expansive context windows, tool invocation, and localization adaptations. Unisound remains committed to delving deeper into the RSI approach, paving the way for a future where AI continuously refines and elevates its own capabilities.