Shanghai-based Goku Technologies submitted a research paper to the top AI conference NeurIPS, known as the “AI Olympics.”

The paper introduces a new training method called Step-wise Adaptive Hybrid Training (SASR), developed with Shanghai Jiao Tong University and its AI unit, Shanghai AllMind.

The research critiques popular training methods, including supervised fine-tuning (SFT) and reinforcement learning (RL).

These methods are often used by companies such as OpenAI and DeepSeek, and the paper claims that SASR provides improved performance.

Goku says SASR is inspired by human reasoning, but did not comment on the paper’s submission.

This development underscores China’s ongoing advancements in AI technology.

🔗 Source: South China Morning Post

🧠 Food for thought

1️⃣ China’s AI training innovations reveal an algorithmic path around hardware restrictions

Goku’s SASR training framework exemplifies how Chinese AI firms are advancing through algorithmic optimization rather than relying solely on cutting-edge hardware.

This approach mirrors DeepSeek’s strategy, which achieved competitive results at significantly lower costs through techniques like Mixture of Experts (MoE), multi-token prediction, and FP8 mixed-precision training 1.

Despite U.S. semiconductor restrictions intended to limit China’s AI progress, these algorithmic innovations have enabled continued advancement, confirming NVIDIA CEO Jensen Huang’s assessment that export controls have been ineffective.

The emphasis on training efficiency and novel frameworks like SASR shows how Chinese AI development has adapted to hardware constraints by prioritizing software optimization, creating models that require less computational power but achieve comparable results.

This algorithmic focus may actually accelerate innovation in resource-efficient AI, potentially yielding models that can run effectively on less powerful hardware, which is a significant advantage for deployment in regions with limited infrastructure 2.

2️⃣ Financial institutions are emerging as unexpected AI research powerhouses

Goku’s AI research breakthrough follows a pattern established by High-Flyer hedge fund, which spun off DeepSeek as an independent AI research entity in 2023.