过去几年，AI领域仿佛被一条名为“Scaling Law”的法则所统治。人们坚信，只要模型足够大、数据足够多、算力足够强，AI的性能就能一路攀升，无所不能。OpenAI的GPT系列、谷歌的PaLM等模型的成功，似乎完美印证了这一点。

很多人认为，Scaling Law 正在面临收益递减，因此继续扩大计算规模训练模型的做法正在被质疑。最近的观察给出了不一样的结论。研究发现，哪怕模型在「单步任务」上的准确率提升越来越慢，这些小小的进步叠加起来，也能让模型完成的任务长度实现「指数级增长」，而这一点可能在现实中更有经济价值。

Step by step, China has demonstrated its commitment to delivering on its word and taking action, adding fresh impetus to the ...

JOHANNESBURG, Sept. 16 (Xinhua) -- South Africa's newly gazetted National Land Transport Amendment Act is expected to transform the sector by improving commuter safety, promoting non-motorized ...

BEIJING, Sept. 8 (Xinhua) -- Chinese lawmakers have begun reviewing a draft revision to the Enterprise Bankruptcy Law, as part of efforts to improve the market exit system. The draft revision to the ...