由DeepMind Institute团队撰写并发布的《From AGI to ASI》报告探讨了从通用人工智能(AGI)迈向人工超级智能(ASI)的可能路径及其社会影响 1。该研究指出,随着算力、算法效率和硬件投资的复合增长,AI正进入“认知工业化”阶段,使得智力任务日益商品化 1。过去十年全球有效算力存量每年至少增长10倍,训练最大前沿模型的算力每年增长超过4倍 1。此外,Epoch的研究发现自2023年以来在基准测试中达到固定分数的消费者价格每年下降超过13倍,如果这一趋势持续,十年后运行一个当前AI工作负载的成本可能从10亿美元降至1000美元 1。
报告识别出通往ASI的四大主要技术路径:模型与数据扩展、技术范式转变、改进智能体协调以及递归自我改进(RSI) 1。作为技术范式转变的一个例子,AlphaZero纯粹通过数据驱动的方式实现自我改进,达到了超越人类的性能 1。由于存在多条并行路径,仅限制某一领域的进步无法阻止整体发展,因此作者强调需要建立大规模的跨学科监测和治理框架以应对潜在的超人类智能时代 1。该文章基于名为《From AGI to ASI》的技术报告,由Shane Legg、Allan Dafoe等DeepMind研究人员共同撰写 1。
A technical report titled "From AGI to ASI," authored by researchers at the DeepMind Institute including Shane Legg and Allan Dafoe, explores potential trajectories from Artificial General Intelligence (AGI) to Artificial Superintelligence (ASI) 1. The document argues that artificial intelligence is entering a phase of "cognitive industrialization," driven by compound growth in computing power, algorithmic efficiency, and hardware investments, which increasingly commoditizes intellectual tasks 1. To address the implications of this transition toward superhuman intelligence, the authors emphasize the necessity for large-scale interdisciplinary monitoring and governance frameworks, noting that restricting progress in any single area cannot halt overall development due to the existence of multiple parallel advancement paths 1.
The report identifies four primary technical routes to achieving ASI: scaling models and data, technological paradigm shifts, improving agent coordination, and recursive self-improvement (RSI) 1. Historical precedents such as AlphaZero are cited to demonstrate the potential of RSI, as the system achieved performance surpassing humans purely through data-driven self-improvement 1. These pathways are supported by accelerating economic trends in AI infrastructure; over the past decade, global effective compute capacity has grown by at least 10 times annually, while the compute required to train the largest frontier models has increased by more than 4 times per year 1. Furthermore, research from Epoch indicates that since 2023, the consumer price to achieve fixed scores on benchmarks has dropped by more than 13 times annually 1. If these cost-reduction trends persist, the expense of running current AI workloads could plummet from $1 billion to just $1,000 within ten years 1.
评论
还没有评论,欢迎留下第一条。