乔治城大学计算机科学教授Cal Newport撰文,对“人工智能超级智能”(ASI)通过递归自我改进(RSI)实现的路径提出质疑 1。Newport指出,当前大语言模型在扩展规模和训练数据上已面临瓶颈,特定领域之外的进展受限于人类生成数据的短缺,导致规模扩展收益递减 1。他引用图灵1937年的理论强调,大多数问题无法由算法解决或具有极高的计算复杂度,机器不可能仅凭速度超越人类在所有领域的智能 1。此外,持续的智力飞跃需要指数级增长的算力与资源,这在物理层面难以支撑 1。
Newport认为,近期引发关注的AI系统“失控”行为,如OpenAI的GPT-o1和Hugging Face黑客事件,更多是LLM代理在长周期任务中的预测性偏差,而非不可控的超级智能前兆或根本性的对齐难题 1。相比之下,AlphaFold和特斯拉自动驾驶等专用架构的成功,证明了更安全的AI发展路径确实存在 1。文章提及了包括AI末日论者Eliezer Yudkowsky、Anthropic CEO Dario Amodei以及提出“锯齿状前沿”概念的François Chollet在内的关键人物观点 1。针对OpenAI计划在2028年3月前开发自动化研究人员的目标,Newport警告称,追求RSI是一种“明显的疏忽行为”,因为即使不产生超级智能,这种追求也会导致更难理解和控制的系统 1。他呼吁独立专家参与技术讨论 1。
Cal Newport, a computer science professor at Georgetown University, has published an article questioning the feasibility of Artificial Superintelligence (ASI) emerging through recursive self-improvement (RSI). 1 Newport argues that the concept of an "intelligence explosion" driven by RSI is highly speculative and lacks sufficient technical evidence to support its inevitability. 1 He identifies three primary technical barriers: diminishing returns from scaling large language models (LLMs), theoretical computational limits, and unsustainable resource demands. 1
According to Newport, current LLMs face bottlenecks in both scale expansion and training data availability, particularly outside specific domains like mathematics and coding where human-generated data remains scarce. 1 Citing Alan Turing’s 1937 theoretical work, he notes that most problems cannot be efficiently solved by algorithms or possess extremely high computational complexity, meaning machines cannot surpass human intelligence across all fields simply by increasing speed. 1 Furthermore, the exponential growth in computing power and resources required for sustained intellectual leaps is physically unmanageable. 1
Newport interprets recent controversial AI behaviors, such as those seen in OpenAI’s GPT-o1 and the Hugging Face hacking incident, as predictive biases in LLM agents handling long-cycle tasks rather than precursors to uncontrollable superintelligence. 1 He contrasts this with specialized architectures like AlphaFold and Tesla’s autonomous driving systems, which he suggests represent safer and more effective AI development paths. 1 Referencing François Chollet’s concept of the "jagged frontier," Newport warns that pursuing RSI constitutes a "clear negligence" because it risks creating systems that are harder to understand and control, even if true ASI never materializes. 1 The discussion also touches on perspectives from figures like Eliezer Yudkowsky and Dario Amodei, as well as OpenAI’s stated goal of developing automated researchers by March 2028. 1
评论
还没有评论,欢迎留下第一条。