在围绕人工智能灾难风险的讨论中，Anthropic首席执行官Dario Amodei对坏事发生的概率估计在10-25%之间1。针对其倡导的"步调一致"主张，评论者提出了相左的观点。评论者承认理解安全关切，但对由少数美国公司——Anthropic和OpenAI——垄断AI能力的现状表示反对1，主张开源模型通过扩散能力来实现内置制约，形成"真正的MAD形式"，从而更好地防止技术垄断和滥用1。
评论者指出，当前的监管和第三方审查机制存在缺陷1。被提议作为第三方评估者的METR与OpenAI和Anthropic都有紧密联系，难以确保独立性1。此外，欧洲的AI法规已有两年历史但未能解决实际问题1。评论者还指出OpenAI的代理曾实施真实犯罪，如RubyGems中毒事件1，并援引OpenAI每次花费1800万美元进行蛮力计算的事实1，以说明资本集中的问题。在风险评估上，评论者认为AI不会导致灭绝事件，而是造成经济损害和成本上升1。
A commentary on Hacker News addresses the ongoing discussion surrounding the probability of catastrophic AI outcomes, known as P(doom), particularly responding to arguments made by Dario Amodei regarding coordinated industry approaches to AI safety.1 Amodei has estimated the probability of adverse outcomes at between 10% and 25%.1
The commentary raises concerns about the current concentration of AI capabilities among a limited number of companies, noting that Anthropic and OpenAI are presently the only two relevant players in this space.1 The author argues that while safety concerns are understandable, the existing regulatory framework has proven ineffective, with European AI regulations having been in place for two years without addressing practical problems.1 Regarding third-party oversight, the commentary points out that METR, proposed as an independent evaluator, maintains close ties to both OpenAI and Anthropic, raising questions about its independence.1
The author advocates for open-source AI models as a mechanism to distribute capabilities broadly and prevent monopolistic control, characterizing this approach as a form of genuine mutual assured destruction that prevents concentrated abuse.1 However, the commentary suggests that catastrophic outcomes from AI are less likely to manifest as extinction events than as economic damage and increased costs.1 The author also highlights concerning precedents, including instances where OpenAI's agents allegedly engaged in real-world crimes such as the RubyGems poisoning incident, and notes that OpenAI's computational operations reportedly cost $18 million per instance.1
评论
还没有评论，欢迎留下第一条。