行业

Gary Marcus:若对齐问题无法解决,就该暂停 AI 发展

精选理由

Marcus 和 Faggella 就对齐问题吵起来了,一个说解不了就停,一个说这问题本身就问错了。双方都引用了 Paul Christiano 的访谈,能一次看清 AI 安全圈内部的主要分歧。关心 AI 治理和风险讨论的话值得看看这两边的完整论证。

Gary Marcus 在 X 上转发并回应 Daniel Faggella 对 Paul Christiano 访谈的评论,提出核心观点:如果对齐(alignment)问题真的无法解决,AI 发展就应该暂停。Faggella 则持相反立场,认为"先解决对齐再前进"在根本上不可能实现,人类需要转向"参与变化"而非"控制终点"。

正文

Faggella 的论据来自 Paul Christiano 在 Hinton 讲座中的表述:未来人类可能不得不用 AI 去控制另一些无法被控制的 AI,也就是说"控制"本身依赖越来越多我们不理解的技术。他主张把问题从"如何让 AI 永远服从人类"改为"哪些发展路径能让这场变化对生命整体有利",接受"影响轨迹"而非"选定目的地"。

Marcus 在引用时保留了更谨慎的立场,坚持"解不了对齐就暂停"。这场对话反映了 AI 安全界内部的分歧:以 Christiano 为代表的"技术对齐优先"路线,与 Faggella 这类"对齐概念本身需要重构"的立场,正在围绕同一批访谈内容展开争论。

原文 · Gary Marcus

if we truly can’t solve alignment, we should pause. period. Daniel Faggella @danfaggella these christiano talks are excellent / we *should* slow down - but the solution of "solve alignment" is flatly impossible and we need to move beyond it the best we can so is to *participate* in change (in terms of what we build [AGI], and become [BCI]) in the changing regime of earth-life that's coming there is no "all posthuman minds do what humans want until the end of time." clearly, non-biological intelligence will have more and more causal purchase in the world, and in order to prevent that, we'll need more AIs to do more things we don't understand to retain "control" (loosely) over the other ones we don't understand paul himself addresses this in the hinton lectures: how we'll need AIs to control the uncontrollable AIs, etc, etc, etc we need to deal with the fact that ALL is flux, that we are part of a living process that is flux, and that the BEST we can do is evolve WITH AGI for a short period (we probably won't have long), and to ensure that the AGIs we are building are capable of carrying forward sentience and self-transcendence beyond us, as we've carried it beyond the sea snails before us i wish this wasn't happening at the speed it's happening but I genuinely beleive that that "we have no solution, but we need to solve alignment" take isn't where things can stop we must ask: "If we can't eternally stop or control it, what trajectories of development would make it a kind of flourishing of the living process (as we are), vs a net detriment to all of life?" we need to accept that influencing the trajectory is all we have, not "picking destinations" (i.e. "eternal robo-world where humans don't have to work and all posthuman intelligences eternally obey us"). we need to move beyond "solve alignment" as the END of the discourse, it hasn't been reasonable from the beginning 🔗 View Quoted Tweet📊 2 ⚡

产业坐标与信源核验

本条情报归属 AITOP「行业」核心栏目,重点聚焦北美前沿实验室与硅谷科技巨头。涉及核心实体与技术议题包括:Gary Marcus、Paul Christiano、AI安全、对齐,读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。

信源核验说明:本文原始线索采编自「Gary Marcus」(@GaryMarcus),经 AITOP 多维评分管线结构化提炼。