Yann LeCun 发文反驳 AI 灭绝论，称 p(doom) 估计毫无意义
LeCun 这次把话说得挺直白：p(doom) 全是拍脑袋数字，Hinton 那套灭绝论他不买账。他还顺带介绍了自己的 Objective-Driven AI 架构，主张靠推理时护栏保安全。关注 AI 安全路线之争的值得看看原文。
Meta 首席 AI 科学家 Yann LeCun 在社交平台发文，直接反驳 Geoffrey Hinton 一方的 AI 灭绝风险论。他表示灾难场景（灭绝、就业末日）对了解现实世界如何运作的人来说完全荒谬，p(doom) 概率估计都是作者凭空捏造的，对人类拥有完全能动性的事件赋予概率没有意义。
LeCun 从两个层面展开论述：一是对人类的信心不足只是暴露了提出者的悲观，这是他与 Hinton 分歧之一；二是智能不等于权力和支配，病毒、毒蛇、老虎都能放倒人类，现实中大量愚蠢的人支配着更聪明的人。他认为超级智能 AI 会在安全护栏约束下为人类工作，不会有接管世界的欲望。
他还介绍了自己提出的技术方案 Objective-Driven AI 架构，主张在推理时通过满足安全护栏来保证系统安全。这篇发言延续了图灵奖得主之间关于 AI 安全风险的公开争论，表明 LeCun 与以 Hinton 为代表的生存风险派在技术路线上仍在持续对立。
In democratic societies, everyone has *some* responsibility on how the use of technology maximizes benefits and minimizes bad side effects. The benefits of AI are real today, and will greatly expand in the future. The bad side effects have, so far, been minimal. It's on all of us to make sure it remains that way. The catastrophe scenarios (extinction, employment apocalypse, etc) are totally ridiculous for anyone who understands how thing work in the real world. And the p(doom) estimates are all pulled out of their authors' back ends. It makes no sense to attribute a probability to events over which we have complete agency. Whatever probability you come up with merely reveals your lack of confidence that your fellow humans will do the right thing. Confidence in humanity is one thing Geoff and I disagree on. The other thing we disagree on is the primacy of intelligence. In the real world, intelligence is not synonymous with power and dominance. A virus, a poisonous snake, or a tiger can bring down a human. Lots of very stupid humans dominate much smarter ones (just look at the international political scene). Superintelligent AI will work for us and do our bidding, subject to safety guardrails. It will have no desire to take over. I have made detailed technical proposals for AI architectures that can be made safe by satisfying safety guardrails at inference time. I call it Objective-Driven AI Architecture.📊 11 ⚡
This report is categorized under AITOP's "Industry" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Yann LeCun, Geoffrey Hinton, AI safety, p(doom).
Source attribution: curated from Yann LeCun (@ylecun).
