LeCun has "zero concerns" about AI wiping out humanity, recent "rogue" incidents
Fears that advanced AI could wipe out humanity are challenged by critics who argue current systems are just tools whose real dangers stem from human misuse, bad incentives, and poor engineering rather than autonomous malevolence. Others counter that superhuman models, tightly integrated with critical infrastructure and capable of large‑scale persuasion or bioengineering, could plausibly cause catastrophic or even existential harm, especially given weak governance and opaque corporate control. The exchange broadens into skepticism over billion‑dollar “AI safety” efforts, concern about regulatory capture and social decay, and a call to focus on concrete risks like disinformation, job displacement, and fragile socio‑technical systems.
Extent of AI Extinction Risk
- Many commenters agree with the “near zero” extinction-risk view: current systems are text autocompleters that can’t act without brittle, human-written harnesses; you can “pull the plug.”
- Some argue it’s almost impossible to literally wipe out all humans; even nukes, pandemics, and climate change leave survivors. They demand at least one concrete, detailed mechanism before treating “human extinction” as serious.
- Others counter that you don’t need Skynet: highly capable AI could design novel pathogens, orchestrate multi-vector attacks (bio, infrastructure, persuasion), or exploit digitized control systems. Even “only” billions of deaths or civilizational collapse is bad enough.
- Several note that humans are already cognitively hackable via engagement algorithms, propaganda, and social media; smarter systems could scale this dramatically.
Near-Term and Non-Existential Risks
- Many think the real issues are social and economic: job loss, inequality, political manipulation, targeted persuasion in elections, education collapse, and general “brainrot” as people offload thinking and socializing to bots.
- Government/military misuse, autonomous weapons, state surveillance, and small-scale bioterror are repeatedly cited as more concrete threats than rogue superintelligence.
Human vs Model Responsibility
- Widespread view: recent “rogue agent” hacks resulted from bad sandboxing, mis-specified rewards, and management recklessness, not malign intent from models.
- LLMs are described as symbol-in/symbol-out functions; real-world harm comes from tools they are given (file systems, shells, networks).
- Others push back that once you connect models to powerful toolchains and infrastructure, “it’s just the harness” is cold comfort; controlling such systems at scale is nontrivial.
AI Safety Movement & Industry Incentives
- Long, detailed criticism of the AI safety / effective altruism ecosystem: billions in funding, dense cross-links with frontier labs, claims of regulatory capture, self-dealing, and using extinction rhetoric to justify monopolies and high valuations.
- Defenders argue this is sincere risk-mitigation and fundamental theory work; money flow alone doesn’t imply bad faith.
Capabilities, Architectures, and Uncertainty
- Debate over whether current models truly have spatial reasoning or common-sense physics versus just pattern-matching benchmarks.
- Some think current LLM-based approaches are a dead end for AGI; others see emergent abilities and huge remaining headroom.
- Overall, commenters converge more on “we’re highly uncertain” than on any single forecast.