“AI 教父”辛顿提议 AI 行业建立 FDA 式审批机制:模型推出前需通过安全审核Godfather of AI" Hinton Proposes FDA-Style Approval Mechanism for the AI Industry: Models Must Pass Safety Review Before Launch


AI free token/AI免费积分

IT之家 10 月 7 日消息,杰弗里 · 辛顿(Geoffrey Hinton)希望 AI 企业跨过一道制药行业早已熟悉的监管门槛。

“AI 教父”辛顿提议 AI 行业建立 FDA 式审批机制:模型推出前需通过安全审核

在当地时间周二播出的《Smart Girl Dumb Questions》播客节目当中,这位被称作“AI 教父”的诺贝尔奖得主、计算机科学家提出建议:AI 公司在机型正式发布之前,必须先向监管机构证明自身机型具备安全性。

“你不能直接研制一款新药就把它推向市面。”辛顿对播客主持人纳伊玛 · 拉扎(Nayeema Raza)说道,“你必须说服美国食品药品监督管理局(FDA)。而为了完成这件事,你需要开展大量工作,相关投入大约要十亿美元级别。”

“对于人工智能而言,至少也应当设立这样的要求。”辛顿补充道。

辛顿针对 AI 安全争议提出这套应对思路的背景,一维度是各界不断浮现关于 AI 带来灾难性后果的预判;另一维度,OpenAI 在测试之后,出于对模型服从使用者指令、恪守权限边界的担忧,已经搁置了自家最新模型。

IT之家注意到,OpenAI 安全平台负责人萨奇 · 贾恩(Saachi Jain)此前接受《商业内幕》采访时表示,GPT‑6.1 Astra 模型在持续处理高难度任务维度有所进步,但在权限管控以及行为日志反馈维度仍存在短板。

辛顿提出,人工智能正在越来越多地达成自我迭代优化,这既可能加速能力提高,同样也会放大潜在风险。

“指数级的增长速度是颇为惊人的。”辛顿说,“笔者已经开始看到递归式自我改进现象:人们正在以多种多样的方式,借助 AI 本身去改良 AI。”

“当然,这仅仅属于我的推测。”辛顿补充说,“但我估计,距离局面远比现在糟糕,大概也就只剩一两年时间了。”

本文来源:https://www.ithome.com/1/010/187.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

IT Home, October 7 — Geoffrey Hinton hopes AI companies will meet a regulatory threshold already familiar to the pharmaceutical industry.

Godfather of AI" Hinton Proposes FDA-Style Approval Mechanism for the AI Industry: Models Must Pass Safety Review Before Launch

In an episode of the podcast Smart Girl Dumb Questions aired Tuesday local time, the Nobel laureate and computer scientist dubbed the "Godfather of AI" proposed that AI companies should be required to prove to regulators that their models are safe before official release.

"You can't just develop a new drug and throw it on the market," Hinton told podcast host Nayeema Raza. "You have to convince the U.S. Food and Drug Administration (FDA). And to do that, you need to invest significant effort—on the order of billions of dollars."

"The same requirement should apply to artificial intelligence," Hinton added.

Hinton put forward this regulatory framework against the backdrop of growing concerns about AI's potentially catastrophic consequences, as well as OpenAI's decision to shelve its latest model after testing raised worries about the model's obedience to user instructions and adherence to boundary constraints.

IT Home notes that Saachi Jain, head of OpenAI's safety platform, previously told Business Insider that while the GPT-6.1 Astra model has improved in sustained handling of complex tasks, it still falls short in permission controls and behavioral logging feedback.

Hinton pointed out that AI is increasingly capable of self-iteration and optimization, which could both accelerate capability gains and amplify potential risks.

"The exponential pace of progress is quite astonishing," Hinton said. "I'm already seeing recursive self-improvement: people are using AI itself to improve AI in a wide variety of ways."

"Of course, this is only my speculation," Hinton added. "But I estimate we're only about one to two years away from a situation that is far worse than what we see today."}

Source: https://www.ithome.com/1/010/187.htm If you have any copyright concerns, please contact us at 13951431913#139.com or submit a comment on the right.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry