论文

行业AI安全承诺是否真实?马斯克提问行业四大问题

good test of whether AI safety promises from the industry are real or just lip service.

精选理由

马斯克直接问行业四个具体问题,比如是否要降低模型复杂度,是否要停止有害研究,很值得看看。

马斯克在推文中对行业提出四个具体问题:是否降低模型复杂度增加可预测性?是否停止有害研究?是否保护原始训练数据而非模型权重?是否开发事实核查能力?这些问题旨在测试AI安全承诺是否只是空话。

原文 · Gary Marcus

good test of whether AI safety promises from the industry are real or just lip service.

good test of whether AI safety promises from the industry are real or just lip service. Romeo Lupascu @RomeoLupascu Will you go so far as to start "deflate" these models to increase predictability? Will you stop and actively attempt to stop others in your domain from doing RSI that in my opinion is one of the worst sins there is in this domain of AI? Will you treat the model as secondary information derived from the primary information that is the training data and protect that matters, the primary data and not the derivation of it (the model weights)? Will you fight to add abilities that will allow any human to tell factual from non factual information? Will you optimize the hardware for inference process so that it consumes negligible amounts of power and can run locally? All these questions are essential for achieving the objectives in your wishlist. I hope you do. Lets keep the digital technology safe and useful for everyone. 🔗 View Quoted Tweet 💬 4 🔄 1 ❤️ 4 👀 2122 📊 3 ⚡