推荐mzl · 2025-12
“I think the measures of intelligence per watt and intelligence per joule from Intelligence per Watt is very interesting.”
觉得其每瓦智能和每焦耳智能的度量很有意思
提出每瓦智能(IPW)指标,评估本地小模型能分担多少云端请求
论文提出「每瓦智能」(IPW,即每单位功耗的任务准确率)作为衡量本地推理能力与效率的统一指标,并评测了 20 多个本地语言模型、8 种加速器和 100 万条真实的单轮对话与推理查询。结果显示本地模型能成功回答 88.7% 的查询;2023 至 2025 年 IPW 提升 5.3 倍,可由本地处理的查询比例从 23.2% 升至 71.3%;本地加速器的 IPW 比运行相同模型的云端加速器至少低 1.4 倍,仍有很大优化空间。
评论者喜欢它提出的「每瓦智能、每焦耳智能」框架,认为这是思考 AI 基础设施长期可持续性的有用指标,也有人让大家直接读它的原始数据而不是转述文章。
“I think the measures of intelligence per watt and intelligence per joule from Intelligence per Watt is very interesting.”
“I like the intelligence per watt and intelligence per joule framing in Intelligence per Watt It feels like a very useful measure for thinking about long-term sustainable variants of AI build-outs.”
“Read the paper the author talks about yourself instead”