标签： DeepSeek

让知识蒸馏便宜到可大规模运行

让知识蒸馏便宜到可大规模运行

Multiverse Computing 等研究团队提出两种系统级优化——离线缓存教师模型 Top-K logits，以及融合分块 KL 散度损失——把大模型知识蒸馏的显存峰值从约 250GB 降到约 128GB，让蒸馏可以在单张 H200 上跑通，大幅降低压缩大模型的门槛。

The only way I’ve ever gotten good at anything is by being consumed by it for a while. This has been a pattern in my life – meditation, standup comedy, family, work, LLMs. I’ll spend a few years going ridiculously dee…

The only way I’ve ever gotten good at anything is by being consumed by it for a while. This has been a pattern in my life - meditation, standup comedy, family, work, LLMs. I’ll spend a few years going ridiculously dee...

AI 开发者 Madhu Guru 提出了一种反共识的学习观：想在任何领域（包括大模型技术）做到真正擅长，不是靠每天平衡分配时间，而是用几年时间“被它吞噬”，直到知识内化为直觉。这为 AI 行业里普遍的“工具焦虑”提供了另一种选择。

Find someone who loves you as much as Fable loves putting all features behind flags as environment variables.. It got so bad that I eventually had to put “defaults matter, no hedging” in the Claude.md file. Lesson in…

Find someone who loves you as much as Fable loves putting all features behind flags as environment variables.. It got so bad that I eventually had to put “defaults matter, no hedging” in the Claude.md file. Lesson in...

开发者 Nikunj Kothari 在 X 上吐槽 AI 游戏框架 Fable 把所有功能都藏进环境变量开关里，导致他不得不把“默认值很重要，不要摇摆不定”写进 Claude.md 项目指令文件。这条看似吐槽的帖子，实际点中了 AI 辅助编程时代一个核心问题：当产品没有明确默认行为时，AI 生成的代码会越来…

@ChatGPT Trying to convert normies to go from the web to the desktop app to enjoy the power of agents and even with my manual onboarding it’s confusing.

@ChatGPT Trying to convert normies to go from the web to the desktop app to enjoy the power of agents and even with my manual onboarding it's confusing.

OpenAI 旗下产品 ChatGPT/Codex 的“共享项目”功能被曝出现基础故障，同时有开发者反映，让普通用户从网页端迁移到桌面端以体验 Agent 功能，过程十分困难。两个问题叠加，暴露出 AI 产品在从“极客工具”走向“大众工具”时的真实短板。