Miceren
79 posts




Excited to share our latest work: CHESS♟️ (Context-aware Hierarchical Efficient Semantic Selection)! Long-context LLM inference is severely bottlenecked by massive KV caches. Prior pruning methods often sacrifice generation quality or fail to deliver real speedups. We solved this with a novel algorithm-system co-design: 🔥 Key Highlights: • 1% KV Cache: Achieves Full-KV generation quality by dynamically selecting semantically coherent tokens step-by-step. • 4.56x Throughput: Translates algorithmic sparsity into massive wall-clock acceleration via a custom zero-copy inference engine built atop PagedAttention. No more trading off between accuracy and speed in long-context tasks! 📖 Paper: arxiv.org/abs/2602.20732 💻 Code: anonymous.4open.science/r/CHESS-9958 Huge thanks to my brilliant co-authors Guozhong Li, Chenxi Liu, and Prof. Panos Kalnis. Would love to hear your feedback! 👇 #LLMs #MachineLearning #GenerativeAI #SystemOptimization #KAUST

Google 最应该溢价五倍收购 cursor agent editor 也不是谁都能做的,这个编辑错误从反重力诞生之初就没解决过。


手动刷推太耗时?试试用 Grok 的推文自动监控与筛选功能,可节省大量时间! 它可以定期扫描指定账号的 24 小时内新推文,自动过滤掉闲聊、吐槽、生活琐事,仅提取并总结与 Crypto、AI、Web3、DeFi、编程、安全 等相关的重点内容,帮助你快速跟进热点。 使用步骤: 1/ 打开 Grok,点击左上角个人资料图标 → 展开侧边栏 → 进入 Tasks 页面。 2/ 点击 "+" 新建任务,在提示词输入框粘贴以下模版: 获取这些账号的24小时内新推文,过滤掉闲聊、吐槽、个人生活,仅保留与Crypto、AI、Web3、DeFi、编程、安全相关的有价值信息,无相关内容的账号不列出,需基于帖子核心观点简洁总结: @star_okx @cz_binance 具体账号可按需替换~ 3/ 设置运行频率(推荐每天运行) → 开启通知(可推送至 X 或邮箱)→ 保存并激活任务。 4/ 完成后,Grok 会自动执行并推送结果,你只需查看总结即可掌握重点,无需逐条翻阅时间线。 目前这套提示词特别适合 DeFi、AI、Web3 和投资,能显著提升信息筛选效率。 大家有更具体的需求,可尝试修改提示词!

分享一个冷知识,在美国、中国等所有主流国家,根据法律规定,你存入银行的钱,已经不属于你自己而是银行,银行并不是在帮你保管这笔钱,而是只给了你对应的债权。

美国的流浪汉在街头被自由冻死。 中国的流浪被政府“强制”学技能,最后成功翻盘。 这就是中美两国体制的区别。

AAVE is generating over 85% of the total revenue in the on-chain lending sector. Insane how much market share a single platform can capture.

found a pretty major data bug it turns out almost every major dashboard has been double-counting Polymarket volume (not related to wash trading) this is because Polymarket's onchain data contains redundant representations of each trade. receipts ⬇️⬇️











