LogoThread Easy
  • Explorar
  • Criar thread
LogoThread Easy

Seu parceiro completo para threads do Twitter

© 2025 Thread Easy All Rights Reserved.

Explorar

Newest first — browse tweet threads

Keep on to blur preview images; turn off to show them clearly

苹果新搞了一个RAG框架：ml-clara，解决长上下文处理效率低下、检索与生成优化过程的分离问题 其核心思想是，不要把整段文本塞给大模型，而是把“检索”和“生成”全部压缩到同一个可微的连续向量空间里，统一训练、一次推理 以此解决，1 上下文越来越长计算量爆炸，2 检索器和生成器独立训练导致优化目标不一致，3 梯度断流的问题 在NQ、HotpotQA、MuSiQue、2Wiki上，不同压缩比4×/16×/32×均保持领先，压缩到32×时仍优于未压缩的纯检索基线 上下文长度最高可压32×–64×，同时保留了生成准确答案所需的基本信息 具体是，1、首先压缩预训练，把文档压成32～256维向量，保留QA/复述语义 2、然后指令微调，让压缩向量适配下游问答任务 3、再端到端联合训练，检索器+生成器一起优化 #RAG #mlclara

苹果新搞了一个RAG框架：ml-clara，解决长上下文处理效率低下、检索与生成优化过程的分离问题 其核心思想是，不要把整段文本塞给大模型，而是把“检索”和“生成”全部压缩到同一个可微的连续向量空间里，统一训练、一次推理 以此解决，1 上下文越来越长计算量爆炸，2 检索器和生成器独立训练导致优化目标不一致，3 梯度断流的问题 在NQ、HotpotQA、MuSiQue、2Wiki上，不同压缩比4×/16×/32×均保持领先，压缩到32×时仍优于未压缩的纯检索基线 上下文长度最高可压32×–64×，同时保留了生成准确答案所需的基本信息 具体是，1、首先压缩预训练，把文档压成32～256维向量，保留QA/复述语义 2、然后指令微调，让压缩向量适配下游问答任务 3、再端到端联合训练，检索器+生成器一起优化 #RAG #mlclara

github：https://t.co/vf7TFI1Wna

avatar for AIGCLINK
AIGCLINK
Fri Nov 28 02:57:51
In London there's no phone signal on the Underground train. In Korea there's not only signal, but it's extremely fast!

In London there's no phone signal on the Underground train. In Korea there's not only signal, but it's extremely fast!

Photographer & software engineer, prev @Ghost. Loves building w/ Nodejs, React, Rails, Django! DM for collabs. ❤️ @xxxxjiwon

avatar for Ronald
Ronald
Fri Nov 28 02:57:46
RT @visegrad24: 🇪🇺 EU moves forward with the controversial “chat control” law EU member states have agreed on amendments to the so-called…

RT @visegrad24: 🇪🇺 EU moves forward with the controversial “chat control” law EU member states have agreed on amendments to the so-called…

Privacy Experience @PrivacyEthereum | Events @ethereum | OSS 🛠️ Indie Maker @eth_gastracker https://t.co/xNlTAwRRp1

avatar for Wesley— oss/acc
Wesley— oss/acc
Fri Nov 28 02:52:31
I'm a simple man, in school I was taught about «Mongol-Tatar Yoke», and nothing about China. Took me some time to discover that there was such as thing as «Yuan Dynasty» claiming supremacy over the Horde. I think it's really hard to not see that early Yuan is not quite Chinese

I'm a simple man, in school I was taught about «Mongol-Tatar Yoke», and nothing about China. Took me some time to discover that there was such as thing as «Yuan Dynasty» claiming supremacy over the Horde. I think it's really hard to not see that early Yuan is not quite Chinese

We're in a race. It's not USA vs China but humans and AGIs vs ape power centralization. @deepseek_ai stan #1, 2023–Deep Time «C’est la guerre.» ®1

avatar for Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)
Fri Nov 28 02:51:01
Realistically I think the only way is to convince the EU member states to either 1) pressure the EU to redesign itself structually, and if not 2) exit the EU Once enough exit there will be restructuring of the EU or a whole new design or it disappears And the individual EU nations will just make deals with each other like they used to I'm a big believer in the concept of the EU but the current EU design especially with the unelected and authoritarian European Commission just doesn't work and it's not democratic And now it's seriously going into a direction of totalitarianism with Chat Control Very scary

Realistically I think the only way is to convince the EU member states to either 1) pressure the EU to redesign itself structually, and if not 2) exit the EU Once enough exit there will be restructuring of the EU or a whole new design or it disappears And the individual EU nations will just make deals with each other like they used to I'm a big believer in the concept of the EU but the current EU design especially with the unelected and authoritarian European Commission just doesn't work and it's not democratic And now it's seriously going into a direction of totalitarianism with Chat Control Very scary

🇪🇺https://t.co/NdorAWrhrB @euacc 📸https://t.co/lAyoqmT9Hv $125K/m 🏡https://t.co/1oqUgfDEsx $40K/m 🛰https://t.co/ZHSvI2wRou $38K/m 🌍https://t.co/UXK5AFra0o $16K/m 👙https://t.co/RyXpqGvdBB $14K/m 💾https://t.co/M1hEUBB6da $6K/m

avatar for @levelsio
@levelsio
Fri Nov 28 02:50:40
谷歌又发福利了！没地方玩 Nano Banana Pro 的可以试试 刚看到群里朋友分享的 Google  Gemini Enterprise 可以白嫖一个月，不需要绑卡 能免费用谷歌所有的模型包括 Gemini 3 Pro 以及 Nano Banana Pro 进到链接里面点击下面商务版的“开启免费试用”就行

谷歌又发福利了！没地方玩 Nano Banana Pro 的可以试试 刚看到群里朋友分享的 Google  Gemini Enterprise 可以白嫖一个月，不需要绑卡 能免费用谷歌所有的模型包括 Gemini 3 Pro 以及 Nano Banana Pro 进到链接里面点击下面商务版的“开启免费试用”就行

这里领取：https://t.co/clUKMI1hfY

avatar for 歸藏(guizang.ai)
歸藏(guizang.ai)
Fri Nov 28 02:47:19
  • Previous
  • 1
  • More pages
  • 2154
  • 2155
  • 2156
  • More pages
  • 5634
  • Next