Shom

language model | sequence modeling | education | HCI

[Zhiyuan1i @uniartisan]

轻舟已过万重山

[Shom @ShomLinEd]

@flyingpetal472 纯粹inference 效率吧

[Fanqing @FanqingMengAI]

深度学习 -> 基于梯度的下降
RSI -> 基于ICL的下降

[Shom @ShomLinEd]

@zhuzilinallen 分久必合，合久必分，stack的纵向整合

[Shom @ShomLinEd]

@SuJinYan123 有点东西，但ai slop太多了

[KC @ScarletKc]

之前和 Yifan 聊 Multi-Agent Harness，听到了一个很有意思的设计。

他想让多个 Coding Agent 基于同一份项目状态协作，但不依赖 Git Worktree 来隔离每个 Agent。

他的思路是把沙箱隔离粒度缩小到每次 Tool Call。

每次工具调用都有独立的临时 Workspace，执行完可以保存为 Commit，也可以直接丢弃。底层通过 Zero-copy fork、COW 和内容去重，让不同 Agent 共享已有文件状态。

我第一反应是：你干脆重新发明 Git 吧。

但看完他的 LayerFS 设计文档，发现还真有点东西。

现在常见的 Agent 沙箱主要按 Task 隔离，而他希望做到 Tool Call 级别，让每一步操作都能独立执行、追溯、回滚和分叉。

这样不仅能支持更细粒度的多 Agent 协作，还能让 Agent 从任意保留的中间状态继续探索其他方案。

更有意思的是，他说这个想法半年前就写进 README 了。

[Yifan Xu @yifanxu_ephai]

别说agent不是7x24小时执行任务，agent就算在执行任务也80%的时间都在思考。之前cloudflare出了篇blog分析token economics，就指出了这个问题。

如果只有在用到bash时候启动沙箱，像read，write，edit直接通过文件系统提供的api/sdk进行操作。agent估计对沙箱的真实需求只占任务时间的5%。

[Shom @ShomLinEd]

@ScarletKc 很厉害，非常agent native

[nathan chen @nathancgy4]

life update: I've recently left @Kimi_Moonshot!

The past year at Kimi has been the best of my life so far. I'm deeply grateful to have worked with and learned from such amazing and supportive colleagues.

Some of my favorite memories include (in chronological order): optimizing KDA's kernels and trying every design possible with @yzhang_cs; coming up with and refining Attention Residuals with @yzhang_cs @Jianlin_S; working on optimizers, which by far taught me the most about training dynamics (and where I pulled the most all-nighters); creating the cases for K3's release with @yzhang_cs @Xinyu2ML while watching the World Cup; and more, and more.

Quoting my earlier tweet about Kimi K3’s release: it is truly incredible to have built such an open (and strong) model with a group of genuine people.

My days at Kimi are memories I'll forever cherish... thanks for everything! (especially @Yulun_Du for believing in and hiring a 16yo back then :P).

More things to explore :) 🫡

[Shom @ShomLinEd]

https://t.co/WlyPCOouXG
despicable

[Nex @NexEcosystem]

⚡ Meet NexRT, the low-latency decode engine behind Nex-N2.5.
💡 Built for single-request inference on Nex-N2.5-Pro.
• 1,000+ tokens/s at 8K context on 8×H100 with DFlash*
• Standard decoding, MTP and DFlash
• Context parallelism for long contexts, up to 256K

SGLang handles prefill; NexRT handles decode. The two engines share routed MoE weights.
Custom CUDA kernels, full decode CUDA Graphs and direct GPU communication help reduce latency throughout the decode path.
Performance report and demo now available. Source code will follow.
*DFlash throughput assumes an average of 5.5 committed tokens per round; actual throughput depends on draft acceptance.

🌐 Explore Nex-N2.5 and NexRT: https://t.co/uAf9LLd1Zo
🔗 GitHub：https://t.co/IKMdoqx4JG

[Shom @ShomLinEd]

@yifanxu_ephai postgres和s3都要网络，不可能快吧🤔

[Shom @ShomLinEd]

@yifanxu_ephai profile过了吗？直接看CPU热点啊

[jianlin.su @Jianlin_S]

Slow is Fast: Is Muon the end of Shampoo?
https://t.co/9zo5vGKbcH

Orthogonal Slowest Descent — from a convex optimization convergence bound — washes away any Kronecker preconditioner's eigenvectors, leaving only eigenvalues to balance the gradient's spectrum — leads to Muon.

[J.L @Jo1uck]

Could this be a direction for future model architectures?

Introducing Memory Attention (MA).

https://t.co/5bDpP0uQPs

[Shom @ShomLinEd]

@flyingpetal472 何意味，hysparse出很早了，要说也是dpsk抄小米的

[Shom @ShomLinEd]

@GenAI_is_real 软件工程才开始啊，问题的scale上去不知道多少了

[Shom @ShomLinEd]

@JarrettYe i prefer scaling neural network to achieve better performance(applying FSRS-6 for my learning system tho)

[Shom @ShomLinEd]

@geniusvczh agent糊一把什么静态分析写不了🤪