
Exploring Gen AI && Builder + Creator👨💻|vibe coding🧲 K8E Sandbox Creator - https://teapaw.org/t.co/MlGFUg6NEQ Telegram: @xiaods
beijing
Joined September 2008
- Tweets38.4K
- Following1.4K
- Followers4.4K
- Likes1.2K
Tommy Xiao retweeted
亚马逊版 WorkBuddy ,免费用Claude Opus 5.5!
AWS-Quick是亚马逊旗下的办公AI,类似于中国的WordBuddy
背后的模型是 Claude Opus 5.5满血版
根据AI输出的时间来计算的,不是按Token,10小时输出,实际可以用挺久
客户端只有windows 和 mac,免费版每月 10 小时,每月1号重置
重点:有亚马逊账户的授权登录就能用,多个亚马逊账户可以切换用
URL:aws.amazon.com/cn/quick
上周,NSH 有 67 位贡献者,合并了 510 个 PR。
我也发现,越来越多 Longbridge 的同事开始参与 NSH 的改进。
之前,参与 GPUI Kit 开发的团队同事大约只有 5–6 人。除了 Longbridge Pro,公司内部使用 GPUI 的项目也只有 4–5 个。
而现在,很多同事已经在用 NSH 开发各种各样的 App,也在使用过程中不断为它贡献改进。
因为 NSH 真正帮到了大家,大家也能很轻松地参与进来。哪怕只是改进一处文档,也同样有价值。
github.com/longbridge/nsh/in…
Tommy Xiao retweeted
正式向大家汇报一下 LoopX 从 1.x 到 v2.4 的大版本跨越——自 1.0 以来全仓走过 4,100+ 个合入 PR,10 月切入 2.0 架构后一直埋头打磨没正式官宣。今天借社区 @lemomo_ai 做的这支 3D 无穷轨道短片,把从 1.x 到 2.4 的三件核心进化一起交个卷:
1. 从 1.x 自由脚本/笔记到 2.0 结构化通用任务语义:不再靠易漂移的 Markdown 或纯机械队列硬撑,抽象出领域无关的 Goal / Acceptance / Todo / Claim & Lease / Gate / Evidence 状态协议,落进 TypeScript + SQLite / 自托管 PostgreSQL 原子事务底座(含默认拒绝 ACL),跨 Codex、Claude Code、Gemini、DeepSeek 中立混编。
2. 同周期 3.0% 首读差分与 18h 真长程标尺:把同压缩周期续跑的首读上下文压到原包 3.0%(失配自动回退完整快照);针对 10h 后模型爱改局部测试“刷分”的通病,引入同父反事实标尺(Yardstick)与证据驱动步幅(Stride v2),在 LHTB(46 题 +17.3%)和 EdgeBench 18h(ANN 7,159 QPS 等)跑出扎实对照,并在 HuggingFace 开源了 68 个检查点。
3. 走向 v2.4 与 22 位社区开发者共创:我现在家里一台 Mac 常驻挂着十来个 7×24h agent(主控轨迹刚过 11.15 GB / 3,598 turns),配合社区近期共建的一句话建 Goal、LooChat 多 CLI 工作台与飞书管家门禁,出门手机点一下就能放行。日常英文进展与每周「开源场景 Loop」也会同步更在官号 @LoopXAgent 👇
Tommy Xiao retweeted
Harnesses used to be a crutch for weak LLMs. In 2026 and onward, with stronger models, harnesses now offer a mechanism for composing an LLM’s existing and new capabilities to generalize well beyond their training environments.
In our new paper, LM Harnesses are Compositional Generalizers, we explore how various harness abstractions enable LM systems to generalize over task length, task strategies, recursive task depth, and more through composing existing LLM capabilities. These abstractions provide a means for more data-efficient training, especially over long tasks, by keeping trajectories locally in-distribution (LID) — i.e., shaping all individual LLM calls to be in-distribution.
For tasks that take a very long time to roll out, this implies that we can even perform surrogate harness training, meaning we train an individual LLM call instead of doing the very long harness rollout, using purely synthetic tasks, with fake tool calls + sub-agents!
A common idea floating around is that LLMs will absorb harnesses and make them obsolete. We disagree, and argue that we haven’t sufficiently explored what harnesses can provide.
The full paper, logs, codebase, and data are available: alphaxiv.org/abs/2611.harnes…
w/ @SOURADIPCHAKR18 @lateinteraction
Tommy Xiao retweeted
In the 1960s, computing was centralized until the PC put power into everyone's hands. Now Al is trained in datacenters and rented back to you token by token.
Today we open source Linger (@HeyLingerAl), your personal agent living on your own computer.
We believe intelligence should be distributed. Your context, your data, and the intelligence that powers them should be owned entirely by you.
Github: github.com/simplexai-labs/li… Homepage: heylinger.com
对 Anthropic 股票感兴趣的,还是要看看它的发家史,然后在决策到底投不投。youtube.com/watch?v=sYdN4b14…
1
257
Tommy Xiao retweeted
I wrote the first line of code Friday, had most customers migrated over by Sunday.
Ask your agent to try it out: github.com/comp-work/gke-san…
网易的UU远程要收费了,吃相难看,给大家推荐一个开源神级远控 Airhand github.com/sqlpxc/airhand
实测非常好用,可以自己建中转服务器,网络好的时候是打洞直连的,有很多特性,QUIC 传输 · 硬件编解码 · 最高 120 fps,强推!!!
Tommy Xiao retweeted
Karpathy 开炮了。
这位前 OpenAI 联合创始人、前特斯拉 AI 负责人,在一段 16 分钟的对谈里,把当下 Agent 热潮的底裤扒了个干净。
他说得很直接:Agent 的本质就是大规模蒸馏。你花大价钱跑的那个大模型,99.99% 的容量都浪费在了它压根用不上的数据上。听起来刺耳,但你仔细想想,事实就是这样。
然后他给出了一条真正能跑通的技术路线:先用大模型把能力蒸馏到小模型里,再给小模型配上正确的工具,最后用一个闭环把整个系统串起来。小模型、正确的工具、一个闭环,三件事,够了。
这里面最关键的一环,是闭环。
小模型本身并不可怕,可怕的是它一旦接入闭环,就能自我校正、自我改进,越跑越强。模型的参数量可以小,但只要反馈回路搭对了,它就能长出远超参数规模的实战能力。
所以别再为更大的模型交税了。真正让一个 Agent 变得危险的,从来都不是参数量,是闭环。
这 16 分钟的信息密度,比今年市面上任何一门 500 美元的 AI 课都高。看完之后,再去读那篇 Harness, Loops, Graphs,你就会明白:闭环就是那副马具,套上了,小模型才能真正跑起来。
太震撼了。晚上 ANT 要搞起来
Anthropicr招股书中一些震惊的数字:
1、现金王:帐上有1319亿美元现金及短期投资,仅次于伯克希尔和谷歌。
2、赚钱狂:狂烧了3年钱,其实累计消耗是100亿美元,但至今年9月,现金流为净流入65.7亿美元。
3、投入猛:未来的算力帐单签到了2045年,各种承诺加起来9009亿美元。公司Q2起1美元算力能跑出2.08美元收入
4、B端强:世界500强中年付费10万美元以上客户占到一半以上,年付100万美元的11.9万家,而一年前仅为950多家;年府1000万美元的超级大客户250家。
八成营收来自API付费(真是B端用户最豪横)——而世界500强中那超一半的客户,其在AI上的开销目前仅占各家企业总开销的0.1%。
5、人效高:公司5300名员工,按Q3营收,人均年产值为1200万美元。
6、编程多:使用CLAUDE聊天的用户消费较少,仅为使用CLAUDE CODE的高频用户消费的1/20,大客户都是通过API或Claude code工作。
目前看确实很强很强,不过Google我相信一定会赶上来的。
531
Tommy Xiao retweeted
网易UU远程收费啦,计划于 2026 年 12 月 1 日推出高阶增值服务。在此日期前注册成功的用户,可领取年卡五折专属优惠
VIP30(连续包月24)、SVIP98(连续包月88),折合最低9.9/月,更新版本后免费版将只支持1080p30帧
免费的午餐该醒了,毕竟网易也不是慈善家,不过P2P打洞连的也要限画质吗?
Tommy Xiao retweeted
👏 Congratulations to the @baseten and @sgl_project teams on pushing low-latency inference forward with NVIDIA Vera Rubin.
Today we’re sharing our latest results in pushing low-latency performance for Deepseek v4.1 Flash. It’s running at ~2,500 tok/s on Vera Rubin, batch size 1, TP4, 1.9ms per step
It’s up to 1.7x faster than our GB300 numbers, and we’re working with @sgl_project and @radixark to get it all into OSS.
Here’s what @TriDao2 and @stel4th shared at the SGLang summit last week:
Tommy Xiao retweeted
These are literally the best visual resources for learning how Transformers, LLMs, embeddings, diffusion, and inference work ↓
1/ Transformer Explainer - Watch GPT process text through embeddings, attention, MLPs and next-token prediction.
poloclub.github.io/transform…
2/ Brendan Bycroft’s LLM Visualization - Explore an LLM from architecture down to tensors and operations.
bbycroft.net/llm
3/ 3Blue1Brown - Visual intuition for linear algebra, neural networks, backprop, attention and Transformers.
3blue1brown.com/topics/neura…
4/ The Illustrated Transformer - One of the clearest visual explanations of embeddings, Q/K/V and attention.
jalammar.github.io/illustrat…
5/ TensorFlow Playground - Watch neural networks learn as you change layers, activations, features and learning rate.
playground.tensorflow.org
6/ Google PAIR AI Explorables - Interactive explainers on LLMs, generalization, interpretability and model behavior.
playground.tensorflow.org
7/ Distill - Exceptional visual essays on t-SNE, feature visualization, GNNs and interpretability.
distill.pub
8/ Abhik Sarkar’s Transformer Visualizations - RoPE, KV cache, FlashAttention, MQA, GQA and more.
playground.tensorflow.org
9/ Visual Guide to Attention Variants - MHA, MQA, GQA and MLA visually compared.
magazine.sebastianraschka.co…
10/ Visual Guide to Mixture of Experts - Routing, experts, sparse activation and load balancing.
newsletter.maartengrootendor…
11/ Modular LLM Inference Handbook - Prefill, decode, KV cache, batching, quantization and speculative decoding.
handbook.modular.com/llm-inf…
12/ Apple Embedding Atlas - Explore clusters, neighborhoods and outliers in large embedding spaces.
apple.github.io/embedding-at…
13/ Diffusion Explainer - Follow Stable Diffusion step by step.
poloclub.github.io/diffusion…
14/ Neuronpedia - Explore features, activations, SAE latents and attribution graphs inside real models.
neuronpedia.org
15/ Seeing Theory - Probability, Bayes, distributions, regression and inference made interactive.
seeing-theory.brown.edu
16/ CNN Explainer - Visualize convolutions, feature maps, activations and pooling.
poloclub.github.io/cnn-expla…
17/ GAN Lab - Train a GAN in your browser and watch its generated distribution evolve.
poloclub.github.io/ganlab/
Save this. There’s a serious AI curriculum hiding inside these links.
A free online book: LLM Inference Handbook by Sherlock Xu (2026).
This book is a practical technical reference for engineers deploying, scaling, and operating language models in production.
It brings together the concepts and decisions involved in turning a trained model into a responsive, cost-aware service, from model behavior and hardware selection to application interfaces and ongoing operations.
Readers can use it as a guided progression or as a lookup resource for specific engineering questions.
Read the book online on @ChapterPal: chapterpal.com/book/3fef12a8…
Tommy Xiao retweeted
Introducing Qwen3.8-27B-pi-adapted 🤗
We adapted Qwen's model specifically for the pi coding agent using reinforcement learning and a LoRA adapter.
The results, measured as pass@1:
🚀 SWE-bench Pro (Hard 51): 29.4% → 62.7%
💻 Terminal-Bench 2.1: 39.5% → 55.8%
⚡ 55–62% fewer tool calls per task
This is the power of open models: you can adapt them to YOUR tools and YOUR use cases. Let’s build! 💛
Open weights: huggingface.co/harness-hub/Q…
Methodology: huggingface.co/blog/harness-…
































