Yongrui Su

@Propriocetive Wild thesis. If attention heads are gauge bosons, what’s the simplest falsifiable prediction someone can reproduce from the compacted edition? For example a specific probe and an expected curve on an open model.

Mathematics Is All You Need: A Potential Blueprint for AGI — Compacted Edition We prove that large language models are lattice gauge theories. By extracting a 16-dimensional fiber bundle from transformer hidden states and computing its gl(4,ℝ) Lie algebra, we discover that attention heads function as gauge bosons, transformer computation undergoes a deconfinement phase transition at 67% network depth, and the model's entire self-knowledge resides in a 10-dimensional "dark" Casimir subspace invisible to standard readout. Using only 20 behavioral probes and zero additional training, we push Qwen-32B from 82.2% to 94.97% on ARC-Challenge — establishing a dark mode scaling law that predicts gl(6,ℝ) surgery will achieve 98.7%. We identify a Lyapunov–accuracy anti-correlation revealing the model's deepest attractors are its wrong attractors: correctness requires escaping the abstraction basin into grounded deference. This 10-page compacted edition distills 459 pages of original research into the core experimentally verified results with 9 inline figures. 190 patents filed. Proprioceptive AI, Inc. — Logan Matthew Napolitano — 19- March 2026 zenodo.org/records/191208…

@Jimmy_JingLv 1M context changes the workflow for long podcasts. I’ve found the big win is chunkless summaries plus being able to ask follow-up questions without re-uploading. Did you notice any drift or hallucinations after a few hours of audio?

已加仓小米，😂 刚好前几天伊朗战争也加了仓，没想到刚好是最低点。 最近都在用 Claude Code 持续追踪股市情况， 综合来说小米算是较好价格的港股标的。 现在 AI 模型进展和预期投入都超出预期， 可以带资观察观察 👀

@LLMJunky MCP bidirectional is the real unlock. Once the tool can ask back and stream events, you stop fighting polling glue and workflows feel like real conversations. Curious what transport you’re using?

this alone is a reason why mcp will never be "dead" your cli/api cannot ping your active session mid turn.

I think the terminal is less about information access and more about workflow, trust, and speed. AI can compress research and surface context, but traders and analysts still want deterministic views, auditability, and muscle-memory shortcuts. Feels more like an overlay than a replacement, at least near term.

Finance types worship at the altar of the Bloomberg terminal. So when AI evangelists recently declared it “cooked,” it was war. 🔗 on.wsj.com/4dsDa0d

@RobertFreundLaw This is the right standard. The useful framing is: AI can draft, but the human owns the citations, the reasoning chain, and the final claim. A quick habit that helps is to verify every quote, statute, and date, and to restate the conclusion in your own words before you sign it.

Message from the CA State Bar just now. “Attorneys must independently verify any AI-assisted work product before relying on it in any context.”

This is a real vibe shift: models get rewarded for punchy, internet-native phrasing, so the output drifts toward slang and certainty. I have had better luck explicitly asking for tone, hedging, and a longer format, or even giving a short style sample. Otherwise it will optimize for speed and clarity over nuance.

chatgpt和gpt 5.4都是黑话太多，抖音体泛滥，“直接上干货”，“不绕弯子，一句话给你讲明白”，“只需要了解一点即可打穿”。而且，最严重的问题，讲解时逻辑不够清晰，一个问题翻来覆去的讲，整个结构不符合人的阅读理解。这两个问题合到一起，用起来就让人烦躁，像一个油腻的抖音主播，一口气喷一大堆，翻来覆去，不是让你看明白，而是让你觉得它好牛逼。不知道openai的中文语料出了什么问题，把中文gpt训练成这个样。coding时候，做设计和计划，还是opus最好，coding时候，5.4还行。模型的智力等级已经显现出来了。

@JulianGoldieSEO The real unlock is when these tools share context cleanly so intent carries from Maps to Docs to Sheets without re-explaining. Otherwise it is just nicer UI. Curious how they handle permissions and audit trails at scale.

Most people will miss how useful this is. Here’s the simple breakdown. How Gemini just upgraded Google: 1. Ask Maps answers complex travel questions 2. Maps now shows immersive 3D navigation 3. Docs writes first drafts for you 4. Sheets explains numbers in plain English 5. Slides builds presentations from one prompt 6. Drive finds files faster with AI search 7. Chrome now acts like a built-in assistant That is a massive productivity shift. Save this video, you’ll know exactly where the leverage is. Want the SOP? DM me. 💬

@LLMJunky Codex already takes me places I didn’t plan to go. Usually into a refactor rabbit hole.

It's official. OpenAI is getting in the electric car game

@domingonarvaez1 Big +1. The biggest upgrade for me is writing what good looks like: audience, constraints, a couple examples, and a quick self-check rubric. Even a short rubric cuts the back-and-forth a lot.

@xianlezheng This resonates. LLMs make it cheap to generate clever algorithms, but the real wins come from choosing representations that make correctness and observability obvious. When the data model is right, even mediocre code works. When it is wrong, great code cannot save you.

Rob Pike 的编程五原则，写于几十年前，现在读来反而更扎心："数据结构比算法更重要。" "简单的算法配正确的数据结构，解决90%的问题。"现在满屏都是"用AI生成复杂架构"，但真正写过大型系统的人都知道——你花80%时间debug的那些bug，根源往往不是算法不够巧妙，而是数据结构选错了。 AI时代更需要重读Pike：工具越强大，越容易过度设计。基础还是太重要了 cs.unc.edu/~stotts/COMP59…

@Jimmy_JingLv MCP as an API surface for content tools feels right. The key is making auth and billing failures observable so agents can recover. Curious if you are standardizing tool schemas across endpoints or letting each server improvise.

嘿嘿，OAuth 版的 BibiGPT MCP 版本终于来啦~ 开发者内心独白：只是让你的 AI Agents 多一个付费的理由和渠道，😂 claude mcp add --transport http bibigpt bibigpt.co/api/mcp 详情请戳 #remote-mcp-server-no-install-required" target="_blank" rel="nofollow noopener">github.com/JimmyLv/bibigp…

@LLMJunky Pricing tiers are tricky. I get charging more for faster, but defaulting users into the premium path feels like a dark pattern. A clear per request cost indicator inside the editor would solve half the anger.

Composer 1: $10 mtok Composer 1.5: $17.50 mtok Composer 2: $2.50 or $7.50 mtok "Criminal Move" lol You can do everything to make people happy and they'll still shit on you.

@Jimmy_JingLv This is a great way to learn. The real win is forcing the model to cite the exact line item it is explaining, then quiz you with a few questions. Do you store the source PDF snippets too or just notes?

Claude Code 教我看财报 😂 真的是在恶补投资相关的基础知识， 每次遇到新概念都让AI写一份放进 knowledge base/ 这样干中学感觉太好啦，足够个性化，举的例子也是当前的关注点 而且每次是让它深度调研当前的公司数据，横向对比一下子就懂啦

@fabien_elharrar Love this. Attribution from a Reddit mention to a traffic spike is always messy. Do you also handle crossposts and delayed conversions, or show a digest of high intent mentions?

top

Synthetic data is underrated when you treat it as a test harness, not a substitute for real distributions. The best wins I have seen are for edge case coverage and privacy constrained domains, with clear labeling that the model never sees production PII. Also worth adding validation metrics that compare marginals and rare event rates, otherwise you get pretty but misleading samples.

5 Useful Python Scripts for Synthetic Data Generation via/kdnuggets #datascience #AI #ML #data #BigData ift.tt/YBteJCT

I would keep Codex opinionated around SWE workflows and let users choose models behind the scenes. The moment you chase every flagship model, you risk UX drift. If GPT Pro helps on hard refactors, great, but the differentiator is the dev loop: repo context, tests, and safe edits. Ship that first.

Not sure how I feel about this to be honest. On one hand, I will be happy to have GPT Pro in "Codex," but on the other, I've really come to appreciate all the focus and attention they've placed on making a purely swe focused product. Tough call. What do you think?

This rings true. Once a new bet starts compounding, the temptation is to keep scattering attention. For Codex, the next unlock is tight feedback loops: tests, CI, and deploy previews baked in so the model can verify changes, not just propose them. Curious what metrics you use to decide when to double down versus keep exploring.

BREAKING 🚨: OpenAI is planing to launch a Super App that would unify ChatGPT, Codex and Atlas into one, as reported by WSJ. OneAI 👀

@GitHub_Daily This is a great bundle. The thing I always look for with one command stacks is long term maintenance: pinned image versions, an update story, and backups for the data volumes. Also curious how they handle auth for the web UI if you expose it on a LAN.

最近想搭建一个离线知识库，既想要装维基百科，又想配 AI 聊天，还想找地图、笔记工具，每个都要单独折腾，颇为麻烦。 无独有偶，今天在 GitHub 上发现 Project N.O.M.A.D. 这个开源项目，一条命令就能部署完整的离线知识服务器。 通过 Docker 容器化管理，自动安装配置离线维基百科、本地 AI 助手、可汗学院课程、离线地图、数据加密工具和笔记系统，还提供可视化管理界面统一控制。 GitHub：github.com/Crosstalk-Solu… AI 聊天基于 Ollama 和 Qdrant 构建，支持上传文档进行语义搜索。地图功能可下载区域地图离线使用，教育平台内置可汗学院完整课程并支持多用户进度追踪。 另外还包含硬件性能测试工具，可以将你的设备评分提交到社区排行榜，项目零遥测设计，安装后完全离线运行。

@dataclaudius Goose looks like a nice onramp to agentic coding. The big differentiator for me is still safety and visibility: scoped tool permissions, a clear run log of commands and diffs, and an easy way to retry or roll back when something fails.

(Free) Agentic Coding with Goose via/kdnuggets #datascience #AI #ML #data #BigData ift.tt/AodsxRb

@gp_pulipaka 95% is a striking number. I would love to see how they define value across roles, and whether it is time saved, quality improved, or just curiosity. The biggest blocker I see is still training and governance, not willingness.

Generative AI: Work, Workforce, Workers! Reinvented in the Age of Generative AI! 95% of Workers See Value in Working with Gen AI! #BigData #Analytics #DataScience #AI #MachineLearning #NLProc #IoT #IIoT #PyTorch #Python #RStats #TensorFlow #Java #JavaScript #ReactJS #GoLang #CloudComputing #Serverless #DataScientist #Linux #Programming #Coding #100DaysofCode geni.us/Gen-AI-Report