CHEN Dakeng

Postdoc @CUHKofficial | PhD @CUHKofficial | Computational Social Science, Comparative Politics, AI4SS |

[CHEN Dakeng @dakengdevinchen]

Does digital surveillance make citizens more compliant? We (with @jvzhan1) find yes, at first, but the effect fades within a couple of years, even as the state adds more cameras. Surveillance shapes behavior, but with a time limit. Read more @Journal_PoPs: https://t.co/YMncdyBj8w

[CHEN Dakeng @dakengdevinchen]

We are hiring a full-time RA with:
✅ Deep academic passion
✅ Strong computational skills
✅ Good understanding of Chinese politics
Join us at CUHK!
Apply here 👇 https://t.co/UBJEGgr7O3

[Anthropic @AnthropicAI]

New Anthropic Fellows Research: a new method for surfacing behavioral differences between AI models.

We apply the “diff” principle from software development to compare open-weight AI models and identify features unique to each.

Read more: https://t.co/VAsu2PSgCX

[AJPS @AJPS_Editor]

Using large language models to analyze political texts through natural language understanding by Kenneth Benoit, Scott De Marchi, Conor Laver, Michael Laver, and Jinshuai Ma is now available in Early View. @kenbenoit https://t.co/8tE6nqVci9

[Hugging Models @HuggingModels]

Meet a game-changer in AI reasoning: Qwen3.5-35B-A3B. This model isn't just another text generator. It's been distilled from top-tier reasoning models to think step-by-step like the best. Perfect for complex problem-solving!

[向阳乔木 @vista8]

把卡比大神的Opencli，封装成Skill，方便Claude Code和龙虾调用。

安装：npx skills add joeseesun/opencli-skill

场景如下：
① 一句话查看HackerNews的热帖，写入Obsidian或飞书。
② 一句话查看知乎私信，一句话查看微博热搜，发微博等。

最酷是Opencli支持自迭代生成任意网站Cli。

比如让AI给Producthunt写了个Cli，Github开源见评论

[CHEN Dakeng @dakengdevinchen]

Built a Claude Code skill for R/Python data science — inspired by
@posit_pbc
's AI prompts in Positron Assistant. Makes Claude think and behave like a tidy person: tidyverse, ggplot2, polars, notebook-aware protocols, etc. https://t.co/lv0EXpiDu1

[OpenCode @opencode]

OpenCode Go is a low cost ($10/month) subscription designed to bring agentic coding to programmers around the world

it provides generous limits and reliable access to the most capable open source models

run /connect and select OpenCode Go to get started

[宝玉 @dotey]

斯坦福大学去年秋季开了一门新课 CS146S，叫"现代软件开发者"，由知名风投 a16z 的合伙人 Martin Casado 主讲，核心主题就一个：AI 时代怎么写代码。

课程的基本观点是，软件开发已经从"从零手写代码"变成了"规划→AI生成→修改→循环"的新工作流。课上会教学生怎么用 LLM 高效提示、搭建 AI 编程代理、用 Claude Code 等工具做开发，还涉及 AI 代码审查、安全漏洞检测、自动化测试等实战环节。

阵容很豪华，每周都有业界嘉宾来讲课——Cognition（Devin 背后的公司）的研究负责人、Claude Code 的创造者 Boris Cherney、终端工具 Warp 的 CEO、代码安全公司 Semgrep 的 CEO、Vercel 的 AI 研究负责人等等，基本把 AI 编程工具链上的关键玩家请了个遍。

课程为期一个学期，3 学分，要求有编程基础，每周大概 10-12 小时的投入。最终项目是做一个展示现代开发实践的作品。

https://t.co/6ecEpIQ7A2

[CHEN Dakeng @dakengdevinchen]

@imocone @mranti 我也想问… 价格应该是一模一样的

[CHEN Dakeng @dakengdevinchen]

@mranti 看来国际版贵有贵的道理，但是我测试网页的国际版kimi，censorship和国内版一样敏感

[Yiqing Xu @xuyiqing]

1/ 🐎 Our gift for the Year of the Horse:

An AI-assisted workflow that scales reproducibility in empirical research. w/ @YangYang_Leo

Paper: https://t.co/fslLN0zTQO

[CHEN Dakeng @dakengdevinchen]

@Agimon_AI That’s right, and data scientists have to deal with “‘mixed” language context, switching between python, r, (maybe Julia) and stata for different tasks.

[CHEN Dakeng @dakengdevinchen]

Got early API access to Kimi k2.5 (https://t.co/gWYwFO3p9V) and spent the past few days testing it extensively in Claude Code for data science and coding tasks. Compared it against GLM-4.7, Claude Sonnet 4.5, and Gemini 3 Pro. Here are my findings (1/6)

[CHEN Dakeng @dakengdevinchen]

The official platform now supports multi-sub-agent coordination, but I couldn't test this with API access.
Among China's three major coding models (MiniMax, GLM, Kimi), I use GLM most for volume, but if I had to pick one top performer, it would be Kimi k2.5.
(6/6)

[CHEN Dakeng @dakengdevinchen]

Task consistency is robust. For automation tasks (outputting series of MD files with specific format/naming), GLM-4.7 tends to forget instructions in long contexts. k2.5 maintains consistency throughout.
Others noted major aesthetic improvements in code generation as well.
(5/6)

[CHEN Dakeng @dakengdevinchen]

Multi-program-language support has improved. I primarily work with R for data science, and previous Kimi versions would jump to Python (sometimes hallucinating non-existent R packages). k2.5 stays in R reliably, better than GLM-4.7, approaching Gemini, though behind Claude.
(4/6)

[CHEN Dakeng @dakengdevinchen]

Exceptional stamina for long-running tasks. k2.5 easily handled 15+ minute sessions, executing autonomously based on the plan. Outperforms both Gemini 3 Pro and GLM-4.7 here, though not at Claude subscription levels. Token generation is noticeably quicker than GLM-4.7.
(3/6)

[CHEN Dakeng @dakengdevinchen]

I am developing a keyboard/musical note matching game for my daughter. Native multimodal capability is nice. Unlike GLM-4.7 where you need to configure vision MCP, k2.5 handles images out of the box. Its ability to understand coding intent from screenshots is impressive. (2/6)