DailyVOL 2026.09.18246 items

Recursive Self-Improvement and Infrastructure Advances Drive Domestic Large Model Evolution

Singapore·Friday, 18 September 2026·DAILY · 08:00 SGT

2 Sep 2026. Recursive Self-Improvement (RSI) and infrastructure optimization emerged as the dominant themes today, with domestic large models demonstrating significant self-improvement capabilities. Zhipu GLM revealed the first RSI practice for domestic models, where an Infra Agent driven by GLM-5.3 completed system design, debugging, and optimization. Meanwhile, Zai.org optimized its GLM-5.3-Flash model for production with a 3.2x speedup in two weeks. In product developments, OpenAI launched Astra for Law, Google Gemini added file APIs, and Anthropic merged Claude Cowork into its main interface. Industry-wise, Huawei emphasized hardware monetization as core to its AI strategy, while Microsoft's Nadella highlighted AI agent security risks. Security concerns were raised with the discovery of the AI-powered RatHat malware targeting Chinese users.

Models

4 items

Models BuildersSelected12:21

Andromeda 2 improves drug formulation design with autonomous lab

For researchers in pharmaceutical R&D, Andromeda 2 demonstrates a significant improvement over previous agentic systems in designing high-performance drug formulations.

#Andromeda 2 #drug formulation #autonomous lab #RAG

arXiv cs.LG@Michael M. Craig et al. (6)Source

Models US vs ChinaOfficialSelected16:08

Scalabot HERON-CRA model adds context memory and RL engine

For robotics developers, this model's improved performance on a specific task shows a concrete upgrade in its capabilities.

#Scalabot HERON-CRA #robotics #embodied ai #context memory

官方一手pandaily@[email protected] (Pandaily)Source

Models BuildersSelected78°17:45

Zai.org optimizes GLM-5.3-Flash for production

This matters because it demonstrates a practical approach to optimizing large language models for real-world production environments, which is crucial for companies looking to deploy such models efficiently.

#Zai.org #GLM-5.3 #Infra Agent #model optimization

shao__meng@shao__mengSource

Models China to ASEANSelected78°17:04

Zhipu GLM reveals first RSI practice for domestic large models

This is the first time a domestic Chinese large model company has publicly used AI self-improvement for production environments, showing AI's ability to participate in the entire engineering loop of a reasoning system.

#Zhipu (Z.ai)#GLM-5.3 #Recursive Self-Improvement #AI self-improvement

IT之家Source

Products

4 items

Products China to ASEANSelected17:19

Kimi launches financial AI solution for banks and securities firms

For users in the financial sector, this solution can significantly reduce the time spent on data processing and analysis, from days to hours, by integrating multiple data sources and automating report generation.

#Kimi #financial AI #banking #securities

IT之家Source

Products US vs China78°05:04

Chinese contentOpenAI 推出 Astra for Law 法律专用版，通过 ChatGPT 和 Codex 推出

OpenAI 推出了 Astra for Law 法律专用版，通过 ChatGPT 和 Codex 推出，API 访问即将到来，方便开发者使用。

#Astra for Law #ChatGPT #OpenAI #法律专用版

官方账号OpenAI@OpenAISource

Products Builders02:52

Chinese contentGoogle Gemini 新增文件 API，支持对话中上传/下载数据

Google Gemini 新增文件 API，方便上传和下载文件，对话中就能操作，比之前更方便。

#Google Gemini #文件上传 #文件下载 #API

Philipp Schmid@_philschmidSource

Products US vs China00:59

Chinese contentAnthropic 合并 Claude Cowork，推出文档与幻灯片功能

Anthropic 把 Claude Cowork 合并到 Claude 主界面了，现在 Claude 能自己判断任务难度，还新增了文档和幻灯片生成功能，挺方便的。

#Claude #Anthropic #文档生成 #幻灯片生成

Techsauce@ Techsauce Team Source

Industry

4 items

Industry PolicySelected11:00

Huawei轮值董事长汪涛：华为AI战略的核心是算力，坚持硬件变现

For readers interested in the hardware side of AI, this explains Huawei's strategy of focusing on computing power and its specific product roadmap.

#Huawei #昇腾 #AI #算力

IT之家Source

Industry Policy06:19

Chinese content新型安卓恶意木马 RatHat 引入 AI 识别能力，可辅助黑客远程控制设备

朋友，Zimperium 发现了一种叫 RatHat 的新型安卓木马，它用 AI 来分析手机界面，让黑客能远程控制设备，比传统木马更难被查杀。

#RatHat #Zimperium #AI安全

IT之家Source

Industry Policy16:12

Chinese content台积电 COO 称 AI 不太适用于尖端芯片设计

台积电 COO 米玉杰说，AI 像三岁超人，在尖端芯片设计上不太实用，和 OpenAI、Anthropic 的事件有关。

#台积电 #AI安全 #芯片

IT之家Source

Industry Policy17:45

Chinese content微软CEO谈AI安全：AI代理成内部风险隐患

微软CEO说AI代理是内部风险，避免供应商锁定很重要，和普通AI工具不一样。

#Microsoft #AI安全 #智能体

Tech in Asia@Gilang KharismaSource

Papers

3 items

Papers BuildersOfficial23:39

Chinese content苹果发布 REVERSAL-BENCH 测试环境，用于评估无重置强化学习

苹果搞了个新测试环境叫 REVERSAL-BENCH，专门用来测强化学习里环境能不能恢复原状，对做机器人控制的人可能有参考。

#REVERSAL-BENCH #强化学习 #可逆性 #基准测试

官方一手Apple ML ResearchSource

Papers BuildersSelected12:22

Model growth and recursion influence scaling exponents

This research provides a new method to improve the efficiency of large language models, potentially reducing the computational cost of training them.

#scaling laws #looped transformers #computational efficiency #pre-training

arXiv cs.LG@Zixi Chen et al. (4)Source

Papers BuildersSelected12:12

Preventing Model Collapse with Synthetic Data

For researchers working with large language models, this provides a more rigorous mathematical understanding of the stability of training on synthetic data, which is increasingly common as human data becomes scarce.

#LLM #model collapse #synthetic data #training dynamics

arXiv cs.LG@Matteo Marchi et al. (4)Source