Chinese contentNemotronLabs 开源全双工语音模型 VoiceChat，支持工具调用
NemotronLabs 开源了全双工语音模型 VoiceChat，能边聊边调工具，被打断也接得住，做语音智能体的可以看看。
DailyVOL 2026.09.22112 items
OpenAI's internal AI automates model training, including GPU kernel writing
Singapore·Tuesday, 22 September 2026·DAILY · 08:00 SGT
2 Sep 2026. OpenAI's internal AI has automated most of the new experiment model's construction and training processes, including writing GPU kernels and optimizing training code. According to The Information, internal AI agents now collaborate without human intervention, reducing experimental cycles from years to weeks. Meanwhile, Anthropic reported that 26% of its AI R&D work in August was completed by AI, up from less than 1% in February, with over 90% of work reaching 'AI collaborative' levels. The company also committed up to $517 billion in compute resources, with 30,000 AI agents handling research and engineering tasks. Additionally, various industry players released new AI models and tools, such as NemotronLabs' open-source VoiceChat model and LangChain's Jev for agent evaluation, showcasing advancements in both internal AI automation and external product innovation.
01
Models
3 itemsXiaomi's MiMo-V2.6 RL training
MiMo-V2.6 demonstrates massive RL scaling with 2B tokens per step, offering insights for developers working on large-scale AI training.
Chinese contentSpaceXAI 发布编程模型 Grok 4.7：百万词元输入 2 美元起
主打写代码的 Grok 4.7 上线了，官方称速度同类两倍、价格一半，Cursor 可直接调用。
IT之家Source
02
Products
5 itemsEXL Medical IDP solution on AWS
Healthcare providers in ASEAN can reduce medical claims processing time by over 100 minutes per case using this AWS-based solution.
Jev matches human reviewers in agent evaluation
Jev offers 98% cost savings and 5x faster evaluation compared to Claude Sonnet 4.6 for developers testing AI agents.
DocJev library classifies documents 6x faster than gpt-5.6-luna
Free and open-source on GitHub, DocJev classifies documents 6x faster than gpt-5.6-luna at equivalent accuracy, giving developers a lightweight alternative to VLM-based pipelines like LlamaParse.
Zhipu (Z.ai) open-sources ZCode desktop app and terminal agent on GitHub
Apache 2.0 licensing lets developers anywhere, including Southeast Asia, self-host and audit ZCode, and the open code lets teams verify there is no hidden cloud-upload logic before trusting it with proprietary codebases.
LangSmith offers free SemIf model
Developers in ASEAN can access SemIf free for testing without cost during this limited period.
03
Industry
5 itemsChinese contentOpenAI 内部 AI 已能自动化实验模型训练，包括写 GPU kernel
OpenAI 让内部 AI 自己写 GPU kernel、优化训练代码，原本几年的实验一周就能跑完。
Chinese contentOpenAI 组建数学顾问小组，其 AI 已解决 100 多个开放数学问题
OpenAI 的 AI 已解决 100 多个开放数学问题，还组了个数学顾问团提建议，但顾问管不了研究进度。
Chinese contentAnthropic 披露内部数据：26% 的 AI 研发工作已由 AI 完成
Anthropic 把内部账本公开了：3 万个智能体在做研发，26% 的活由 AI 干，还讲了监控系统怎么拦风险。
Chinese contentAnthropic 算力承诺最高达 5170 亿美元，为此前预期的近 3 倍
Anthropic 算力合同最高 5170 亿美元，就为喂饱 Claude Code 的需求，是此前预测的近 3 倍。
Chinese contentAnthropic 在湾区建湿实验室，让 Claude 指挥机器人做药物实验
Anthropic 在湾区建了个湿实验室，要让 Claude 指挥机器人做实验，专攻难成药的罕见病，还刻意避开药厂赛道。
04
Papers
5 itemsChinese contentCodeMidas：以源代码为输入自动构建编码智能体 RL 训练环境
把现成开源代码库自动变成智能体 RL 练习题，MiMo-V2.5 训完 ProgramBench 涨了 17%。
Softmax Attention Study Identifies Abstention and Noise Filtering as Missing Primitives
Abstention explains nearly all gating gains at 10M parameters while noise filtering dominates by 350M, a concrete scaling rule for teams training or fine-tuning small language models on standard softmax attention.
AutoViewMem Reorganizes Conversational Memory into Self-Configuring Semantic Views
Builders running memory-augmented agents on open Qwen3 models can adopt this approach, since it beat strong memory baselines on LoCoMo and PersonaMem without adding retrieval routing or iterative search.
λ-Controlled GRPO Stabilizes Flow-Matching RL for Image Generation
Compared with the strongest empirical stabilizer for Flow-GRPO, λ-Controlled GRPO raises OCR text accuracy and preference reward while keeping late-step path variance on budget, replacing per-step hand-tuned fixes with a single calibrated rule.
05
How-to
5 itemsComputer Use Agent Automates E2E Testing for Mole Mac Uninstaller
Mac users get a replicable recipe for stress-testing uninstallers: the 90-plus cleanup paths Mole gained make it scrub leftovers more thoroughly than typical Mac cleaning apps.
Benchling multi-tenant AI agents security
The architecture provides a blueprint for secure multi-tenant AI deployments in regulated industries like life sciences.
Cloudflare Quick Tunnels Bypass Router Port Forwarding via Outbound-Only Reverse Tunnels
Home developers can expose local servers without touching router settings, unlike Cloudflare Named Tunnels, which require a registered domain and dashboard configuration before use.
JEV model generates 3D indoor scenes in one second
Developers building interior-design, game or AR tools can replicate this workflow by pairing JEV with pre-built model libraries instead of generating 3D meshes from scratch.
Jev Agent Harness Patterns
Practical patterns for building typesafe coding agents that developers can implement immediately.
112Stories
23First-party
21New models
43Sources
AITOP · generated by the editorial system