Skip to content

Dev weekly 2026-Jul-13

Published: at 02:02 AM

AI

Harness Engineering for Self-Improvement 利用工程化手段实现自我提升

Better Models: Worse Tools 较新的 Claude 模型有时会在嵌套的 edits[] 数组中调用 Pi 的编辑工具时,加入额外且凭空捏造的字段。

文译 专注于将多语言 EPUB、FB2 或 TXT 小说翻译成中文,并尽量保留 EPUB 原排版、图片、目录和跳转。

Getting started with loops

How Schneider Electric Built Their LLMOps Foundations At Enterprise Scale With LangSmith langsmith architecture

正式开源!美团 LongCat-2.0 同步开放国产卡推理代码 https://github.com/meituan-longcat/LongCat-2.0

LiteRT.js, Google’s high performance Web AI Inference

Tau A small, readable terminal coding agent — and a working example of how coding agents are built, from huggingface.

Improving Agents is a Data Mining Problem langchain

Gemma 4 Technical Report

Fine-Tuning Financial LLMs for Customer Intent and Experience 中文演说。Jui Sheng Chang and Walter Hsieh from Cathay Financial Holdings and Cathay United Bank will share how financial LLMs can be built and fine-tuned for customer intent classification and improved customer experience.

Machine Learning Systems 人工智能系统工程的原理与实践

Architecture 2.0

Recreate Our 1-Minute Seedance 4K Film - The Full Seedance 4K Breakdown With All Prompts 重现我们那部 1 分钟的 Seedance 4K 短片——Seedance 4K 全流程解析及所有提示词

The State of the AI Economy

Radar Trends to Watch: July 2026 值得关注的雷达趋势:2026 年 7 月

欢迎来到 Learn Harness Engineering

CS294/194-280 Advanced Large Language Model Agents

Agentic AI Architecture infoq电子书

Achieving operational excellence with AI

Fable is not a useful model

Pi: The Coding Agent That Hides Nothing

LLM-as-a-Verifier A General-Purpose Verification Framework

The new GPT-5.6 family: Luna, Terra, Sol

Programming

Announcing TypeScript 7.0

Rewriting Bun in Rust

Benchmarking Coding Agents on Databricks’ Multi-Million Line Codebase 模型所依托的工具框架对成本和质量有着巨大影响。在许多情况下,像 Pi 这样简单的框架在我们的工作负载中表现最好。

How Prompt Tuning Improved GPT-5.5 in VS Code

Teaching the Agent Our Craft: Structured Agentic Development on a Real Codebase

astryx - design system

SentencePiece 文本分词工具。快,轻,不用人工标注。给神经网络生成系统用

AgentHarness Evaluation harness for Apodex-1.0 on public deep-research benchmarks.

MarketFish - Don’t guess. Simulate 多智能体市场模拟引擎。不是问一个大模型”这个产品能行吗”,而是建造一个由 128+ 个 AI 消费者组成的数字市场——每个人都有独立的身份、预算、情绪和偏见——让他们在 30 轮中自由买卖。他们的购买决策、流失模式、社交影响,揭示真实用户会怎么做。

Maka 桌面 AI 工作台 基于pi

Learning to Replicate Expert Judgment in Financial Tasks 总而言之,LLM 在资金调配中表现出高信念、叙事驱动的自由交易风格。结果更多由集中度、叙事执念和不对称下行风险决定,而非交易层面的优化。

Palantir Ontology的Neo4j+业务规则引擎平替实现及其如何”驱动”Agent?

Learning to Replicate Expert Judgment in Financial Tasks 学习在金融任务中复刻专家判断 我们测试过的前沿模型在处理相对简单的金融任务时表现吃力,且模型本身的进步并未带来多少性能提升。相比之下,我们已经证明,使用由专家投资者标注的高质量专属数据集进行微调,可以产生在我们的任务上超越前沿性能的定制模型。我们发现,这一结果不仅适用于我们在本文中讨论的六项任务,在更广泛的领域也同样成立。

如何萃取老師傅的知識做成 Agent Skill? 六條路線的方法論框架

你的 Prompt 通得過 The Mom Test 嗎? 如何避免 LLM 的迎合問題

What is the future of software engineering with Adam Bender, Principal Software Engineer at Google

Six Things Enterprise Architects Get Wrong About the Operational–Analytical Chasm — and How to Close the Agentic Loop Snowflake

UML — See Your Codebase Point it at a folder. Get a diagram back.

Ten Years Away: Designing an Interactive Comic for Studio375’s Tenth Anniversary The story behind “Ten Years Away”, a scroll-driven interactive comic created to celebrate Studio375’s tenth anniversary, from its printed origins to its WebGL implementation.

Your Kafka Infrastructure Just Became Optional Datastream:兼容 Kafka 的原生端点

JavaScript in-page GUI agent. Control web interfaces with natural language.

Anomaly detection using dynamic thresholds and two-year-long alerts in Cloud Monitoring

CS25: Transformers United V6

Stanford CS25: Transformers United V6 I From Language Models to Native Multimodal Intelligence

Evaluating ChatGPT as a Portfolio Decision-Maker in Micro-Cap Equities

Heads of Enterprise Architecture Must Rewire EA for AI

Flint: A Visualization Language for the AI Era A Microsoft Research project

Elastic 开源了基于认知科学的 Atlas Agent Memory https://www.elastic.co/search-labs/blog/agent-memory-elasticsearch

Survey Says: Agentic AI Penetrates the Enterprise, but Some ROI Questions Remain 《Battery Ventures 企业技术支出状况调查》。该调查涵盖了 100 位资深技术领导者的反馈,他们所代表的年度技术支出总额超过 660 亿美元

Your LLM Has a Spending Problem - How Prompts, Context Windows, and Agent Loops Quietly Torch Your Budget 你的大模型(LLM)有“乱花钱”的问题——提示词、上下文窗口和智能体循环是如何悄悄掏空你的预算的 AIFinops

Others

AI 季报 26Q2:从 coding 到 RSI,强者愈强的未来?

播客|从「西印度毁灭述略」到冷战核危机:漫谈古巴与加勒比海五百年

《晋书》专题|林昌丈:魏晋史上的南与北

Product Design Psychology by Wouter de Bres Understand the minds you’re designing for and the mind you’re designing with.

iPhone Photography Awards - The 2026 Winners

mobile01 旅遊美食

浓雾与犯罪 华生的回忆录里不止有风和雨(福华的秋窗风雨夕》,还有伦敦那尽人皆知的雾。

How to ask for help from people who don’t know you

宏观经济学 7 天速成课教学实验复盘报告

播客|与柏琳漫谈巴尔干千年史:从古南斯拉夫诸国到近代科索沃危机

朱雀仿宋/朱雀宋朝 「朱雀仿宋」是璇玑造字的开源仿宋字体计划

ML Job Interviews: The Ultimate Guide

RED FLAG - LOCOMOTIVE ACT 1865 1865年的英国,针对道路上日益增多的自行驱动车辆,立法做出的回应是《机车法案》(有时被称为《红旗法案》)。该法案包含多项条款,其中规定自行驱动车辆前方必须有一名步行者,手持红旗,以提醒其他道路使用者车辆即将到来。 既得利益者的目的是阻碍进步, 以保护他们在马匹、马车和火车上的投资。