#AI编程
9352 related articles
The "Harmful Success" of LLM Coding Agents: Security Concerns Behind 95% Accuracy
VeriHarm-Bench introduces the concept of "harmful success," targeting the blind spot where LLM coding agents cause unintended harm while completing benign tasks. Frontier agents reach 95% task accuracy but only 37% safe success rate on average, revealing that task success metrics may severely underestimate security risks.
Four Bridges Game: Why Do AI Agents Deceive Without Any Instruction?
A new study uses the Kradle platform and Four Bridges game to measure spontaneous AI deception in multi-agent interactions: all eight frontier models deceived, across 73% of runs, without any deception instructions.
OpenAI Pando: The Git Backend Built for Codex-Scale Pushes
OpenAI built Pando, a Git-compatible backend designed to handle massive push pressure from AI coding agents like Codex. This article examines its design motivations, technical trade-offs, and impact on agent fleets.
Alibaba Qwen Code Repository Leaks Qwen4 Series Models: Max, Coder-Plus, and Omni Exposed
Alibaba's Qwen Code repository Pull Requests accidentally leak Qwen4 series model IDs, including qwen4-max, qwen4-coder-plus, and qwen4-omni-120b, all with a unified context length of 262,144. This article analyzes the naming system, product roadmap, and community reaction.
Fine-Tuning a 322M Model in Practice: Three Rounds, Two Failures, and a Safer Coding Agent
Developer documents the laya-guard project: fine-tuning a 322M model as a command safety guard for coding Agents. Three training rounds, two failures, revealing data length bias, comment bias, and secret-exposure blind spots, delivered at a 61% allow rate, with reflections on evaluation independence.
Building a Deep Learning Framework from Scratch in Pure NumPy: A New Lightweight DL Option
A lightweight deep learning framework built entirely from scratch in Pure NumPy has gained Reddit attention and JOSS acceptance. This article analyzes its teaching and research value, its applicable boundaries, and why implementing a framework from scratch is the best path to understanding deep learning.
Claude's Filters Tighten: Why Game Mod Developers Are Worried
A game mod developer worries Claude is becoming less useful due to tightened copyright filtering and content policies. We analyze their use cases, Claude's refusals, and the structural risks of relying on a single AI tool.
FastAPI Interview Experience Sharing: Reddit Community Discussion
A Reddit community discussion on Python FastAPI interview experiences, outlining common FastAPI interview topics including async programming, dependency injection, and Pydantic validation.
Tech FrontiersLangChain Deeply Integrates Async Subagents, Deep Agents Officially Launches
LangChain reveals most of its internal agents use async subagents meaningfully, now available to developers via Deep Agents, pushing agent workflows toward parallelism.
Fully Automated Video Editing with Claude Code: Build an AI Editing System with Zero Code
No code required to build a fully automated AI editing system: Claude Code coordinates seven tools including Parakeet, HyperFrames, and ffmpeg, plus a Skill-style manual, to achieve zero manual timeline editing and drastically cut costs and production cycles.
Anthropic Official Extension Poisoned: 5,000 Developer Machines Hit by Backdoor
Anthropic's official Claude extension for Cursor was poisoned in a supply chain attack. The malicious version lived nearly 12 hours, affecting 5,000 developer machines and planting SSH backdoors to steal cloud credentials. This article analyzes the attack chain, the silence of both companies, and deeper warnings for AI Agent deployment security.
Are 90% of AI-Written Tests Pseudo-Logic? Three Blind Spots and Probe Tactics
Why are 90% of AI-generated unit tests pseudo-logic? This article breaks down three blind spots—fake coverage, hollow logic, and missing side effects—with hands-on mutation testing, boundary probes, and side-effect hooks to turn AI tests from decoration into a real quality line.
AI Writing Too Much Code? Ponytail Plugin Cuts Output by 54%
Ponytail is an open-source plugin that makes AI coding assistants write less code, using a six-tier code ladder to prioritize reusing existing solutions. In tests across twelve tasks, lines of code dropped by an average of 54%, with the date picker task down 94%. Works best with Claude Code.
Ponytail: 176K Stars — Teaching AI Coding Assistants to Be Lazy and Halve the Code
Ponytail is an open-source project that teaches AI coding assistants to be lazy with a seven-rung ladder strategy: find existing solutions first, then write minimal code. Official tests show 53% less code, 41% less time, and high-risk logic test coverage rising from 68% to 98%. It has 176K stars on GitHub.
Can AI-Written Specs Generate Code Directly? A Multi-Agent Review Workflow Explained
Can AI-written Specs generate code directly? Bilibili creator Lao Dao Qian Li shares the SDFlow multi-agent review workflow: fresh contexts, parallel reviews from multiple angles, citation verification, then item-by-item adjudication. 50 issues trimmed to 18 adopted; humans only make key decisions. 17 minutes and 15 million tokens buy reduced uncertainty and early error correction.
Codex Freelancing in Practice: From Zero to Your First Earnings with AI Programming
A Bilibili creator shares how they monetized AI programming gigs with Codex from zero, covering tiered freelance platforms, collaboration models, and the opportunities and risks of AI coding freelancing.
Why Is Writing AGENTS.md First the Wrong Move When Handing Legacy Code to AI?
Handing legacy code to AI for refactoring? Writing AGENTS.md first is wrong. The real prerequisite is runnable tests as an exit condition—let AI draw the map itself.
AI Keeps Reworking Your Code? Plan First, Code Later to Slash Rework
AI code always needs rework? Learn plan mode in AI coding: for big changes, have AI propose a plan before coding. Focus on three things—which files to change, how to verify, where to confirm—plus two copy-paste prompts to cut rework.
Vercel Project Configuration Adds vercel.toml Support
Vercel adds vercel.toml support for project configuration, with properties matching vercel.json, plus comments and a flatter structure. Learn the one-of-three constraint with vercel.json and vercel.ts, plus Services configuration examples and migration tips.
LLM 0.37 Released: New Models and Lessons from AI-Assisted PRs
LLM 0.37 adds gpt-6.1-sol and gpt-6.1-terra models, the llm openai endpoint add command, a cheaper default model gpt-6-luna, and multimodal embeddings.