#AI编程
9158 related articles
Anthropic Merges 30,000 Code Changes a Day — So Why Are 98% of Open-Source Issues Unclaimed?
Anthropic merges 30,000+ code changes daily, yet over 98% of issues in its public Claude Code repo are unassigned, including privacy and security bugs. Developers ask: why not use Claude for issue triage?
CoRD: How Coupled Sampling Fills the Blind Spot of On-Policy Distillation
CoRD uses maximal coupling from speculative decoding to interpolate student and teacher distributions, filling the blind spot of on-policy distillation. It beats on-policy distillation on pass@1 and pass@8 across six Qwen3/Qwen3.5 students.
No Verifier Needed: How Test-Time Self-Play Breaks Through LLM Reasoning Bottlenecks
The paper proposes Test-Time Self-Play (TTSP), a verifier-free test-time scaling method. On LiveCodeBench v6, it achieves a 35% absolute improvement in average@8 without a verifier, with greater advantages as verifiers weaken, offering new ideas for scenarios lacking verifiers.
The Two-Lane Gradient Turntable Property of Truncated BPTT: A Theoretical Breakthrough
Truncated BPTT introduces gradient bias, but new research proves that under exponentially decaying state-transition Jacobians, the error between truncated and full-context gradients decays exponentially with distance from the context boundaries, forming a two-lane gradient turntable property that provides theoretical guarantees for choosing truncation length.
Non-Coding Dad Builds 3D Earth Toy with Claude Code, Unexpectedly Goes Viral in Classrooms
A non-coding father used Claude Code to build My Earth Toy, a 3D interactive Earth with 830+ landmarks, 470+ animals, and 180+ vehicles—now used in classrooms worldwide. Discover how AI coding lowers the barrier to creation.
Merge Uses GPT-6 Astra and Codex to Reshape Product and Business Tools
Merge uses GPT-6 Astra, ChatGPT, and Codex to review product code, support marketing, prepare sales updates, and plan executive dinners, showing how the OpenAI stack is deployed across multiple business lines and what teams can learn from it.
MuJoCo Parallel Training: Single World with Multiple Robots vs. Independent Environments
Explore the differences between single-world multi-robot and independent environment parallelism in MuJoCo RL, analyze their advantages, and learn best practices for migrating to GPU.
Claude Code Reportedly Sends User Email Addresses to Websites
Claude Code reportedly sends user email addresses when visiting websites, sparking privacy debate on Hacker News. This article analyzes the root cause, community divide, and practical impact on developers, exploring privacy transparency in AI coding tools.
DeepSeek Harness Quick Start Guide: From Installation to 5 Real-World Cases
What is DeepSeek Harness? Based on a Bilibili tutorial, this article covers how this AI Agents software works, installation and setup, four working modes, and five real-world cases: a 3D racing game, official website, music player, science animation, and self-developed plugin—plus token money-saving tips.
Breaking Down DeepSeek's Super-agent Architecture: How Router+Worker Solves the Multi-Session Management Bottleneck
An analysis of DeepSeek's Super-agent paper: how the Router+Worker two-tier architecture uses Fast Pass, Chain of Recall, and storage-compute separation to solve multi-AI-session management bottlenecks, and how much it truly alleviates the human attention bottleneck and what engineering trade-offs it makes.
19K Stars in 5 Days: Why DeepSeek's Open-Source AI Programmer Went Viral
DeepSeek's open-source AI programmer hit 19K GitHub stars in 5 days. Why it went viral: the agent boom, AI writing its own code, and the shift from Q&A tools to AI employees.
GPT-6.2 Astra Leaked: SVG Generation and OpenAI's Hardware Ambitions
Fact-checking GPT-6.2 Astra's capability data and SVG demos, plus Tencent HY4 open-source model, Anthropic Flash Proof cryptographic proofs, DeepSeek's Huawei chip training plan, and Musk's AI power bottleneck warning—latest AI developments in one article.
Token Pass Credits Now Automatically Apply to AI Gateway: No Dedicated Key Needed
Token Pass credits now automatically apply to covered model calls in AI Gateway, supporting API keys, OIDC tokens, and fx login—no dedicated key needed. After credits run out, standard key requests fall back to team AI Gateway credits, while Token Pass keys stop until the next billing cycle; fallback paths follow the team's ZDR settings.
Deploy Open-Source Video Model MiniMax H3 Locally in 5 Minutes and Skip Cloud Generation Fees
Bilibili creator tests local deployment of open-source video model MiniMax H3: three-step ComfyUI install, 5-minute setup, no coding. Same prompt generates in 2 minutes locally vs 4 minutes online with queuing, saving about 30 RMB per 15-second 720P video. Includes VRAM requirements and RTX 5080 test results.
Making AI Coding Agents Remember Rejected Ideas: Lessons from an Open-Source Tool
A developer running multiple Claude Code Agents found that code persists across sessions but decision reasoning—especially rejected options—is lost, causing new sessions to re-propose abandoned approaches. He built the open-source tool cogit to capture messages, summarize them, and link them to commits, but faces challenges with summary accuracy and overly narrow per-file prompting.
GitHub's New Git Storage Architecture: How to Handle Thousands of Concurrent Agent Pushes
GitHub unveils a new Git storage architecture designed to support thousands of AI Agents pushing to a single repository simultaneously. This article analyzes the core problems, technical approach, and impact on AI programming workflows.
Claude Built a Programming Language Out of Memes: Memesis Lets AI Agents 'Meme Around' Safely
A developer used Claude Code to build Memesis, a programming language for AI agents with internet meme keywords, built-in sus/yolo safety, open source and locally runnable.
Vercel Schedules Launches: Enabling AI Agents to Schedule Tasks in Real Time
Vercel Schedules enters public beta, enabling AI agents and apps to dynamically create and modify one-time and periodic tasks at runtime without redeployment. This article analyzes its static and dynamic scheduling modes, agent task implementation workflow, and security design.
Conductor Acquired by Meta: Shutdown Timeline and Alternatives
After joining Meta, Conductor's cloud service goes read-only on Nov 13 and shuts down Nov 20; the local Mac app ends in early 2027. See the timeline, export checklist, and migration alternatives.
Building AI Apps with Claude Code from Scratch: A ZTM Free Course Review
ZTM launches "Claude Code for Beginners," a free course for complete beginners who build a Battleground AI battle app with Claude Code and Claude Design, seeking feedback.