#AI编程
9354 related articles
AI Writes More Code, But System Design Is Still an Engineer's Job
Hacker News debate: AI writes more code, but system design still needs engineers. We analyze the real difference between coding and software engineering, why engineering judgment can't be replaced, and how AI coding affects team collaboration and review.
RTL-Interact: Teaching LLMs to Design Chips by Interacting with EDA Tools
RTL-Interact proposes a multi-task post-training framework that trains a single RTL-specific model to complete the generation, EDA checking, and revision loop, and introduces PAPO reinforcement learning, surpassing the teacher model DeepSeek-V3.2 on CVDP while reducing tool calls by 63.8%.
Human-AI Collaboration Discovers Entanglement-Assisted Quantum Code Families Through Evolution
The study proposes a researcher-in-the-loop discovery pipeline combining program evolution and formal verification to discover entanglement-assisted quantum code families, filling 54 table gaps and providing 115 verified code records.
Muon-RT: Reusing Teacher Trajectories for Momentum-Free LLM Self-Distillation
Muon-RT reuses teacher-student displacement instead of an independent momentum buffer, eliminating persistent memory overhead in self-distillation. On Qwen2.5-7B-Instruct with SDFT, it exceeds AdamW by 2.6 points while saving 42% memory, and exceeds momentum Muon by 5.4 points while saving 27% memory.
UpLoop: Training Recurrent Transformers with Zero Extra Compute, Significant Performance Gains
UpLoop proposes a pretraining scheme using upcycling to train recurrent Transformers. Without extra training FLOPs, a 1.3B model improves average score on 22 downstream tasks from 22.99 to 27.52, with enhanced math reasoning.
Deep Learning Speedrun: The Minimum Math Needed to Train an MNIST Model
Explore the minimum math needed to train an MNIST model. Based on the Deep Learning Speedrun discussion on Twitter, this article summarizes key math concepts for getting started with deep learning, including matrix multiplication, derivatives, and gradient descent, and analyzes the debate between practice-first and theory-first learning paths.
Vercel CI Launch: Turborepo Smart Task Skipping Delivers 2-6x Speedup for Internal Repos
Vercel launches Vercel CI with Turborepo smart task skipping, delivering 2-6x CI speedup for internal repos. Optimized for agents via Vercel MCP and CLI, now in Beta on all plans.
When You Hit a Dependency Error, Delete uv.lock: Lessons from a Python 3.15 Migration
Hit dependency build errors after switching to Python 3.15? Here's a practical tip: when uv.lock causes compatibility issues, deleting the lock file and re-running uv often quickly resolves errors like PyO3 version mismatches.
How the Claude Code Team Does Code Review: Risk Tiers and a Cap on Review Rounds
Anthropic reveals the Claude Code team's code review process for high-output agents: layered automated checks, risk-scaled human approval, a cap on review rounds, and turning production incidents into evals. We break down the core principles and what teams scaling AI coding tools can learn.
Understanding DeepSeek Super-agent in 3 Minutes: Architecture and Boundaries of a Team AI Coding System
What is DeepSeek Super Agent? This three-minute overview of the open-source AI coding system covers its positioning: running in DeepSeek Harness, connecting to GitHub and Feishu for team tasks, organizing coding Agents through entry, scheduling, and execution layers, distinguishing tasks from conversations, and enforcing permission boundaries. Currently a research prototype.
Claude Opus 6 Suspected Leak: Multiple Next-Gen Models Under Secret Testing
Claude Opus 6 suspected leak: Opus 5.5 requests routed to a new checkpoint; GPT-6.1 Sol default context raised to 500K tokens; Gemini 4 Barium-B vs Argon comparison exposed. This article summarizes recent secret testing signals of multiple next-gen flagship models, 3D scene demos, and Karpathy's observation on AI usage tiers.
GPT-6.5 Leaks, Claude Fable 5.5, and Chinese Models in a Frenzy: This Week's AI Headlines
This week in AI: questionable GPT-6.5 signals, Microsoft PR showing GPT-6.1, Anthropic's intent-reading model, Gemini 4 Carbon falling short of rumors, MiniMax M3.1 Flash's ultra-low pricing, and Kimi K3.1 and new GLM models on the horizon.
AI Briefing: GPT Reset and Terra Leak, Claude Unban Wave, Qwen 4 Preview
Weekly AI roundup: GPT reset and Codex reasoning intensity switching, GPT-6.1 Terra tested in GitHub Copilot, Claude unban wave and Claude Code team departures, Qwen 4 and Gemini 4 Carbon leak signals, and Kimi K3.1 Pro's cost and capability on Terminal Bench 4.
GPT-6.5 Leak Rumors and MiniMax M3.1 Flash: The AI Model Arms Race Escalates
This week in AI: GPT-6.5 leak rumors, GPT-6.1T in Microsoft code, Anthropic's model that spots whole classes of bugs, Gemini 4 Carbon underperforming, and MiniMax M3.1 Flash disrupting coding with ultra-low costs.
OpenAI, Anthropic, and Google Reveal New Moves on the Same Day: Model Iteration and Ecosystem Shifts
OpenAI reportedly testing GPT-6.2, two new Anthropic Claude models leaked, Google retires Juice coding assistant, Cognition unveils Gemini 4 coding benchmark, Cursor reports malicious extension, Apple AI integrates Tongyi Qianwen.
AI Built Me an Android Notes App: With a Built-in Thinking Partner
A Reddit user used the AI tool Hark Pro to develop a native Android notes app with Kotlin and Jetpack Compose, featuring a built-in AI thinking partner, connection engine, idea development, and semantic search, showcasing new possibilities for AI programming and personal knowledge management.
ChatGPT and Codex Quota Pools Merge: Chat and Coding Share Quota Starting December
ChatGPT and Codex quota pools merge in December, sharing quotas for chat and coding. This article covers merger rules, official responses to coding model quality controversy, tightened device login limits, NVIDIA's weight compression, rumors about next-gen domestic models, and open-source tool progress.
ChatGPT and Codex Share a Quota: How Dual-Use Users Should Split the Bill
ChatGPT Plus, Pro, and Business plans will share a token budget with Codex; Instant model chats are excluded. Here's how dual-use users can split writing and coding budgets.
15-Year-Old Developer Builds a Hardware-Aware SLM Trainer from Scratch: Python and Rust in Practice
A 15-year-old developer builds a hardware-aware SLM trainer from scratch using Python and Rust, automatically estimating trainable model size (up to 100M parameters) and handling arXiv paper data pipelines while protecting LaTeX code blocks. The project is open-source, sharing technical decisions and learning insights.
Training a PPO Agent in UE5: 72,000 Nights of Guarding the Campfire
A developer shares a UE5 project training a PPO reinforcement learning agent to guard a campfire, covering ~72,000 nights of training and key issues in environment design, reward functions, and training efficiency.