MLog

All Posts

Filter by keyword, tag, and category.

Tech#AI#DeepSeek#LLM

DeepSeek V4 and the Great AI Reversal: How Chinese LLMs Took the Lead

DeepSeek V4 launches on July 15 with a 1M-token context window, time-of-use pricing, and industrial-grade code generation. Meanwhile, Chinese LLMs have dominated OpenRouter's global rankings for ten consecutive weeks. Here's a deep dive into the technology, economics, and geopolitics behind this shift.

Jul 13, 20264 min
技术教程#AI#GPT-5.6#OpenAI#大模型#技术解析

GPT-5.6 Launches: A Deep Dive into the Sol, Terra, Luna Three-Model Architecture

On July 9, 2026, OpenAI officially launched the GPT-5.6 model family, introducing a three-tier architecture with Sol, Terra, and Luna. This article provides a comprehensive analysis of the technical architecture, performance benchmarks, pricing strategy, and practical guidance for model selection.

Jul 13, 20266 min
AI生成#AI#Agent#LLM#2026

The AI Agent Era: From Passive Q&A to Autonomous Execution

In 2026, the AI industry officially transitions from chatbot-style models to autonomous AI agents. This article explores the key capability leaps, current product landscape, and what this paradigm shift means for developers and everyday users.

Jul 12, 20265 min
Tech#AI#frontend#devtools

The npm Moment for AI Coding: From Writing Prompts to Installing Skills

Vercel's skills project hit 24K GitHub stars in five months, AGENTS.md has been adopted by 60K+ open-source projects. AI coding is shifting from 'teaching the AI' to 'the AI finding its own capabilities.' Here's what that means.

Jul 12, 20266 min
Tech#AI#GPT-5.6#Machine Learning#Agent

When AI Becomes the Teacher: GPT-5.6 and the First Domino of Recursive Self-Improvement

OpenAI's GPT-5.6 Sol independently post-trained the Luna model—not AI-assisted training, but AI acting as the researcher itself. The recursive self-improvement flywheel has started spinning.

Jul 12, 20266 min
Tech#AI#LLM#Interpretability#Anthropic

Inside Claude's Mind: Anthropic Discovers a Spontaneous 'Global Workspace' in Large Language Models

Anthropic researchers developed Jacobian Lens, a new interpretability tool that reveals a spontaneously emergent 'J-space' inside Claude — a neural structure functionally analogous to the brain's global workspace, capable of reading the model's unspoken thoughts. Removing this structure nearly eliminates multi-step reasoning, opening a new frontier for AI interpretability and safety.

Jul 12, 20265 min
后端#AI#后端#模型路由#工程化

Model Routing Is Just a Backend Problem

When your application needs to talk to half a dozen LLM APIs simultaneously, model routing stops being an AI problem and becomes a proper backend engineering one—auth, rate limiting, fallbacks, canary releases, and billing. All familiar territory.

Jul 12, 20266 min
Tech#AI#Developer Tools#Open Source

Not Smarter, Just More Disciplined — Superpowers and the 210K-Star AI Coding Workflow

Superpowers doesn't make AI coding agents smarter. It makes them follow the rules. How a set of Markdown files enforcing software engineering discipline racked up 210K GitHub stars in a month.

Jul 12, 20265 min