https://rexai.top/en/ai/llm/Recent content in LLMs on 梦兽编程Hugo -- 0.163.3en梦兽编程Sun, 19 Apr 2026 10:00:00 +0800https://rexai.top/en/ai/llm/2026-04-19-opus-47-token-cost-shock/Sun, 19 Apr 2026 10:00:00 +0800https://rexai.top/en/ai/llm/2026-04-19-opus-47-token-cost-shock/497 anonymous developer submissions reveal Claude Opus 4.7 consumes 37.3% more tokens than 4.6 on average, with API costs rising proportionally. Here's what caused the 'hidden price hike' and what you can do about it.https://rexai.top/en/ai/llm/2026-03-26-ghost-agent-memory-postgresql/Thu, 26 Mar 2026 00:00:00 +0000https://rexai.top/en/ai/llm/2026-03-26-ghost-agent-memory-postgresql/AI agents can reason, plan, and converse—but forget everything once the session ends. The Ghost project solves this with a pure PostgreSQL-based infrastructure, turning the database into the agent's memory palace.https://rexai.top/en/ai/llm/2026-03-24-apple-llm-in-flash-moe-local-inference/Tue, 24 Mar 2026 00:00:00 +0000https://rexai.top/en/ai/llm/2026-03-24-apple-llm-in-flash-moe-local-inference/An Apple paper from 2023 made it possible to run a 400 billion parameter model on an ordinary MacBook. The core technologies—MoE and quantization—hide an engineering philosophy built around on-demand loading.https://rexai.top/en/ai/llm/2026-03-23-omlx-apple-silicon-llm-inference/Mon, 23 Mar 2026 00:00:00 +0000https://rexai.top/en/ai/llm/2026-03-23-omlx-apple-silicon-llm-inference/oMLX is built for Apple Silicon, using the MLX framework, SSD-backed KV cache, and continuous batching to cut TTFT from 90 seconds to 1-3 seconds in long-context scenarios, comprehensively outperforming Ollama.https://rexai.top/en/ai/llm/2026-03-19-ramp-one-agent-financial-automation/Thu, 19 Mar 2026 00:00:00 +0000https://rexai.top/en/ai/llm/2026-03-19-ramp-one-agent-financial-automation/Ramp, America's fastest-growing enterprise finance platform valued at $32B with 50,000+ customers and $100B+ in annual transaction volume, chose a 'one Agent + a thousand skills' architecture over building many agents. This is a deep dive into Ramp's AI实战经验.https://rexai.top/en/ai/llm/2026-03-17-leanstral-intro/Tue, 17 Mar 2026 00:00:00 +0000https://rexai.top/en/ai/llm/2026-03-17-leanstral-intro/Mistral AI releases Leanstral, the first open-source Lean 4 code agent that lets AI both write code and prove its correctness. Let's talk about formal verification and AI programming.https://rexai.top/en/ai/llm/2026-03-09-ai-literate-programming-agent-era/Mon, 09 Mar 2026 10:30:00 +0800https://rexai.top/en/ai/llm/2026-03-09-ai-literate-programming-agent-era/Literate programming has been around for 40 years but never caught on — because maintaining parallel narratives of code and prose is exhausting. AI agents change that equation entirely.https://rexai.top/en/ai/llm/2026-03-09-agent-safehouse/Mon, 09 Mar 2026 00:00:00 +0800https://rexai.top/en/ai/llm/2026-03-09-agent-safehouse/A macOS tool that sandboxes AI agents like Claude Code and Codex, preventing accidental file deletions and credential leaks once and for all.https://rexai.top/en/ai/llm/2026-03-02-microgpt-200-lines-gpt-from-scratch/Mon, 02 Mar 2026 08:00:00 +0800https://rexai.top/en/ai/llm/2026-03-02-microgpt-200-lines-gpt-from-scratch/Andrej Karpathy has done it again! This time he implemented a trainable, inferable GPT model in just 200 lines of pure Python with no dependencies. This might be the most concise large language model implementation ever.https://rexai.top/en/ai/llm/2026-02-28-neural-network-reverse-engineering-puzzle/Sat, 28 Feb 2026 00:00:00 +0000https://rexai.top/en/ai/llm/2026-02-28-neural-network-reverse-engineering-puzzle/Jane Street released a special ML puzzle: given the complete neural network weights, find an input that makes it output non-zero. This isn't a typical black-box attack - it requires truly understanding what the network is computing. A college student spent days, going from linear programming to SAT solvers, and finally discovered the network was hiding an MD5 hash function.https://rexai.top/en/ai/llm/2026-02-18-glm5-vs-claude-opus-coding/Wed, 18 Feb 2026 10:00:00 +0800https://rexai.top/en/ai/llm/2026-02-18-glm5-vs-claude-opus-coding/Zhipu AI's GLM-5 scores 77.8% on SWE-bench with a 744B MoE architecture, MIT license, and $1/M token pricing — just 3 points behind Claude Opus 4.6 at 1/5 the cost. We ran real benchmarks and code tests to help you decide where your AI budget should go.https://rexai.top/en/ai/llm/ai-buddy-chat/Sun, 11 Jan 2026 00:00:00 +0000https://rexai.top/en/ai/llm/ai-buddy-chat/After a month and $400, I went from being a Claude Code fanatic to a Codex user. This is a real user experience report.https://rexai.top/en/ai/llm/ai-skills-personal-assistant-guide/Wed, 17 Dec 2025 00:00:00 +0000https://rexai.top/en/ai/llm/ai-skills-personal-assistant-guide/Step-by-step guide to using Skills so Codex and Claude shift from generalists into assistants that follow your playbook.https://rexai.top/en/ai/llm/mistral-3-official-release-europe-ai-giant-latest-work/Thu, 04 Dec 2025 15:19:00 +0800https://rexai.top/en/ai/llm/mistral-3-official-release-europe-ai-giant-latest-work/While OpenAI, Google, and Anthropic battle it out across the Atlantic, Europe’s AI power is quietly rising. On December 2, Paris-based Mistral AI officially released its third-generation model family, Mistral 3.https://rexai.top/en/ai/llm/kimi-k2-thinking-dynamic-gguf/Tue, 11 Nov 2025 07:41:51 +0000https://rexai.top/en/ai/llm/kimi-k2-thinking-dynamic-gguf/Step-by-step guide to running Unsloth's Dynamic 1-bit GGUF build of the 1T-parameter Kimi K2 Thinking model on high-end PCs, covering install, download, inference, serving, and troubleshooting.https://rexai.top/en/ai/llm/tokencake-kv-cache-vllm/Thu, 30 Oct 2025 11:49:25 +0800https://rexai.top/en/ai/llm/tokencake-kv-cache-vllm/Beihang/Peking/Alibaba introduce Tokencake, a KV-cache-centric serving framework for multi-agent apps. With time+space scheduling plus CPU buffering and progressive GPU reservation, it trims end-to-end latency by 47%+ versus vLLM and lifts GPU cache utilization by ~17%.https://rexai.top/en/ai/llm/agno-go-high-performance-agent-framework/Sat, 04 Oct 2025 00:00:00 +0000https://rexai.top/en/ai/llm/agno-go-high-performance-agent-framework/Rewriting AI Agent framework in Go brings 16x performance boost, 180ns agent startup, and only 1.2KB memory footprint - this is the extreme experience Agno-Go delivershttps://rexai.top/en/ai/llm/deepseek-v32-exp-sparse-attention-breakthrough/Wed, 29 Jan 2025 10:00:00 +0800https://rexai.top/en/ai/llm/deepseek-v32-exp-sparse-attention-breakthrough/DeepSeek-V3.2-Exp released with groundbreaking DSA sparse attention technology, 2-3x faster inference, 30-40% memory reduction, and API prices cut by over 50%