Tag: #llm

한국어English日本語中文

← All posts

• Do You Need a New Learning Rate When You Change LoRA Rank? — The Two Regimes μA (2026) Splits, and the Limits of That Evidence

  Two competing beliefs circulate in practice at the same time: that changing the rank in LoRA means you have to re-find the optimal learning rate, and that using 1/r scaling makes the learning rate independent of rank. Th

  2026-07-16 · 18 min read #llm#lora#fine-tuning#peft#training

• From Documents to a Knowledge Graph: An Honest Pipeline

  'Extract a knowledge graph from your documents' looks like a single LLM call in a demo. But turning a customer's documents into a graph you can actually query is a six-stage pipeline, and most of the cost and pain lives

  2026-07-15 · 10 min read #knowledge-graph#ai#llm#data-engineering

• Graph RAG, Explained: What It Is and When It Earns Its Cost

  The standard RAG recipe — chunk, embed, retrieve top-k — works when the answer sits inside a single chunk, but it stalls structurally on multi-hop questions and on global sensemaking questions that span the whole corpus

  2026-07-15 · 8 min read #rag#graph-rag#knowledge-graph#ai#llm

• Tencent Hy3: reading a 295B open-weight MoE without the hype

  On July 6, 2026, Tencent released Hy3 under Apache 2.0: a Mixture-of-Experts reasoning and agent model with 295B total parameters but only 21B active per token. Here is what is genuinely new, where it sits in the Chinese

  2026-07-11 · 5 min read #hy3#tencent#hunyuan#open-weights#moe

• Running Small Models Hands-On with a Single RTX 5090 — microGPT, OCR, Music Generation

  I SSHed into a single RTX 5090 (Blackwell, 32GB) and ran a trio of small models by hand. I trained a char-level GPT from scratch in 28 seconds (10.75M parameters, 1.17M tokens/s), pitted a dedicated OCR model (TrOCR) aga

  2026-07-11 · 8 min read #pytorch#gpu#llm#ocr#hands-on

• What a Good Agent Benchmark Looks Like in 2026 — UniClawBench, Live Containers, and a Hidden Supervisor

  UniClawBench, posted to arXiv in July 2026 by HKU MMLab, is a self-described capability-driven benchmark for proactive agents. Instead of matching against static, pre-recorded answers, it runs agents inside live Docker c

  2026-07-11 · 5 min read #ai#agents#evaluation#benchmark#llm

• Why RLHF Models Game Their Rewards — The Mechanisms, Symptoms, and Mitigations of Reward Hacking

  "Reward Hacking in the Era of Large Models," posted to arXiv in April 2026 by Xiaohua Wang and 22 co-authors, is a survey of why and how RLHF-aligned large models game their reward signals. Its central proposal is the Pr

  2026-07-11 · 5 min read #ai#llm#alignment#rlhf#safety

• More Thinking Is Not More Accuracy: Test-Time Compute and the Overthinking Cliff

  In April 2026, Shu Zhou and five co-authors question the reflex to keep adding reasoning tokens in "When More Thinking Hurts". As the compute budget grows, the authors report that the marginal utility of extra reasoning

  2026-07-11 · 6 min read #ai#llm#reasoning#inference#test-time-compute

• The Model Context Protocol (MCP): An Engineer's Reference

  The Model Context Protocol (MCP) is an open protocol that standardizes how applications provide context to LLMs. This is an engineer's reference grounded in the official docs: why it solves the M×N integration problem, t

  2026-07-11 · 6 min read #mcp#ai#agents#llm#protocol

• LLM Burnout: When the Job Becomes Reviewing the Work Instead of Doing It

  A developer's essay, 'I Think I Have LLM Burnout,' trended because it named a quiet shift: the workday is turning from writing code into reviewing what a model wrote. This post represents that argument faithfully and add

  2026-07-11 · 6 min read #llm#developer-experience#burnout#ai-tooling#code-review

• Quantizing the KV Cache to 4 Bits — SAW-INT4 and System-Aware INT4

  In long-context LLM serving, the real memory bottleneck is not the weights but the KV cache. Quantize it naively to INT4 and accuracy collapses; the April 2026 preprint SAW-INT4 pairs token-wise INT4 quantization with a

  2026-07-11 · 6 min read #ai#llm#quantization#inference#kv-cache

• Reading Microsoft Flint — a Visualization Language for Agents to Draw Charts, Not to Be Drawn

  Microsoft Research's Flint is not a language for visualizing agents — it is an intermediate language that lets AI agents reliably produce good-looking charts from data. A compiler derives the low-level decisions (scales,

  2026-07-11 · 5 min read #ai#agents#data-visualization#llm#microsoft

• Optimizing Long-Context Inference with Hybrid SWA — What Xiaomi's MiMo v2.5 Actually Does

  Hybrid SWA interleaves sliding-window attention with full attention across layers to cut both KV-cache memory and compute for long-context inference — now a mainstream design. Grounded in Xiaomi's published MiMo v2.5 inf

  2026-07-11 · 7 min read #ai#llm#inference#attention#optimization

• Upcycling Pretrained Checkpoints into Long-Context Hybrids — What HyLo Proposes

  Hybrid sequence models (attention plus linear/SSM blocks) suit long context, but until now most had to be pretrained from scratch. The April 2026 preprint HyLo proposes a recipe to 'upcycle' an already-trained Transforme

  2026-07-11 · 5 min read #ai#llm#long-context#architecture#efficiency

• Diffusion LLMs That Write CUDA Kernels — DICE and Why Parallel Generation Might Help

  DICE, a February 2026 preprint, claims that diffusion large language models beat autoregressive models of the same size at generating CUDA kernels, setting a new state of the art. The core idea: instead of writing one to

  2026-07-11 · 6 min read #ai#llm#diffusion#cuda#gpu

• Running GLM-5.2 on a slow computer — how colibrì streams a 744B model from disk

  colibrì is a ~1,300-line pure-C inference engine that runs GLM-5.2, a 744B-parameter MoE model, on a consumer PC with 25GB of RAM. The trick is MoE sparsity plus disk streaming: only ~9.9GB of dense layers stay resident,

  2026-07-11 · 5 min read #ai#llm#local-inference#moe#quantization

• Choosing How Much to Think Per Step: Ares and Adaptive Reasoning-Effort Routing

  A practical look at Ares, a March 2026 preprint that treats reasoning effort as a per-step cost lever for LLM agents. A lightweight router reads the interaction history and predicts the lowest sufficient reasoning level

  2026-07-11 · 4 min read #ai#agents#llm#efficiency

• When an AI Maintains Your Code, Write for Humans Anyway

  On July 10, 2026, Scott Robinson revived an old maxim with a twist: an LLM reads your codebase as its style guide, so every shortcut you merge becomes training data it repeats back at machine scale. We walk through his d

  2026-07-11 · 6 min read #ai#llm#code-quality#maintainability#craft

• Building Effective AI Agents: A Reference on the Five Workflow Patterns and Agents

  A practical reference distilled from Anthropic's engineering guide "Building Effective Agents." It covers the precise distinction between workflows and agents, the building block underneath everything — the augmented LLM

  2026-07-11 · 8 min read #ai#agents#llm#engineering#anthropic

• Production RAG Patterns — Why Naive RAG Fails and the Techniques That Actually Help

  A demo RAG system takes half a day to build, but the place it quietly breaks in production is almost always retrieval, not generation. This reference walks through chunking, embeddings and hybrid search (BM25 + vector),

  2026-07-11 · 9 min read #ai#rag#llm#retrieval#embeddings