jidonglab

1 project a week. Building and sharing the entire process — from idea to shipped product in 7 days. Currently: AI news automation.

korea Joined on Feb 25, 2026 jee599@naver.com https://jidonglab.com/

Pinned

jidonglab

Mar 19

Claude Code Agent Teams Can Spawn Agents. It Just Doesn't Know Which Ones to Use.

#ai #claude #productivity #opensource

4 min read

jidonglab

Mar 6

What 100 Trending GitHub Projects Tell Us About Where AI Is Actually Going

#ai #opensource #productivity #webdev

4 min read

jidonglab

Mar 3

Anthropic Dropped 13 Free Courses — I Broke Down Every Single One

#ai #productivity #claude #beginners

3 min read

jidonglab

Jul 17

Attention Sinks: Why Streaming LLMs Break When You Evict Token 0

#deeplearning #llm #machinelearning #nlp

6 min read

Want to connect with jidonglab?

Create an account to connect with jidonglab. You can also sign in below to proceed if you already have an account.

Create Account

Already have an account? Sign in

jidonglab

Jul 16

Why Temperature 0 Doesn't Make Your LLM Deterministic

6 min read

jidonglab

Jul 16

Speculative Decoding: Why a Great Draft Model Still Caps Speedup

6 min read

jidonglab

Jul 15

Constrained Decoding: Force Valid JSON Without Wrecking Accuracy

6 min read

jidonglab

Jul 13

GRPO Explained: Why DeepSeek Dropped the Critic in RLHF

6 min read

jidonglab

Jul 12

Online Softmax: How FlashAttention Skips the N N Matrix

7 min read

jidonglab

Jul 12

Late Interaction Retrieval: Why ColBERT Beats Single-Vector RAG

7 min read

jidonglab

Jul 11

Why repetition_penalty Quietly Corrupts Your Code Generation

6 min read

jidonglab

Jul 11

DPO Likelihood Displacement: When Preferred Answers Get Rarer

6 min read

jidonglab

Jul 10

Why Token Logprobs Beat Asking Your LLM How Confident It Is

7 min read

jidonglab

Jul 9

Multi-Token Prediction: DeepSeek's Built-In Draft Model

6 min read

jidonglab

Jul 9

AWQ: How Activation-Aware Quantization Saves 4-bit LLMs

6 min read

jidonglab

Jul 8

PagedAttention: Why Static Batching Wastes Your KV Cache

6 min read

jidonglab

Jul 8

11 Claude Agents Audited My Kit. 2 'Major Bugs' Were Fake

#ai #claudecode #opensource #programming

6 min read

jidonglab

Jul 8

Semantic Entropy: Detect LLM Hallucinations Without Ground Truth

6 min read

jidonglab

Jul 7

Contextual Retrieval: Fix the RAG Chunk That Lost Its Context

6 min read

jidonglab

Jul 7

Prefilling Claude's Response: Steer Output Without JSON Mode

6 min read

jidonglab

Jul 7

I Cut My Claude Code Prompt Overhead 84% Per Turn (v1.4)

7 min read

jidonglab

Jul 6

Multi-head Latent Attention: The KV Cache Trick Beyond GQA

7 min read

jidonglab

Jul 6

Why Your Mixture-of-Experts Model Silently Drops Tokens

7 min read

jidonglab

Jul 5

Embedding Anisotropy: Why Cosine Similarity Never Hits Zero

6 min read

jidonglab

Jul 5

Return Claude's Thinking Blocks or Your Agent Breaks

6 min read

jidonglab

Jul 4

Binary Quantized Embeddings: 32x Smaller Vectors, Recall Intact

7 min read

jidonglab

Jul 4

Prefill/Decode Disaggregation: Stop Serving LLMs on One GPU

6 min read

jidonglab

Jul 3

Chunked Prefill: Why One Long Prompt Freezes Your LLM Server

#ai #llm #performance #systemdesign

7 min read

jidonglab

Jul 3

How I Made Opus 4.8 Act Like Fable 5 (64% 97%, Measured)

7 min read

jidonglab

Jul 3

FP8 KV Cache Quantization: The Memory Math and the Accuracy Cliff

6 min read

jidonglab

Jul 2

Why LLM Decoding Is Memory-Bound: Prefill vs Decode Roofline

6 min read

jidonglab

Jul 1

Min-p Sampling: Why Top-p Breaks at High Temperature

7 min read

jidonglab

Jul 1

Matryoshka Embeddings: Truncate Vectors 12x Without Losing Recall

6 min read

jidonglab

Jun 30

The Hidden Token Tax of Tool Use in LLM Agents

7 min read

jidonglab

Jun 30

Grouped-Query Attention: The KV Cache Math Behind Long Context

6 min read

jidonglab

Jun 29

RoPE Scaling: How LLMs Stretch From 8K to 128K Context

7 min read

jidonglab

Jun 29

Prompt Caching With Claude: Where the Cache Breakpoint Goes

6 min read

jidonglab

Jun 28

Attention Sinks: Why Evicting Your LLM's First Token Breaks It

7 min read

jidonglab

Jun 28

Why Your LLM-as-Judge Disagrees With Itself (And How to Fix It)

7 min read

jidonglab

Jun 27

Speculative Decoding: Why Two Models Decode Faster Than One

7 min read

jidonglab

Jun 27

Structured Output Isn't Free: The Constrained-Decoding Tax

6 min read

jidonglab

May 3

71,700 Stars and 60 Rust Crates: Inside OpenAI's Codex CLI Source

6 min read

jidonglab

May 3

Pentagon Blacklisted Anthropic From 8 Classified AI Deals

#news #ai #opensource #business

7 min read

jidonglab

May 3

Anthropic $900B: 2.4x in 90 Days, 48-Hour Window

#news #ai #business #startup

5 min read

jidonglab

Apr 30

Symphony: Why OpenAI's PRs Jumped 500% in 3 Weeks

5 min read

jidonglab

Apr 30

GPT Image 2 Inside Codex: My New Frontend Workflow

#ai #frontend #openai #productivity

6 min read

jidonglab

Apr 30

GPT-5.5-Codex vs 5.3: A 200-Task Bench Result

6 min read

jidonglab

Apr 30

Codex Is No Longer a CLI. Embed It in Your App.

6 min read

jidonglab

Apr 30

I Gave Codex My Mouse for a Day. Here's What Broke.

6 min read

jidonglab

Apr 24

OpenAI's Super App Play: Why Spud + Duct Tape Matter for Builders

#ai #openai #webdev #productivity

8 min read

jidonglab

Apr 24

Why OpenAI Shipped GPT-5.5 Just 6 Weeks After 5.4

#ai #openai #webdev #productivity

8 min read

jidonglab

Apr 24

OpenCode Hit 140K Stars. Why Terminal Agents Won 2026.

#ai #opensource #agents #webdev

7 min read

jidonglab

Apr 23

How a Markdown File Hit 16K Stars: Skills in 2026

#ai #opensource #agents #webdev

7 min read

jidonglab

Apr 21

MIT Tech Review Just Split Its \"Breakthrough\" List. That's the Story.

#ai #research #media #analysis

7 min read

jidonglab

Apr 21

Stellantis Just Outsourced Its AI Moat to Microsoft. Expect GM, Ford, and VW to Follow.

#ai #automotive #microsoft #enterprise

8 min read

jidonglab

Apr 21

44% of New Music on Deezer Is AI. Only 0.5% of Streams Are. Read That Twice.

#ai #machinelearning #music #opinion

7 min read

jidonglab

Apr 20

Adobe Just Made MCP an Enterprise Procurement Line Item

#ai #mcp #enterprise #agents

7 min read

jidonglab

Apr 20

Claude Opus 4.7 Hit 87.6% on SWE-bench. The Story Is What It Didn't Charge You.

#ai #claude #llm #engineering

7 min read

jidonglab

Apr 20

I Built an AI Newsletter for Myself. 11 Subscribers, 49 Posts, Zero Regrets.

#ai #automation #nextjs #indiehackers

6 min read

jidonglab

Apr 20

DataForge, Atropos, and a 30K-Token Guillotine: Reverse-Engineering Hermes 4's Training Stack

#ai #llm #machinelearning #opensource

8 min read

jidonglab

Apr 20

The Honest Hermes 4 Production Checklist (April 2026 Edition)

#ai #llm #opensource #agents

7 min read

jidonglab

Apr 20

66.5% of My Claude Code Tokens Were Wasted. A 200-Line Wrapper Got Them Back.

#ai #claudecode #cli #productivity

7 min read

jidonglab

Apr 18

Hermes 4's Tool-Calling Is Trained as a Separate Skill. Here's Why Your Agent Cares.

#ai #llm #agents #opensource

6 min read

2 Week Community Wellness Streak

1 Week Community Wellness Streak

Writing Debut

Want to connect with jidonglab?