Research

Why LLM Wiki? 🧠 Future Of Knowledge For Agentic AI & Humans

A budget method for batch-producing French cleat hook mounts: glue and nail a cleat plus a stop strip to an 8-ft 2×6, then cut it every 5 inches to yield 19 modular blocks. The blocks hang on an existing French cleat wall so garage hooks can be rearranged freely.

The Best MCP Servers of 2025

A creator benchmarks the locally-run Qwen 3.6 model on easy/medium/hard coding tasks and finds the 27B *dense* variant beats the larger 35B mixture-of-experts model on coding despite being smaller. Qwen 3.6 one-shots a todo app, a sorting visualizer, and a planned-out Kanban board with no errors, which he calls an outstanding result for a small local model.

TradingAgents is an Apache-2.0 Python framework that simulates a hedge-fund's org chart with one LLM agent per role (analysts, bull/bear researchers, trader, risk team, portfolio manager) wired together on LangGraph. Its real value is as a clean, fully auditable reference implementation of a multi-agent LLM decision graph — not as a money-making trading bot.

Two ways to swap Claude Code's "engine" for free/cheap open-weight models: run one locally via Ollama, or route through OpenRouter's free/cheap models by overriding Claude Code's API environment variables. Both work and are allowed by Anthropic, but "free" is qualified — local needs hardware, the cloud paths have rate limits or per-token costs, and quality/speed trail the native Claude models.

The video walks through Andrej Karpathy's "LLM wiki" pattern: instead of RAG retrieving chunks at query time, an LLM incrementally builds and maintains a persistent, interlinked set of markdown files (raw sources → wiki → schema) that compounds over time. The creator then live-builds one for trading strategies in Claude Code with Opus 4.6, demonstrating ingest, query, and web-search backfill.

A Claude-Code-specific guide to cutting "invisible" starting context — MCP servers, bloated CLAUDE.md files, verbose skills, and stale `settings.json` defaults — plus day-to-day habits like `/clear`, plan mode, and right-sizing the model. The advice is mostly aligned with Anthropic's own cost guidance, though the video overstates MCP token cost for current versions and wraps the takeaways around a free-skill lead magnet and a paid marketplace waitlist.

How to Replace a Casement Window Operator

Matt Pocock argues AI has accelerated "software entropy," turning codebases into balls of mud, and offers his `improve-codebase-architecture` skill as a cure that scouts for "deepening opportunities" using Ousterhout's deep-module vocabulary. The catch: the skill only surfaces candidates and grills you — the human must make every strategic call, so it's a thinking aid wrapped around classic software-design fundamentals, not an autopilot.

How I Stopped Getting Tick Bites Completely - Everything You NEED to Know

How Anthropic Engineers ACTUALLY Prompt Claude Code

The video shows how to pair Nous Research's self-hosted Hermes Agent with the open-source AionUi ("Ion UI") desktop app to create a multi-agent "agentic OS" that runs autonomous tasks — Excel dashboards, file organization, research reports, simple games — locally on your machine. It's a hands-on install-and-demo tutorial, light on caveats, that positions the Hermes + AionUi combo as a way to deploy several autonomous agents simultaneously with visual real-time oversight.

WorldofAI walks through a wave of Hermes Agent updates from Nous Research: background computer use on macOS (powered by Cua), a free-for-now Qwen3.6-Plus model in the Nous portal, a Light Panda browser backend, a revamped Kanban for multi-agent orchestration, and a new `/goal` autonomous-objective command. Most marquee features are macOS-only and/or beta, and the video carries a paid Tiny Fish sponsor segment, so treat the enthusiasm accordingly.

Hand Woven Rope Tree House (Tree Net) Timelapse

Graphify scans a project once, builds a queryable knowledge graph (code via local AST parsing, audio/video via local Whisper, docs via a one-time Claude pass), and injects a graph summary at the start of every AI session so the assistant makes a few targeted reads instead of re-reading everything. The presenter's own A/B test found modest token savings (~7–8%) but noticeably better answer quality, and he debunks the repo's "71.5x" claim as measuring an unrealistic baseline.

DeepSeek-V4 is an open-weights MoE model (1.6T-param Pro, 284B Flash) shipping a 1M-token context window with three stacked KV-cache compression tricks that cut long-context memory ~90% versus V3.2. It roughly matches recent frontier models at a fraction of the price, but it's unimodal, degrades near its context limit, and parts of its training aren't fully understood even by its authors.

Dangerously Lemon Pie (1990)

CloakBrowser is an open-source (MIT wrapper + free proprietary binary) stealth Chromium that applies fingerprint patches at the C++/source level while keeping the native Playwright/Puppeteer API, letting existing scrapers migrate with a one-line import change. It claims to pass 30+ bot-detection tests and a 0.9 reCAPTCHA v3 score, positioning it as a free replacement for $50–$300/mo anti-detect browsers — but it doesn't solve CAPTCHAs or bundle proxies, and you assume all legal risk.

Anthropic doubled Claude Code's 5-hour rate limits, removed peak-hours throttling for Pro/Max, and sharply raised API rate limits for Opus models — all backed by a new compute deal giving it access to SpaceX/xAI's Colossus 1 data center (300+ MW, 220,000+ NVIDIA GPUs). The video walks through what each change unblocks for builders and flags a speculative future interest in "orbital" AI compute.

Project Farm benched eight 21" walk-behind blades (Oregon high-lift & G5, MaxPower 3-in-1 & mulching, 8TEN high-lift & mulching, Arnold Extreme, Craftsman OEM) across airflow, bagging, leaf-mulching, discharge, mulching, and a rebar-impact durability test. The Oregon G5 Gator mulching blade took the overall win (1.6 average) on the strength of bagging and mulching, but it demands the most horsepower of any blade tested — repeatedly stalling the engine — while the two MaxPower blades were the close runners-up and the MaxPower mulching blade was the toughest against rebar.

The video walks through Andrej Karpathy's April 2026 "LLM wiki" idea: instead of RAG, you let Claude Code ingest raw documents and build a structured, interlinked folder of markdown files that compounds over time, optionally visualized in Obsidian. It's cheap and simple at the scale of ~100 sources, but the creator openly admits it doesn't scale to enterprise/millions-of-documents use.

A self-described "never uploaded" 2021 clip in which B. Dylan Hollis bakes 1919 "ammonia cookies" leavened with ammonium carbonate (baker's ammonia), the same compound found in smelling salts.

AirLLM is an open-source Python library that runs 70B-parameter models on low-VRAM hardware (down to 4GB) by streaming the model into the GPU one layer at a time from disk, plus flash attention to flatten memory on long inputs. It makes large models technically runnable and fully private, but the video skips the catastrophic speed cost of constant disk loading.

9 Claude Code Plugins to Build 10x Faster

6 Claude Code GitHub Repos That Change Everything

32 Tricks to Level Up Claude Code in 16 Mins