Research

Callum (Wanderloots) explains knowledge graphs from first principles (nodes, edges, triples), shows how his 3-year Obsidian vault works as a personal knowledge graph, and argues for keeping a second, AI-maintained "LLM Wiki" vault — per Andrej Karpathy's recent pattern — as shared structured memory across all your AI tools. The pitch: plain RAG fails when answers live *between* documents, so a persistent, interlinked markdown wiki that agents build and maintain outperforms per-tool siloed memory.

Colibri is a ~1,300-line pure-C inference engine that runs GLM-5.2 — a 744B-parameter mixture-of-experts model — on a 25 GB-RAM laptop by pinning the ~10 GB dense core in memory and streaming the ~370 GB of routed experts from SSD on demand, exploiting the fact that only ~5% of parameters activate per token. The catch is speed (0.1–1 token/second), so it's useful only for overnight or air-gapped jobs, but it signals a broader shift: antirez's ds4, llama.cpp, and vLLM are all starting to treat the SSD as a memory tier.

A 76-second build video: mount a French cleat and a stop strip to the back of an 8-ft 2x6, pre-drill and screw every 5 inches, then crosscut on those marks to mass-produce 19 small cleat-mounted blocks for garage hooks. It's a batch-production trick for anyone who already has a French cleat wall and wants cheap, rearrangeable hook mounts.

Ziroth interviews Heat Geek founder Adam Chapman on why their heat pump installs average ~430% seasonal efficiency (SCOP 4.3) against a UK average of ~2.8, walking an "efficiency pyramid" from SCOP 2.5 to 5.5. The gains come almost entirely from system design and operation — low-and-slow running, good pipework, weather compensation, oversized low-temperature hot water cylinders, and accurate heat-loss-based sizing — not from the heat pump unit itself.

A year-end roundup of six MCP servers the creator considers most impactful for AI-assisted development: Context7 (live library docs), Docker MCP Toolkit (a two-tool gateway that fixes context bloat), the shadcn registry MCP (UI components), Google's new managed MCP servers (Maps, BigQuery, Compute, GKE), Notion MCP (workspace management), and Supabase MCP (database/schema automation). The recurring theme is context-window economics — the best MCPs either inject precisely the right context (docs, component specs, schemas) or reduce how many tool definitions the agent must carry.

The creator tests Qwen 3.6 27B (dense) locally in Zed on three coding tasks — a to-do app (5 min, pass), a sorting-algorithm visualizer (20 min, pass), and a plan-then-implement Kanban board (~1 hour, pass first try) — and concludes it's good enough to use regularly as a complement to paid models. The surprising setup finding: the larger 35B MoE variant runs ~3x faster on his hardware, but the smaller 27B dense model benchmarks better on software-development tasks, so he picks the slower one.

Stop Prompting Claude. Use Karpathy's Method Instead.

TradingAgents is an Apache-2.0 Python framework (from a UCLA/MIT-affiliated research paper, maintained by Tauric Research) that simulates a hedge fund as a LangGraph pipeline of LLM agents: four parallel analysts, debating bull/bear researchers, a trader, a risk team, and a portfolio manager with final approval. The video argues its value is traceability — every decision comes with readable reports and debate transcripts — while cautioning that it burns real token money, only trades against a simulated exchange, and is explicitly not financial advice.

The video shows two ways to run the Claude Code harness on non-Anthropic models: locally via Ollama (`ollama launch claude` against a downloaded open-weight model) and via OpenRouter's free model tier (pointing `ANTHROPIC_BASE_URL` at OpenRouter in `settings.local.json`). Both work but come with real trade-offs — slow local inference on consumer hardware, silent context truncation, missing native web search, daily rate limits, and a hidden paid-Haiku fallback unless every model variable is overridden.

Coder announces Mux (pronounced "MX" in the video, short for multiplexer), an open-source desktop and web app for running multiple AI coding agents in parallel, locally or over SSH into remote development environments. It's provider-agnostic, integrates with Coder workspaces for enterprise governance, and Coder is actively recruiting early design partners.

The video explains Andrej Karpathy's "LLM wiki" pattern — instead of RAG re-retrieving from raw documents on every query, an LLM agent incrementally builds and maintains a persistent, interlinked markdown wiki (raw sources → wiki → schema file) so knowledge compounds across sessions. The presenter then builds one live in Claude Code with Opus 4.6 for trading-strategy research, showing ingestion of eight transcripts, wiki-link navigation in Obsidian, and automatic backfilling of the wiki from web searches when a query exceeds what's stored.

The video argues Claude Code usage limits are usually a context-hygiene problem, not a limits problem: MCP tool schemas, bloated CLAUDE.md files, verbose skills, and misconfigured settings silently consume tens of thousands of tokens before you type a message, and every message re-pays that cost. Fixes include pruning MCP servers (or swapping them for CLIs), progressive disclosure in CLAUDE.md, settings tweaks (auto-compact threshold, bash output length, deny rules), and daily habits (`/clear` between tasks, plan mode, editing instead of correcting, right-sizing the model) — plus a promo for the creator's free "context audit" skill and paid marketplace waitlist.

How to Replace a Casement Window Operator

A hand-tightened treenet perimeter that feels tight goes slack the moment you weave into it perpendicularly, so the video builds a gear-free tensioning system from paracord: a prusik-style "press loop" gripping the perimeter rope, alpine butterfly anchor points, and repeated pull–lock–reset cycles with slipped half hitches. The final connection is dressed into a double fisherman's, accepting a small tension loss in exchange for a compact, snag-free knot.

Matt Pocock argues AI-assisted coding accelerates software entropy, and demonstrates his "improve-codebase-architecture" Claude skill, which scans a repo for shallow modules and proposes "deepening opportunities" using a shared glossary (module, interface, seam, adapter, depth, leverage, locality) drawn largely from Ousterhout's *A Philosophy of Software Design*. The workflow is deliberately human-in-the-loop: the agent acts as a "tactical programmer" surfacing candidates and grilling you with design questions, while you make the strategic calls and optionally export the agreed design as an issue for an AFK agent.

A tick-country homeowner walks through practical, low-cost tick prevention: permethrin-treated footwear, strategic clothing, tick tubes that weaponize mice against ticks, packing tape and removal tools for ticks already on you, and diligent daily body checks. Most claims hold up against public-health sources; the notable outlier is his advice to wear minimal clothing so you can feel ticks before they hide, which runs counter to standard guidance.

The video distills Anthropic engineers' published guidance on Claude Code into four rules: prompt reusable skills instead of writing ad-hoc prompts, invest in the tools/scripts layer of skills rather than just the prompt text, build small composable skills instead of one monolith, and update skills after every session so they compound. It's an accessible, mostly accurate summary of Anthropic's Agent Skills material, aimed at non-technical viewers.

The video shows how to combine Nous Research's open-source Hermes Agent with AionUi, a free cross-platform desktop "cowork" app, to build a local "agentic OS" where multiple autonomous agents handle spreadsheets, file organization, report writing, browser tasks and small coding jobs in parallel. It is a setup walkthrough plus three cherry-picked demos (financial dashboard in Excel, desktop file cleanup, a small coin-collecting game), with links promised in the description.

WorldofAI rounds up a week of Hermes Agent updates: background computer use that works with any vision-capable model, Alibaba's Qwen 3.6 Plus temporarily free on Nous Portal, LightPanda as an integrated open-source browser backend with Chrome fallback, a revamped Kanban web dashboard for multi-agent orchestration, and a `/goal` command for long-running autonomous objectives. The information is broadly accurate but shallow — feature names and platform availability are partly wrong or outdated versus the official docs, so verify against the docs before adopting.

A timelapse of hand-weaving a ~95 sq ft "time trial" tree net (floor, sidewalls, and entry portal) between four trees, with an on-screen timer to measure exactly how long it takes and how much material it consumes. Final tally: ~25 hours total (17 for the floor alone), 2,500 ft of paracord, and 120 ft of rope, working out to 3.8–5.6 sq ft per hour — rates the creator flags as faster than normal because he had already woven a net in the same spot.

Graphify builds a persistent knowledge graph of a project (code via local tree-sitter parsing, audio/video via local Faster-Whisper, docs via one-time Claude subagent extraction) that Claude Code auto-loads at session start instead of re-reading files every session. The presenter's own 10-question test found only ~7–8% token savings but subjectively better answers, and he argues the repo's 71.5x claim measures an unrealistic paste-everything baseline — the tool's real value is on large, mixed code+docs projects, research vaults, and long sessions.

Water electrolysis is not one clean H₂O-splitting reaction but two separate electrode reactions — acid forms at the anode and base at the cathode — and exploiting that separation with ~$1 DIY ion-exchange membranes (ground water-softener resin + PVC cement) unlocks hardware-store versions of serious electrochemistry. The video demonstrates four builds on that foundation: making hydrochloric acid and lye from table salt, closed-loop refining of magnetite ore into pure plated iron, an iron sulfate flow battery with tank-scalable capacity, and a hydrogen generator whose output is self-pressurizing.

A 24-second short outlining a speed-run sequence for building a treenet (spacenet): scout trees, erect a temporary scaffold for access, anchor and ratchet a tight perimeter rope using a paracord-only block and tackle, then fill it with a chaos weave while keeping tension in every knot. The steps match established treenet practice, but the "done by sunset" framing dramatically understates the time and material a real first build takes.

DeepSeek V4 ships as MIT-licensed open weights with a 1-million-token context window, using three stacked compression techniques (token-level compression, 128:1 Heavily Compressed Attention, and Compressed Sparse Attention) to cut KV-cache memory by ~90% versus its predecessor. The Pro model roughly matches recent frontier closed models at a fraction of the price, though it's text-only, degrades near the context limit, and some of the video's claims (Engram usage, beating Gemini 3.1 Pro on recall) are shakier than presented.

B. Dylan Hollis makes a 1990 no-bake lemon pie that layers four lemon sources — 6 oz of lemon Jell-O in boiling water, fresh zest and juice, cream whipped with lemonade powder, and sour cream — then chills it in a graham cracker shell of his own choosing, since the recipe never names a vessel.

CloakBrowser is an open-source stealth Chromium fork claiming to patch fingerprint surfaces at the C++/binary level while keeping the native Playwright/Puppeteer API, positioned as a free replacement for $50–$300/month anti-detect browsers. It's real and technically credible, but the current repo has shifted to a free-old-build / paid-current-build model, undercutting the video's "suddenly free" pitch.

Anthropic announced (at Code with Claude 2026, May 6) a compute deal giving it SpaceX's Colossus 1 data center — 300+ MW and 220,000+ Nvidia GPUs — and used the capacity to immediately double Claude Code's 5-hour rate limits on Pro/Max/Team, remove peak-hours throttling, and raise Opus API rate limits (tier 1: 30K→500K input TPM, 8K→80K output TPM). The video's practical read: retest rate-limit-killed workflows, use Opus more freely, and treat multi-agent and 1M-context API workloads as newly viable — while noting Anthropic's longer arc toward orbital compute and community-trust plays like covering local electricity price hikes.

Project Farm tests eight 21" mower blades (Oregon high-lift and G5, MaxPower 3-in-1 and mulching, 8TEN high-lift and mulching, Arnold Xtreme, Craftsman OEM) for airspeed, noise, bagging weight, leaf mulching, discharge, grass mulching, and rebar-impact durability. The Oregon G5 tungsten-carbide mulching blade wins overall (average finish 1.6) but demands significant engine power; both MaxPower blades are close runners-up, with the MaxPower mulching blade surviving rebar impact best.

Nate Herk walks through implementing Andrej Karpathy's viral "LLM wiki" idea: paste Karpathy's gist into Claude Code and it scaffolds a plain-markdown knowledge base (`raw/` sources, `wiki/` pages, index, log, CLAUDE.md) that the LLM auto-organizes, cross-links, and lints — no vector database or embeddings. He demos it on his 36 YouTube transcripts and the AI 2027 article, uses Obsidian purely as a graph/markdown viewer, and closes with an honest caveat that the pattern suits hundreds of well-indexed pages, not millions of enterprise documents.

B. Dylan Hollis bakes a 1919 recipe leavened with ammonium carbonate ("baker's ammonia"), the same compound used in smelling salts, which he withheld from upload since 2021 for fear viewers would think it dangerous.

AirLLM is an open-source (Apache-2.0) Python library that runs 70B-parameter models on ~4GB of VRAM by loading one transformer layer at a time from disk instead of holding the whole model in memory. The video's claims are essentially accurate but omit the critical cost: inference is roughly 10× slower than conventional serving, making it suitable for offline batch tasks rather than chat.

Austin Marchese walks through nine Claude Code plugins spanning output compression (Caveman), research (Exa + Firecrawl), workflow discipline (Compound Engineering), media generation (Higgsfield, the sponsor), Anthropic's official plugins, multi-model fallback (OpenAI's Codex plugin), his own BuildPartner.ai, edit/search acceleration (Morph), and token-cost observability (CodeBurn). The recurring thesis is that most token spend is mechanical rather than reasoning, so compressing output, improving input quality, and offloading mechanical work saves both money and time.

Fireship debunks a set of programmer time-wasters: chasing bleeding-edge tech, paradigm dogma, over-applied Clean Code/DRY, 100% test coverage targets, premature performance and infrastructure optimization, and over-reliance on AI coding tools. The recurring prescription is to default to boring proven technology, abstract only after repetition hurts, and optimize only when production evidence demands it.

The video recommends six GitHub repos to layer onto Claude Code: Superpowers (dev methodology with skills/agents/hooks), Everything Claude Code (hackathon-winner's full config), claude-flow (agent swarms, advanced), Open Design (open-source Claude Design alternative), obsidian-skills (vault bridge), and a Karpathy-derived CLAUDE.md of four coding principles. Install order matters: Superpowers first, claude-flow last, and most install with a single `/plugin install` command.

A 32-item anthology of Claude Code tips spanning beginner habits (CLAUDE.md via `/init`, context hygiene with `/context`/`/compact`, plan mode), intermediate workflow (subagents on cheap models, custom skills, screenshot self-check loops, `/rewind`, hooks) and power-user features (git worktrees, `/loop`, remote control, agent teams, Context7 MCP, permission allowlists instead of `--dangerously-skip-permissions`). The individual tips are largely accurate and verifiable against official docs, but several depend on rollout status, plan tier, or experimental flags the video doesn't mention, and at least one ("ultrathink") is already obsolete.