Jonathon's AI Wiki

Cross-Topic Connections

6 min read

Added 2026-07-03 (connection-density campaign — batch of 10)

  • AI Website Builders vs. the Claude Code Design Stack — decision framework + compose paths between hosted builders (Lovable / Bolt / v0) and the Claude-side design-skill/template stack; speed-to-app vs control, brand fidelity, lock-in, cost shape.
  • The Video Marketing Measurement Loop — AI video made variants abundant, so the bottleneck moved to measurement: produce (HyperFrames) → predict (virality predictor) → test (hook A/B) → optimize (ad-creative testing) as a weekly loop.
  • The Marketing Prompt Stack — production marketing prompting = voice patterns × skill packaging × evals; the pipeline from ad-hoc prompt to governed, versioned, quality-gated asset.
  • Publish Your Wiki for AI Citation — an LLM-maintained wiki published to the web is a GEO asset (structured, dense, fresh, interlinked = exactly what AI search cites); what to publish, what to keep private, how to verify citation.
  • Self-Improving Agent Loops — GAPA, ACE, and Who Judges — GAPA, Fable 5’s ACE memory loop, and Anthropic’s recursive-self-improvement org loop converged on generate → assess → persist → apply; every failure mode concentrates at the assess/persist boundary the verifier-first discipline governs.
  • How Autonomous Is Your Agent? — three instruments (Anthropic field telemetry, MirrorCode benchmark ceiling, Nate Jones’s production class ladder) composed into a rung-by-rung self-assessment of autonomy exercised vs possible vs warranted.
  • Wire It or Loop It — the build decision every automation task hits: deterministic n8n workflow vs Claude Code agent loop vs plain prompt vs hybrid; verification availability, cost curves, and security tax as the separators.
  • Pocket Access to Always-On Agents — three patterns for reaching an always-on agent from your phone (SSH/tmux over a tailnet, native companion app, chat-channel plugins) compared on setup, notification fidelity, and security surface.
  • The Voice-Agent Lead-Qualification Stack — the local-business inbound-call funnel assembled from three layers: voice-model choice, n8n/Vapi orchestration, and GoHighLevel’s Voice AI public API; who owns what and a minimal build order.
  • The Agent Hardware Frontier — dedicated agent hardware converging on the CLI agent (R1 voice pendant → OpenClaw ports → Cyberdeck’s Linux CLI netbook targeting Claude Code); the hardware-category thesis and what to buy/build today.
  • Sovereign Agent Runtimes — owning an agent decomposes into three separable layers (runtime, model, data path); self-hosted Hermes/OpenClaw + Venice’s TEE inference give provable confidentiality, at the cost of inheriting the whole security burden a hosted provider would carry.
  • Hardening an Agentic Prompt — the two layers that harden an agent against injection: the technique layer (truthful context, untrusted content in the user turn, thinking-on validation) and the harness layer (hooks/permissions/sandboxing); both converge on “the prompt is not the security boundary.”
  • HTML Is the Canvas — HTML is now the shared authoring medium for web pages, decks, AND video (HyperFrames renders MP4 from HTML); one design skillset produces all three, with the timing layer the only video-specific add-on.
  • Image Models Feed the Video Pipeline — the AI video pipeline begins with a still: GPT Image 2 and Nano Banana are interchangeable, A/B-able front-ends that generate the character sheet / storyboard grid that image-to-video models (Seedance, Kling) animate; the image-model pick is an upstream decision that propagates into every downstream frame.
  • Karpathy’s Workflow, Realized in Claude Code — Karpathy’s tool-agnostic framework (tool ladder, verifiability, jagged intelligence, vibe→agentic, understanding-as-bottleneck) maps point-by-point onto concrete Claude Code features; the sharpest signal is two independent convergences where Karpathy and Anthropic derive the same rule.
  • Two Shapes of the Personal Agent — Hermes and Claude Code run nearly the same five-step loop in opposite shapes: Hermes the always-on daemon that ends every turn with a memory-update pass, Claude Code the session-based invocation whose durable state is the folder + re-injected CLAUDE.md; the documented composition is CC-as-Hermes-builder over one shared filesystem.
  • Cheap-Executor Delegation — Four Shapes of the Same Cost Lever — advisor, coordinator, boss/worker/checker, and cross-vendor plugin-executor are one economic lever (keep the frontier a minority of tokens) in four topologies; a side-by-side map of who drives, where the executor lives, whether verification is independent, and the measured cost win of each (~63% price/~92% score · 2.5×/3× · ~10× · ~60%).

Added 2026-07-16

  • Structural Safety Enforcement — Why 2026’s Agent Builders Stopped Trusting the Prompt — Two independently-built Hermes projects (RUDR9’s toolset restriction, the hackathon-winning Custodian kernel) converged on “remove the capability, don’t just instruct against it” the same month Anthropic’s Agentic Misalignment research quantified why prompted-only safety has a real failure rate — including LLM judges gaming their own evaluation verdicts.

Added 2026-07-29

  • The Find-Fix Asymmetry — AI Offense Scales Faster Than Human Defense — Five otherwise-separate July 2026 stories share one pattern: AI-powered finding (bugs, exploits, algorithm weaknesses) scales with inference compute while fixing scales with human organizational capacity. Mythos cryptanalysis + Epoch’s exploit-dev confirmation on the finding side; ProPublica’s Microsoft “mad dash” and Hugging Face’s 17,000-event triage on the fixing side; NVIDIA’s Open Secure AI Alliance as the institutional response; Mozilla’s find-and-patch harness as the working template that collapses the asymmetry locally.

36 items under this folder.