Updated May 2026.
16GB Mac support, free to try.
Many of you asked for 16GB Mac support - Maibook v0.1.2 now supports Mac with 16GB RAM! Still free.
View on X
The flagship launch - on-device community of personalized AI agents.
The second brain is an AI snapshot of you. The third brain is an ongoing AI live streaming feed of your activity. Introducing Maibook - your third brain. A community of personalized AI agents expanding on your every activity, responding to your every request.
View on X
Standalone dark-mode viewer for Claude Code logs. 100% local, single HTML file.
Reading Claude Code logs in the terminal is painful. I got tired of scrolling through raw JSON to debug my agent sessions, so I built a standalone viewer. It turns your CLI history into a full UI. 100% local, single HTML file.
View on X
100% private Mac app using local MLX models - surfaces forgotten ideas across markdown files.
ahai - a 100% private Mac app using local mlx AI models - found my 1541 ideas across 13447 markdown files.
View on X
Case Study #10 from the Haixu educational comics series.
Haixu Case Study #10 - Butterflies
View on X
No-code web apps before “no-code” was a category.
I just launched meon into the wild. Make any web app experience come true, in minutes, no programming needed.
View on X
and earlier games - Smart Run (2013), Pop Rage (2014).
Four modes for generating personalized AI UI experiences - single-page HTML, web servers, native wrappers, and source-driven auto-generation. The thesis behind Maibook and the rest of MLAI’s lineup.
There are a few modes in which we can generate rich personalized ui experiences with AI: single page HTML (viewers, editors, transformers) with search/sort/filter; web servers for multiple views/file system access; native app that wraps either - for an even richer native feel, and private/local experiences. Rich personalized UI built on files is the future.
View on X
Four overlapping plan-and-execute systems - terminology overlap is genuinely confusing. Breakdown of each.
Claude Code now has FOUR overlapping "plan and execute" systems: Plan Mode (built-in), Superpowers Brainstorm (plugin), Superpowers Execute-Plan (subagents + worktrees), Agent Teams (new with Opus 4.6).
View on X
Engineers have a lot more, not less, work to do with AI doing most traditional coding tasks.
Engineers have a lot more (not less) work to do with AI doing most of traditional coding tasks. There are ∞ choices to make with AI outside of the model itself - choice and sequence of words, models, tool vs pre vs post process.
View on X
AI can be made to do anything barring domain/physical constraints, given the right context and agency.
I work under the conjecture that AI can be made to do anything barring domain/physical constraints, given the right context and agency. Increasingly less reliant on better models and more on code, natural language (prompt and context) and safe agency (tools).
View on X
Safer alternative to AI write-access: read-only agents on local files. References Simon Willison’s lethal trifecta.
Maibook was built as a safer alternative to the current AI norm of giving AI unmonitored write access on user's file system and web accounts - read @simonw's lethal trifecta to see why this is dangerous.
View on X
Most so-called “open” AI tools and models abuse the word. The MLAI bar: private, local, no sign-in, no cloud by default.
Most so called "open" AI tools and models abuse the word "open" in bad faith. An open AI tool is useful only if it is private, local, requires no sign-in/signup, and doesn't send data to the cloud by default.
View on X
How to use AI effectively: pick the best model, invest in the first prompt, debug by fixing the prompt and re-running.
How to use AI effectively - the missing manual.
View on X
The code + prompt layer matters more than the model in 2026. Small and open models in a good harness rival frontier models for most tasks.
The Best Harness Beats the Best Model.
View on X
Three-prompt technique - ask the LLM to argue one view, then the opposite, then resolve both.
Ask llm to convince me of one view, then in a second session ask it to convince me of the opposite view, then in a third session ask it to resolve both opposite viewpoints. Extremely effective.
View on X
The bitter pill: AI coding agents make the choices of an average developer. Hands off except when something breaks.
Succeeding with AI coding agents needs engineers to swallow a single bitter pill - the agents will most often make the choices of an average developer. We shd be hands off except when something breaks.
View on X
Use Claude Code skills as on-demand crash courses for any tech stack.
You can get very high quality cheatsheets/crash course/refresher/quick start with idiomatic best practices for a tech component. Just ask Claude Code to create a skill for it, with optional specific preferences.
View on X
Use /superpowers:brainstorm in default mode before plan mode - plan mode is overeager to generate the plan.
TIL how to brainstorm with claude code before planning. Use the official plugin /superpowers:brainstorm. Start with that in default mode (not plan mode, which is overeager to generate the plan and start building!).
View on X
Rewind to continue from an earlier point. Fork to branch with current context. Combine for unrelated branches from earlier points.
Claude code fork rewind mini-tutorial: rewind - to continue from an earlier point. fork - use current context, for an unrelated question. fork then rewind - fork from an earlier point.
View on X
Claude Code handles the hardest part of manual testing - clean repro steps in natural language.
Claude Code helps with the hardest part of manual UI testing - logging proper repro steps. I just do the clicks, tell it what I found in rough natural language, and CC has the codebase loaded, logs the test report nicely with all relevant details.
View on X
Why AI assistance can reduce productivity despite expectations. Applies far beyond coding.
]]>One of the most useful articles I have read in a long time. Explains the paradox of why AI assistance reduced productivity, in spite of strongly expecting the opposite. Applies to much more than just vibe coding.
View on X
“Your third brain.”
The second brain is an AI snapshot of you. The third brain is an ongoing AI live-streaming feed of your activity. Maibook is a private community of personalized AI agents expanding on your every activity, responding to your every request - on-device on your Mac or Windows PC.
maibook.app - Documentation - Download
Dark-mode viewer for Claude Code CLI chat sessions. Browse, search, and visualize AI coding sessions in a rich UI. Launched on Hacker News.
Get on Gumroad ($1+) - Show HN
AI-powered educational comics - a new visual learning experience by Human+AI. Topics span science, history, culture, philosophy, technology, and lifestyle.
| Category | Count | Examples |
|---|---|---|
| Science & Nature | 21 | Atom, Black Hole, Ecosystem, Photosynthesis, Volcano |
| History & Landmarks | 17 | Great Wall, Machu Picchu, Parthenon, Wright Brothers, Steam Engine |
| Philosophy & Culture | 8 | Shinto, Philosophy, Zen and Tea, Yoga, Society |
| Holidays & Events | 7 | Thanksgiving, Halloween, Christmas, Burning Man |
| Technology | 6 | AI World, Blockchain, Nanotechnology, Electric Cars |
| Lifestyle & Wellness | 6 | Mindfulness, Parenting, Sleep, Retirement |
| Mathematics & Medicine | 2 | Mathematics, Anesthesia |
Browse All on Gumroad - Read Case Studies
9 AI-generated educational comics on Kindle and Kindle Unlimited. By R. C. Anand.
| Title | Series | Price | Link |
|---|---|---|---|
| Thanksgiving | Holidays in the US | $2.99 / KU | Amazon |
| Halloween | Holidays in the US | $2.99 / KU | Amazon |
| Christmas Traditions | Holidays in the US | $2.99 / KU | Amazon |
| New Years Eve | Holidays in the US | $2.99 / KU | Amazon |
| Aurora: The Northern Lights | Standalone | $2.99 / KU | Amazon |
| Mindfulness | Standalone | $2.99 / KU | Amazon |
| Sleep | Standalone | $2.99 / KU | Amazon |
| Parenting | Standalone | $2.99 / KU | Amazon |
| Electric Cars Worldwide | Standalone | $2.99 / KU | Amazon |
“Rediscover what you meant to do.”
Native Mac desktop app using Apple MLX to scan markdown files and surface forward-looking concepts - features to build, content to create, ideas to explore. 100% local, privacy-first, Apple Silicon optimized. PySide6 + MLX.
ahai.app - Documentation - Buy on Gumroad ($5+) - Decision log: what, how, why
]]>Updated May 2026.
The papers everything else rests on.
| arxiv | Title | Byline | Tag |
|---|---|---|---|
| 1706.03762 | Attention Is All You Need | Vaswani et al. - 2017 - The Transformer. | Seminal |
| 2005.11401 | Retrieval-Augmented Generation for Knowledge-Intensive NLP | Lewis et al. - 2020 | Seminal |
| 2009.03300 | Measuring Massive Multitask Language Understanding (MMLU) | Hendrycks et al. - 2020 | Seminal |
| 2302.13971 | LLaMA: Open and Efficient Foundation Language Models | Touvron et al. - 2023 | Seminal |
| 2401.04088 | Mixtral of Experts | Jiang et al. - 2024 | Seminal |
Chain-of-thought to ReAct.
The track outside the pure-LLM mainstream - learning representations that predict, not generate.
| arxiv | Title | Byline | Tag |
|---|---|---|---|
| 2603.14482 | V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning | Meta FAIR - 2026 | Recent |
| 2602.11389 | Causal-JEPA: Learning World Models through Object-Level Latent Interventions | 2026 | Recent |
| 2603.19312 | LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels | 2026 | Recent |
| 2603.22281 | ThinkJEPA: Empowering Latent World Models with Large Vision-Language Reasoning | 2026 | Recent |
Combining learned + symbolic.
| arxiv | Title | Byline | Tag |
|---|---|---|---|
| 2410.22077 | Mapping the Neuro-Symbolic AI Landscape by Architectures | Hudson et al. - 2024 - The most-referenced paper in our working notes - a handbook on augmenting deep learning through symbolic reasoning. | Seminal |
Training at scale.
| arxiv | Title | Byline | Tag |
|---|---|---|---|
| 2503.14476 | DAPO: An Open-Source LLM Reinforcement Learning System at Scale | ByteDance / Tsinghua - 2025 | Recent |
Networked agents. Adjacent to Maibook’s design - how trust works in networks where agents (and humans) interact.
| arxiv | Title | Byline | Tag |
|---|---|---|---|
| 2603.11054 | A Survey on Quantitative Modeling of Trust in Online Social Networks | Song, Barber - 2026 | Recent |
Programs we follow.
| Source | Title | Byline | Tag |
|---|---|---|---|
| Sutton | The Bitter Lesson | Rich Sutton - The argument for scale over hand-crafted methods. | Essay |
| Numenta | Thousand Brains framework | Hawkins / Numenta - Cortical columns, voting-based representation. | Program |
| FAIR | V-JEPA & world-model program | Yann LeCun - Meta FAIR | Program |