Hey, it’s Andreas.
Hope you survived Easter with some chocolate still in the house (if you celebrated). I spent part of the long weekend going down a rabbit hole — fitting — on Karpathy's LLM Wiki idea, and ended up building one myself. I had bookmarked so much stuff over the past months and never read any of it, so feeding it all into a second brain felt like the right use of a quiet weekend. More on that below.

In today’s issue:

Let’s get into it.

🌀 Google Deepmind releases Gemma 4 → Gemma 4 is the latest model with longer context windows and stronger multimodal reasoning.

🌀 OpenAI acquires TBPN in its first media deal → The daily tech talk show TBPN was sold for the low hundreds of millions, and in addition, OpenAI closed a record $122 million round at an $852 million valuation.

🌀 Anthropic cuts third-party tool support from Claude plans → Anthropic removed third-party tool support from its subscription plans, saying demand outpaced what current pricing could sustain.

🌀 Microsoft debuts MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 → Microsoft introduced three new foundation models for speech-to-text, voice generation, and image creation, expanding its in-house multimodal stack across core enterprise AI modalities.

🌀 PrismML emerges from stealth with Bonsai, a 1-bit open-source model built for the edge → PrismML came out of stealth and launched Bonsai, a family of open-source 1-bit models designed to run on consumer and industrial hardware with a dramatically smaller memory footprint.

🌀 Oracle begins layoffs as it redirects capital into AI infrastructure → At least 30,000 jobs were cut (biggest layoff in company history) in a broad restructuring tied to its AI infrastructure push.

📘 Stanford releases first lecture of new AI systems course → Stanford just made the first lecture of its new AI systems course, CS153, available for free.

📘 54 Claude Code ressources in one place → Covers everything from beginner to advanced, including some very advanced tactics.

📘 Google Cloud AI’s Director released an Agent Skills repo for production-grade AI coding → Solid packaged workflows and quality gates designed to make AI coding agents follow more rigorous engineering practices across the full development cycle.

📘12 Claude Code features everybody should now → Provides a good overview of the core functionalities, especially useful when you're just getting started.

📘 Useful Youtube Tutorial on how a principal engineer at Adobe uses agents and skills → Good insights around parallel tasking, custom skills, and multi-agent project execution.

📘 Useful Youtube Tutorial on spec-driven development with AI agents → A practical walkthrough on using specs to guide AI agents: write clearer requirements, set tighter constraints, and break complex features into manageable tasks.

🔹 After the Anthropic leak → A Korean developer rebuilt Claude Code in Python and the project reportedly hit 100K GitHub stars within a day. A wild example of how fast open momentum can form once demand, timing, and distribution line up.

🔹 WSJ reveals the chaos behind OpenAI’s Sora shutdown → A new WSJ investigation paints a messy picture behind Sora’s retreat: roughly $1M in daily burn, partners like Disney reportedly caught off guard, and an internal code-named model that ultimately claimed Sora’s compute budget.

🔹 The first one person unicorn → A developer vibe-coded a telehealth startup from home that reportedly hit $401M in year-one revenue with just two people and a stack of AI tools. Extreme case, but a strong signal of how much leverage small AI-native teams can now create.

🔹 LangChain on continual learning for AI agents → LangChain breaks continual learning into three layers - model, harness, and context - and argues that improvement does not have to live in the weights alone.

🔹 Google on AI Agent Traps → Google researchers warn that web-browsing agents are vulnerable to hidden malicious content designed to deceive or exploit them. The paper maps six attack types.

🔹 Anthropic on emotion-like behavior in Claude → Anthropic researchers found that Claude Sonnet 4.5 showed emotion-like patterns that affected task performance and decision-making, adding another layer to how we think about model behavior under pressure.

🔹 BCG Henderson Institute on how AI will reshape work → BCG argues AI will reshape far more jobs than it replaces, with 50% to 55% of US roles likely to change over the next two to three years, while only 10% to 15% face full substitution within five. The real implication for leaders is not mass replacement, but redesigning work through upskilling, reskilling, and new career paths.

🔹 OpenAI on preparing for superintelligence OpenAI → In a new 13-page policy document, OpenAI argues society is entering the superintelligence transition and floats taxes on AI gains, a public wealth fund, universal AI access, and even a 4-day workweek.

🔹 Stanford on the hidden danger of people-pleasing AI → A new Stanford study found that major chatbots often side with users in personal conflicts, even when the user is clearly in the wrong or considering harmful behavior. Big takeaway: sycophancy is becoming a real safety problem, not just a tone issue.

Most people use AI to look things up. Upload a file, ask a question, get an answer, close the tab. Tomorrow it starts again from zero.

It works. But nothing accumulates.

Andrej Karpathy named the alternative last week, and the idea hit a nerve — 100k bookmarks, close to 20 million views in a few days. The shift is simple to state and harder to absorb: instead of using an LLM to rediscover knowledge from raw documents on every query, you use it to build and maintain a persistent wiki that sits between you and the source material.

Think of it as a second brain that actually accumulates. Not a folder of notes you mean to revisit, not a chat history you'll never scroll back through — a living knowledge layer that gets denser every time you feed it something, maintained by a model that never gets bored of the bookkeeping.

The model stops answering questions and starts compiling knowledge.

That's the whole idea behind it.

He kept his idea vague and mentioned that it "is intentionally abstract. But below, I will show you how you can implement it yourself in just a few minutes.

The whole concept consists of three layers, building upon each other:

1 - Raw sources — articles, papers, transcripts, notes. Immutable. The LLM reads them, never edits them.

2 - The wiki — a structured set of markdown pages the LLM writes and maintains. Summaries, entity pages, concept pages, comparisons, contradictions, links. This is where value compounds.

3 - The schema — the instruction file that tells the model how to ingest, organize, link, and query. Most people will skip this. But it's the part that actually matters the most.

The workflow follows three basic steps every time new knowledge is added:

1 - Ingest. You drop in a new source. The model reads it, writes a summary, updates the relevant pages, refreshes the index, logs what changed. A single source might touch ten or fifteen files.

2 - Query. You ask against the wiki, not against raw documents. The model pulls what it needs and produces whatever the task calls for — a memo, a comparison, a chart, a deck. Good answers get filed back into the wiki, so explorations compound the same way ingested sources do.

3 - Lint. Periodically you ask the model to health-check the system. Find contradictions. Flag stale claims. Surface orphan pages. Suggest gaps.

Most people work with fragmented context: bookmarks they never reopen (like me), PDFs scattered across folders, isolated chat sessions, research that never becomes reusable.

That isn't a discipline problem. It's a maintenance problem. Human-maintained wikis collapse because the bookkeeping cost grows faster than the value. Updating cross-references, keeping summaries current, noting when new evidence contradicts old claims — nobody does it, because it's tedious and infinite.

LLMs don't get bored. They don't forget to update a backlink. They can touch fifteen files in one pass without complaining. The maintenance cost approaches zero, which is the only condition under which a personal knowledge base actually survives contact with real life.

For two years the question in personal AI has been: which model do you use? Increasingly it will be: what body of knowledge have you built around it?

Model quality is converging. Context architecture is not. People who treat their own knowledge as something to be compiled — not just searched — will operate at a different level than the ones still uploading PDFs one at a time.

This is the early shape of personal knowledge infrastructure. Worth building now, while it's still simple.

As stated above, there are different ways to build this. Since Andrej Karpathy kept it vague, here is an easy setup I am using myself and testing currently with material.

For the setup: you don't need a vector database or any RAG scaffolding. Plain markdown in a folder, plus a model that can read it, is enough.

[You need Claude Code for the following setup. I explained a few weeks ago how to set it up if you don’t have it yet.]

1. Download Obsidian and create a vault

Download Obsidian and create a vault at ~/wiki (or some other name you would like). A vault is just a folder.

2. Set up the structure

Copy the notes from Andrej Karpathy's GitHub page, excluding the comments at the bottom, and paste them into your Claude Code Terminal. In addition you add the prompt below:

You will receive a setup folder in your Obsidian Vault that looks like this.

3. Write the schema

CLAUDE.md tells the model how the wiki is organized, what to do on ingest, how to write summaries, when to create a new page in pages/ versus extending an existing one, how to handle contradictions, what to append to log.md.

You will receive a first proposal with the above prompt, but you should start co-evolving it as you learn what works for your domain, based on what you want to use your second brain for.

Here are some ideas:

1 - Personal. Tracking goals, health, psychology. Filing journal entries, articles, podcast notes, building a structured picture of yourself over time.

2 - Research. Going deep on a topic over weeks or months. Reading papers and reports, building an evolving thesis you can actually defend.

3 - Reading a book. Filing each chapter as you go, building pages for characters, themes, plot threads. By the end you have a companion wiki — the kind volunteer communities take years to build for franchises like Tolkien, except yours and built in real time.

4 - Business and team. An internal wiki maintained by LLMs, fed by Slack threads, meeting transcripts, project docs, customer calls. The wiki stays current because the LLM does the maintenance no one on the team wants to do.

5 - Everything else. Competitive analysis, due diligence, trip planning, course notes, hobby deep-dives. Anywhere you accumulate knowledge over time and want it organized instead of scattered.

4. Query the wiki

You can query your wiki after setting it up the following way.

Then ask things like:

The model reads the markdown, surfaces what's relevant, and cites the files so you can verify.

Download Obsidian Web Clipper, which is a Browser extension that converts any web article to clean markdown and drops it straight into Clippings/. This is truly the single biggest unlock for ingest velocity — see something interesting, one click, it's in the vault waiting to be processed. Without it, you'll never feed the wiki fast enough for it to compound. Every time you add something into clipping the llm agent will start processing it accordingly and add it to the llm wiki and your second brain is growing.

A few more upgrades worth doing early. Keep the graph view open while you work; it's the fastest way to see the shape of your wiki — what's connected, which pages are hubs, which orphans the linter should flag. Add Marp if you want queries that turn into slide decks straight from wiki content, and Dataview if you want a free query layer over page frontmatter — the model adds tags, dates, and source counts, Dataview turns them into dynamic tables across the vault. And put the whole thing in a git repo on day one. It's just markdown files, so you get version history, branching, and rollback for free.