Using Karpathy’s Wiki: How I Turned an AI Agent into a Stock Analyst

If you're still using AI as a chatbot to summarize emails that you won't read anyway, you're using it wrong.

I decided to use mine to research stocks for me as a hobby project.

The Context: News Overload

I spend too much time at the intersection of AI and stocks. For a while, my setup was basic: Hermes agent that summarizes 20+ newsletters into a morning email. It felt productive, but it was a pretty weak use case.

I was reading AI and stock news, but none of it was compounding. I'd find a great AI thesis on Monday in a podcast, and by Friday, I'd forgotten how it connected to the semiconductor bottleneck I read about on X three weeks ago.

Then Andrej Karpathy posted about the "LLM Wiki" — the idea that AI shouldn't just answer questions, it should maintain a persistent, evolving artifact of knowledge.

I realize I never retain any knowledge, and when I did, they were siloed. I needed a brain that remembers and connects the dots for me.

My setup:

Agent harness: Hermes Agent + custom skills
Model: GLM-5-Turbo
"Brain": Andrej Karpathy's LLM wiki

The Build: From Script to Strategy

My first prompt was the following:

Do a daily scan of: https://x.com/aleabitoreddit https://x.com/midascabal and compile a list of stocks that is a buy, speculative, on the radar, or any other category that you can think of. I want to get the signals vs noise from all the posts he's making. Post in "market-updates" channel in Slack

The result after a single prompt was pretty good, but mechanical with no reasoning.

After feeding it my own watchlist, my first daily market briefing and asking it to hunt for asymmetric plays, my 8am "market update" look like a big watchlist and news flash:

The first version of the scanner was like a freshman intern: eager, but not thinking strategically. It could find the tickers, but it couldn't tell me why they mattered in the grand scheme of things.

To move from a basic watchlist to a high-conviction research system, I had to build in these "missing" layers:

Macro Contextualization: Raw stock mentions are useless without context. I forced the agent to layer in "Strategic Headlines" — connecting a ticker mention to things like interest rate shifts, hyperscaler AI CAPEX (the "big spenders"), and specific supply chain bottlenecks.
The Conviction Hierarchy: Not all signals are created equal. I implemented a strict pruning system that categorizes tickers into a hierarchy: Conviction (strong thesis), High Signal (unusual activity), and Watchlist (waiting for a pullback or more data).
Recursive Source Discovery: The agent now actively hunts for its own replacements. It scans for new X accounts that aren't just "shilling" but are mapping out supply chains.
Multimodal Inputs: Since some of the best alpha is buried in an hour-long deep dive, I'm integrating transcripts from YouTube and Podcasts as signal sources.

Subscribe to my newsletter and follow me on X for AI tools and hacks.

Tweaks and fine tuning

After running for a few days, my lists, research, and scoring system started to pile up. It was difficult for my agent to find things.

I told V (my "vibecodingdad.com" assistant) to set this up, and it auto-organizes the content. Simply pointed V to Karpathy's LLM-wiki page.

Self improvements

You're an analyst for asymmetric trades with great fundamentals — what else should we add to our pipeline for research, signals, and macro?

You are constantly improving this as you go. always learning what works what doesn't, and always change things up. remember you're an agent that makes money from asymmetric trades and always hunting for new opportunities

Here is the architecture of how it actually works:

The Signal Scanner: Every morning at 7:45 AM, it scrapes the last 24h of tweets from deep-research accounts (think supply chain nerds, not engagement-baiters).
The Noise Filter: It extracts cashtags and classifies them by conviction. It's trained to ignore "I called this at $X" posts and focus on actual thesis-driven data.
The Convergence Engine: This is the magic. If three different smart accounts mention $INTC's CPU-to-GPU ratio or $SIVE's InP lasers, the agent flags it as a "Convergence Signal."
The Persistent Wiki: Instead of a fresh report every day, it maintains a research/index.md. It cross-links tickers. If a news story hits about the Strait of Hormuz, it doesn't just tell me "oil is up"—it tells me how that affects the margins for $VST based on their PPA deals with Amazon.

Feeding it news

Pressure testing

Adding new X follows

When asked to follow https://x.com/pelositracker on politician trades, V politely pushed back.

Adding new stocks

I also had it pressure-test $LMND (a favorite of a friend). V gave it a 4/9 Asymmetry Score, called out the shrinking surprise percentage, and told me the market was "pricing in perfection."

It's not a "yes-man." It's an analyst with an attitude.

Putting it all together:

Actionable Takeaway: Build for Compounding

If you're building AI agents right now, stop building "stateless" tools. A tool that forgets everything the moment you close the tab is just a fancy calculator.

What actually works:

Create a Log: Maintain an append-only timeline of what was researched.
Force Cross-Linking: Make your agent explain how New Ticker A affects Existing Thesis B.
Define Your "Anti-Goals": Tell your agent what not to look at (e.g., meme accounts, lagging indicators).

Most people overestimate what an AI can do in a single prompt, but they completely miss what an agent can do when it's allowed to build a library over a month.

Follow https://x.com/CalConviction for updates on my evolving stock analyst.

This story is published on Generative AI. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories.

Subscribe to our newsletter and YouTube channel to stay updated with the latest news and updates on generative AI. Let's shape the future of AI together!

Contents