This is f*cking insane. Tesla just teased FSD v15 and asked what if we are just driving in a video game. The man who built Tesla Autopilot, Andrej Karpathy:
"The network is being trained every mile that's driven
The most important lessons come from things you'd never put in a simulator"
In 29 minutes, he shows how a fleet of cars became one of the biggest AI training labs on Earth
Every Mile → Real Data → Edge Cases → Next Version → FSD v15
Most people think AI gets smarter in a lab. Tesla's gets smarter in traffic
Worth more than any $2,000 AI course
Bookmark and watch it today
Then read how to build systems that learn from every run: Harness, Loops, Graphs ↓
Your agent is not autonomous
You are its loop
You check every step because nothing else in the system can
I broke down the 7 levels from babysitting agents to running a fleet ↓ https://t.co/4RacdXA4nE
This is f*cking insane. Andrej Karpathy, the man who built Tesla Autopilot and co-founded OpenAI:
"Stop treating it like a chatbot
Treat it like a system that can improve itself"
In 2 hours, he compresses 8 years of building AI into one free lecture
Agents → Loops → Graphs → Self-Improving Systems
People spend $15K on bootcamps that teach less than this
Bookmark and watch it this week
Then read the self-improving systems guide: Harness, Loops, Graphs ↓
Your agent is not autonomous
You are its loop
You check every step because nothing else in the system can
I broke down the 7 levels from babysitting agents to running a fleet ↓ https://t.co/4RacdXA4nE
This is f*cking insane. Andrej Karpathy could have charged $10,000 for this course. He put it out for free:
"Let's build it from scratch
So you understand every single piece"
In 2.5 hours, the man who built Tesla Autopilot and co-founded OpenAI builds it all from zero
Scratch → Neural Net → Language Model → Agents → How It Actually Works
The gap between people who watch it this week and those who save it for later is not 2 hours
Worth more than any $10,000 bootcamp
Bookmark and watch it this week
Then read how to put it to work: Harness, Loops, Graphs ↓
This is f*cking insane. Andrej Karpathy spent 70 minutes showing how top AI users actually work with LLMs
And most people are overcomplicating it
Prompts → Context → Tools → Workflows → How the Top 1% Work
He co-founded OpenAI and led AI at Tesla. This is the hour he spent on how it's really done
Worth more than any $2,000 AI course
Bookmark and watch it today
Then read how to set up your own: Harness, Loops, Graphs ↓
Your agent is not autonomous
You are its loop
You check every step because nothing else in the system can
I broke down the 7 levels from babysitting agents to running a fleet ↓ https://t.co/4RacdXA4nE
This is f*cking insane. Andrej Karpathy could have charged $10,000 for this course. He put it out for free:
"Let's build it from scratch
So you understand every single piece"
In 2.5 hours, the man who built Tesla Autopilot and co-founded OpenAI builds it all from zero
Scratch → Neural Net → Language Model → Agents → How It Actually Works
The gap between people who watch it this week and those who save it for later is not 2 hours
Worth more than any $10,000 bootcamp
Bookmark and watch it this week
Then read how to put it to work: Harness, Loops, Graphs ↓
Your agent is not autonomous
You are its loop
You check every step because nothing else in the system can
I broke down the 7 levels from babysitting agents to running a fleet ↓ https://t.co/4RacdXA4nE
This is f*cking insane. Andrej Karpathy:
"Prompting is going away. Delete everything and keep the graph"
In 1 hour, he breaks down how to build graphs from scratch and make your agents work and improve on their own
Prompts → Delete Everything → Keep the Graph → Agents That Run Themselves
Not the newest session, but still one of the best ever recorded
Bookmark and watch it today
Then read the full graph guide: Harness, Loops, Graphs ↓
Your agent is not autonomous
You are its loop
You check every step because nothing else in the system can
I broke down the 7 levels from babysitting agents to running a fleet ↓ https://t.co/4RacdXA4nE
This is f*cking insane. Andrew Ng's DeepLearning .AI and Neo4j put out a free 1-hour course on what comes after RAG
How to turn raw documents into memory your agents actually keep:
→ Why bigger context windows are not enough
→ Extract entities, relationships and sources
→ Coordinate specialized agents inside one graph
→ Build it with Google ADK and Neo4j
→ Turn knowledge graphs into persistent memory
Documents → Entities → Relationships → Agents → Knowledge Graph
Most people are still learning RAG. This is the layer after it
Bookmark and watch it today
Then read the graph engineering part: Harness, Loops, Graphs ↓
Your agent is not autonomous
You are its loop
You check every step because nothing else in the system can
I broke down the 7 levels from babysitting agents to running a fleet ↓ https://t.co/4RacdXA4nE
This is f*cking insane. Nobel Prize winner Geoffrey Hinton, the Godfather of AI:
"If you sleep well tonight, you may not have understood this lecture"
In 47 minutes, the man who built the neural networks behind every AI alive explains how it actually works and where it's going
Neural Nets → Language Models → Understanding → Agents → Smarter Than Us
He quit Google so he could say this out loud
Worth more than any $2,000 AI course
Bookmark and watch it tonight
Then read how to put it to work for you: Harness, Loops, Graphs ↓
Your agent is not autonomous
You are its loop
You check every step because nothing else in the system can
I broke down the 7 levels from babysitting agents to running a fleet ↓ https://t.co/4RacdXA4nE
This is f*cking insane. Andrej Karpathy recorded 2 hours of how he actually uses AI every single day:
"Most people use 10% of what these models can do
This is the other 90%"
In 2 hours, he goes from which model to pick to one prompt that writes a full research report
Pick the Model → Thinking Models → Research Reports → Run Code → Memory Across Chats
He doesn't use anything you can't open right now
Worth more than most $1,000 AI courses
Bookmark and watch it this weekend
Then read how to turn that 90% into a system: Harness, Loops, Graphs ↓
Your agent is not autonomous
You are its loop
You check every step because nothing else in the system can
I broke down the 7 levels from babysitting agents to running a fleet ↓ https://t.co/4RacdXA4nE
This is f*cking insane. OpenAI co-founder Andrej Karpathy put everything he knows on GitHub for free:
"Don't read about it. Build the small version yourself, and the big one stops being magic"
In 24 seconds, he explains why every one of his projects is small enough to read in a weekend
From building GPT from scratch to an AI that runs its own research overnight. 6 repos, 300K+ stars
Micrograd → nanoGPT → llm.c → nanochat → autoresearch → You Understand the Whole Stack
Save this before it disappears from your feed
Every repo below is free:
An AI that runs its own experiments overnight and keeps only what works
https://t.co/uLuLfxuP6p
Train your own GPT from scratch, the repo half the AI world learned on
https://t.co/m3LPMWf4cF
Build your own ChatGPT for about $100
https://t.co/fpDw9HiyTU
The one-page idea behind the second brain for AI
https://t.co/IMSGK9LKXC
Train a language model in plain C, nothing hidden
https://t.co/jImptz0iUY
How neural networks learn, in a few hundred lines
https://t.co/d0llOqOLcJ
The tokenizer, built from zero
https://t.co/CHyLKuLJvx
His course on building an LLM end to end
https://t.co/oVmV6Dvdwu
Anthropic's SDK for building your own agents
https://t.co/0lGAjP5Bm6
Skills for Excel, Word, PowerPoint and PDF
https://t.co/1ZkjTNbWZL
A harness with everything built in
https://t.co/TXBJMRALVx
A loop that keeps the agent going until the job is done
https://t.co/iyuXUq4tTs
12 principles of agents that actually work
https://t.co/iIsk4nTkXC
18 free lessons on building agents
https://t.co/V3hPGnz2mp
Total cost: $0
Start with micrograd tonight. By the weekend you'll know what's actually inside the model
95% will keep reading threads about AI, the 5% who build the small version will be the ones who get it
Bookmark it and send it to someone who wants to understand how it works
Then read the guide below and put a harness around what you build ↓
Your agent is not autonomous
You are its loop
You check every step because nothing else in the system can
I broke down the 7 levels from babysitting agents to running a fleet ↓ https://t.co/4RacdXA4nE
This is f*cking insane. Andrej Karpathy just broke the whole idea of what an agent is:
"Agents aren't magic. They're distillation at scale
99% of your LLM's capacity goes to data it never needed"
In 16 minutes, he lays out the full stack: a small model, the right tools and a closed loop
Giant Model → Small Model → Right Tools → Closed Loop → An Agent That Actually Works
The model was never the hard part. The loop around it is
Worth more than any $500 AI course
Bookmark and watch it today
Then read how to close that loop: Harness, Loops, Graphs ↓
Your agent is not autonomous
You are its loop
You check every step because nothing else in the system can
I broke down the 7 levels from babysitting agents to running a fleet ↓ https://t.co/4RacdXA4nE
This is f*cking insane. Andrej Karpathy says most of your model is dead weight:
"A 1B model can be much more effective than most current LLMs
90% of current models waste capacity and money remembering stuff that doesn't matter"
In 17 minutes, he explains the cognitive core: strip the memorized internet, keep the reasoning
Giant Model → Cognitive Core → Tools → Tight Feedback Loop → Small Model + Harness
A raw LLM is a CPU with no OS. The harness is the OS
Worth more than most $1,000 AI courses
Bookmark and watch it today
Then read how to build that OS: Harness, Loops, Graphs ↓
Your agent is not autonomous
You are its loop
You check every step because nothing else in the system can
I broke down the 7 levels from babysitting agents to running a fleet ↓ https://t.co/4RacdXA4nE
This is f*cking insane. Andrej Karpathy, the man who led AI at Tesla:
"90% of what AI Twitter tells you to learn will be dead in 6 months
Frameworks die. The harness around the model stays"
In 39 minutes, he walks through what senior AI engineers already stopped wasting time on, and the short list that survives every new model, Fable 5.2 included
Frameworks → Benchmarks → Prompt Tricks → Harness → Loops → Graphs
Every new model kills another pile of tutorials
Worth more than any $1,500 AI engineering course
Bookmark and watch it today
Then read what survives: Harness, Loops, Graphs ↓
Your agent is not autonomous
You are its loop
You check every step because nothing else in the system can
I broke down the 7 levels from babysitting agents to running a fleet ↓ https://t.co/4RacdXA4nE
This is f*cking insane. Google engineer paid $500K-$1.2M a year:
"I don't write code anymore. My agents write it, test it and review it, and it doesn't break"
In 41 minutes, he shows the whole system: idea, spec, plan, test-first build, then the agent checks its own work in a browser
Vague Idea → Spec → Small Tasks → Test-First Build → Self-Review → Shipped
He open-sourced all 24 skills, and the repo is close to 100K stars
Worth more than most $2,000 agent engineering courses
Bookmark and watch it today
Then read the layer above this: how a fleet of these agents runs on loops and graphs in Harness, Loops, Graphs ↓
Your agent is not autonomous
You are its loop
You check every step because nothing else in the system can
I broke down the 7 levels from babysitting agents to running a fleet ↓ https://t.co/4RacdXA4nE
This is f*cking insane. Andrej Karpathy, OpenAI founding member and ex-head of AI at Tesla:
"Prompting is going away. Delete everything, keep the graph"
In this 2-hour lecture, he builds the graph from scratch and shows why it's the only layer that survives
Prompts → Agents → Loops → Graphs → A System That Writes Its Own Prompts
A loop keeps one agent on the job until it's done. A graph decides which agents exist and what each one passes on
Worth more than any $3,000 AI engineering course
Bookmark it and watch it this weekend
Then read the graph part: how one agent becomes a fleet in Harness, Loops, Graphs ↓
This is f*cking insane. Andrej Karpathy, the man who led AI at Tesla:
"The biggest mistake in AI right now is forcing agents to work before you master the harness around the model
We made that mistake at OpenAI in 2016. It cost us 5 years"
In under 6 minutes, he explains why a demo takes a weekend and a product takes a decade
Prompts → Agents → Graphs → Harness → Agents That Actually Ship
The agent was never the product. The harness under it is
Worth more than most $1,000 agent engineering courses
Bookmark and watch it today
Then read the motion engineering article: 100+ defects across 10 Opus 5.5 films, and the harness that catches them ↓
I fixed the same bug in two films
Opus never remembers my fixes
So I hired an AI employee to remember them for me, then tried to make him break my one rule on deadline
He didn't ↓ https://t.co/i7RRo3bIg2
This is f*cking insane. Andrej Karpathy, OpenAI founding member and ex-head of AI at Tesla:
"Prompting is going away. Delete everything, keep the graph"
In this 2-hour lecture, he builds the graph from scratch and shows why it's the only layer that survives
Prompts → Agents → Loops → Graphs → A System That Writes Its Own Prompts
A loop keeps one agent on the job until it's done. A graph decides which agents exist and what each one passes on
Worth more than any $3,000 AI engineering course
Bookmark it and watch it this weekend
Then read the graph part: how one agent becomes a fleet in Harness, Loops, Graphs ↓
Your agent is not autonomous
You are its loop
You check every step because nothing else in the system can
I broke down the 7 levels from babysitting agents to running a fleet ↓ https://t.co/4RacdXA4nE
This is f*cking insane. Andrej Karpathy, the man who led AI at Tesla:
"The biggest mistake in AI right now is forcing agents to work before you master the harness around the model
We made that mistake at OpenAI in 2016. It cost us 5 years"
In under 6 minutes, he explains why a demo takes a weekend and a product takes a decade
Prompts → Agents → Graphs → Harness → Agents That Actually Ship
The agent was never the product. The harness under it is
Worth more than most $1,000 agent engineering courses
Bookmark and watch it today
Then read the motion engineering article: 100+ defects across 10 Opus 5.5 films, and the harness that catches them ↓
This is f*cking insane. Anthropic Research Lead:
"We don't write prompts anymore. We build the graph around the model and let it write them"
In 20 minutes, she shows how graph engineering turns separate agents into one system that gets better on its own
One Agent → Loops → Plan Mode → Graphs → Self-Improving System
The prompt was never the product. The system around it is
Worth more than most $1,000 agent engineering courses
Bookmark and watch it today
Then read the motion engineering article: 100+ defects across 10 Opus 5.5 films, and the AI hire that catches them ↓
I fixed the same bug in two films
Opus never remembers my fixes
So I hired an AI employee to remember them for me, then tried to make him break my one rule on deadline
He didn't ↓ https://t.co/i7RRo3bIg2
This is f*cking insane. A Google engineer just took a tiny AI model from 46% to 90% accuracy, running offline on a phone
"A 270M model can beat a 70B one on your task"
In 21 minutes, he shows the whole loop
Tiny Model → Synthetic Data → Fine-Tune → Quantize → 2,000 Tokens a Second on Your Phone
No cloud. No API bill. Fully offline, in your pocket
Better than any $1,500 on-device AI bootcamp
Bookmark and watch it today
Then read the "Harness, Loops, Graphs" guide below ↓