Anthropic just dropped 13 free AI courses with certificates.
Not one or two tutorials. An entire learning hub covering Claude Code, APIs, MCP, AI fluency, and building AI applications.
A few months ago Google did the same. OpenAI followed.
Three competitors all teaching people to build with AI for free. That probably isn't charity.
Elon once said that Andrej Karpathy has the second-best computer vision brain alive.
Only Ilya Sutskever is above him.He might not have been exaggerating.
At 22, Karpathy was solving a Rubik’s Cube in 16 seconds. Then he spent the next 15 years working on something much harder - teaching computers to understand the real world.
He left Slovakia at 15 to study quantum computing in Toronto, but quickly switched to AI. Quantum felt too clean and abstract. He wanted something with more real-world mess.
He studied under Geoffrey Hinton, then worked with Fei-Fei Li at Stanford. In 2012 he was saying computer vision was still “really, really far away.”By 2014 he had personally beaten Google’s best model on ImageNet.
- Google: 6.8% error.
- Karpathy: 5.1%.
After that his career took off fast.
He helped start OpenAI. Musk brought him to Tesla to build vision for Autopilot using just cameras. When he left OpenAI, the team wasn’t happy about it.
Later he returned to OpenAI, popularized the idea of “vibe coding,” started his own projects, and now works at Anthropic on Claude’s pretraining.
The person who spent years teaching machines how to see is now helping them reason.
At the same time, Meta is offering $100M+ packages to top AI talent.
But here’s what actually stands out: Karpathy recently admitted he feels more behind than ever.
If one of the strongest researchers in the field feels like he’s struggling to keep up, what does that say about everyone else?
Karpathy can keep his brain in a plain text file, Boris Cherny built Claude Code around markdown, and Obsidian is basically the cheap little corner where both ideas start making uncomfortable sense.
The funniest part is that everyone keeps trying to invent the next “second brain” app, while the best format for AI was sitting there the whole time: dumb .md files in a folder.
No database. No lock-in. No $30/month productivity shrine. Just text files Claude can actually read, edit, connect, and build around.
That’s where Obsidian gets weird.
You dump everything into one vault: book notes, code snippets, strategies, unfinished thoughts, random ideas, old plans, things you wrote once and forgot existed. One idea per note. Links instead of folders. No beautiful system. No productivity cosplay.
Then you open Claude Code inside that vault.
Suddenly your notes are not notes anymore. They’re a codebase for your brain.
Claude can scan the whole thing, find patterns, connect ideas, summarize dead branches, spot repeated themes, and build tools on top of your own thinking.
One builder asked it to turn his vault into a 3D map. Same evening, he had a local force graph running on localhost:8766, with every note as a glowing node and every link as a visible connection.
Click a note, and it pulled the summary, outgoing links, backlinks, and nearby context.
Then he asked for Terrain Mode.
Now his vault renders like a landscape. Red peaks show the ideas he keeps returning to. Green flats show the notes that basically died in silence. The system showed that 73% of his knowledge was still active.
Which sounds like fake productivity nonsense until you realize it’s actually useful.
Because most people don’t have a note-taking problem. They have a “my thoughts are scattered across hundreds of files and I pretend search is enough” problem.
Obsidian gives you the raw material. Claude Code gives it hands. Markdown makes the whole thing portable enough that AI doesn’t choke on it.
Karpathy picked plain text because it lasts. Cherny built around files because models can work with them. Claude turns the folder into an engine.
Your notes were never useless.
They just never had someone technical enough to wire them together.
This video shows a fully animated 3D site built in 2 minutes with claude code. one terminal command. but the prompt was almost irrelevant.
the real work was the pre-load: ui/ux pro max skill, a hero component from https://t.co/wld7XqRAok, framer motion stacked in before claude touched anything
the toolkit did the job. the prompt just pulled the trigger.
that's also why most people's claude output looks generic. they skip the setup and blame the model.