The thread said last week. The lecture is eighteen months old.

I bookmarked it before I watched it. That is exactly what the thread asked for: “You probably don’t have 2 hours right now. Don’t let this disappear from your feed.” Andrej Karpathy, eight years at OpenAI and Tesla, had supposedly packed everything he knows into one free two-hour lecture last week. Agents, loops, graphs, self-improving systems. That evening I went back to it and found a Community Note underneath. The lecture in the video is called “How I Use LLMs” and it has been online since February 27, 2025.

Eighteen months old, and I had filed it as something from last week.

How the thread is built

Four layers stacked on each other. Only together do they show what the thing is for.

Here is the post, word for word:

Andrej Karpathy spent 8 years at OpenAI and Tesla

Last week, he packed everything he knows into one free 2-hour lecture

Agents → Loops → Graphs → Self-Improving Systems

People pay $15K for bootcamps that teach less than this

This lecture is better than most paid AI engineering courses

You probably don’t have 2 hours right now

Don’t let this disappear from your feed

Watch it

Then read the guide below

Source: @0xwhrrari on X, August 17, 2026.

And here is the Community Note that ended up underneath it:

The lecture shown in the video, ‘How I use LLMs,’ was actually released by Andrej Karpathy on February 27, 2025, not ‘last week’.

The rest of what hangs off the thread is below.

Layer 1 borrows the authority, layer 2 cashes it in, layers 3 and 4 arrive late. That is the whole machine.

What the thread claims and what the source shows

The post by @0xwhrrari went up on August 17, 2026. When I checked it, it stood at 559,500 views, 2,700 likes, 422 reposts, 85 replies and 8,000 bookmarks. Four claims, and they hold up to different degrees.

The second row is the one worth sitting with. The date is wrong, and the table of contents belongs to a lecture that is not in the video. “How I Use LLMs” is a general-audience tour through ChatGPT, Claude and Gemini. Agents, loops and graphs never come up. Those words come from the article the thread is actually selling, “Graph Engineering: How to Build AI Agent Systems That Don’t Break at Scale”, published a week earlier by the same account. The video supplies the authority, the article is the destination.

One template, several accounts

While I was checking this, I ran into the same construction in a handful of variants, all posted between August 13 and 19, 2026:

• @0xCodila on August 13, 332,200 views: “Anthropic’s Andrej Karpathy just released 1-hour Stanford lecture.” Karpathy has never worked at Anthropic; he founded Eureka Labs in 2024.
• At least four accounts carrying the identical breakdown: 10% LLM, 30% Prompt, 50% Agent, 70% Loop, 100% Graph, followed by the quote “Delete everything, keep Graph.”
• The same copy showed up on LinkedIn on August 14, as “This 1-hour Stanford lecture will teach you more about AI engineering than 100 random AI courses.”

That quote is the sharpest piece of evidence in the whole pile. In the Stanford lecture these posts point at, CS25 V2 “Introduction to Transformers” from 2023, Karpathy explains how the transformer came about by removing the recurrent layer: delete all the RNN, just keep attention. In the reposts that has become “delete everything, keep Graph”, which happens to match the subject of the article each post links to. A line about architecture from 2023 now reads as a slogan about agent graphs in 2026.

Why the format works

Part of the explanation is the bookmark button. The August 17 post has three times as many bookmarks as likes. The copy asks for that in so many words. Saving is cheaper than watching, and it costs a reader nothing to pretend they will free up two hours. Of the 559,500 people who saw this, almost none finished the video, and the construction does not need them to.

What strikes me more is how little labor was involved. The video already exists, it is free, and it sits on the channel of someone whose reputation you can borrow. The only work added was inventing a table of contents and attaching a date.

One number argues against my own story. Karpathy’s original post on February 27, 2025 pulled 956,700 views, 14,000 likes and 10,000 bookmarks. The original outperformed the recycled version. The correction landed too. Within a day two replies carried the right date, and the Community Note followed. So the system does catch this. It is just slow relative to the first thousand bookmarks.

I also do not want to pretend the video is bad. “How I Use LLMs” remains one of the better introductions out there, and anyone who watched it because of this thread got two useful hours. The damage lives in the year and in the invented chapter list, not in the material.

The caveat

The standard shifts from date to mood. Once “last week” becomes a figure of speech rather than a fact, a timeline loses its one distinguishing property. In AI coverage that matters more than elsewhere, because eighteen months is roughly two model generations. Someone learning in August 2026 how you worked with an LLM in February 2025 is getting advice about tools that behave differently now.

Responsibility lands on volunteers and on the source. A Community Note is written by readers, unpaid and unscaled, and it only appears once enough people rate it helpful. Karpathy is not involved at any point and still carries the reputational cost of a chapter list he never spoke. That is the same asymmetry I wrote about in the piece on watermarks and provenance: we keep getting better at measuring which machine made something, and worse at recording who claimed what.

What drains away is the instinct to look. I saved that post without checking a single thing, and I write about this subject every week. Saving feels like diligence while it is deferral that your brain files as a verdict.

What I do now before I save anything

Three actions that together cost me under a minute.

1. Get the date from the source, not from the poster. YouTube shows the upload date under the video, arXiv shows the v1 date. When someone writes “just released” without linking the original, that absence is the signal.
2. Search the source for one line from the table of contents. If the breakdown in the thread appears nowhere in the video or its transcript, the breakdown was invented. For this post that took two minutes to establish.
3. Read the replies before the Community Note exists. The correction was there hours before the label, in the form of two people who recognized the video.

This is the same habit I apply to vendor claims in my work at Novum: ask for the measurement, look at the date, and only then look at the percentage. I skipped it here because it was a YouTube video instead of a procurement document.

Frequently asked questions

Which video is actually in that thread?

“How I Use LLMs” by Andrej Karpathy, 2 hours and 11 minutes, published by him on his own YouTube channel on February 27, 2025. The Community Note under the post points to the same video.

Does the lecture about agents, loops and graphs exist at all?

Not in that form. The progression from agents to loops to graphs comes from the article the poster is promoting. The Stanford lecture that related posts point at is CS25 V2 “Introduction to Transformers” from 2023, which covers the transformer architecture rather than agent graphs.

Does this break any rule on X?

Not as far as I can tell. Sharing an old video is allowed, and there is no outright lie about authorship, since the video really is Karpathy’s. What goes wrong sits in the words “last week” and in a chapter list that does not match the video. Community Notes is the only correction mechanism available for that.

How do I spot this pattern in my own feed?

Watch for the combination of a well-known name, a time reference with no date, a numbered breakdown you can read without watching anything, and a link to the poster’s own article at the bottom. Those four together are a more reliable signal than any one of them alone.

Sources

• @0xwhrrari on X — the post carrying the Community Note, August 17, 2026, accessed August 21, 2026
• Andrej Karpathy, “How I Use LLMs” on YouTube — February 27, 2025, 2 hours 11 minutes
• Karpathy’s own announcement on X — February 27, 2025, including the full chapter list
• @0xCodila on X — August 13, 2026, the “Anthropic’s Andrej Karpathy” variant
• Stanford CS25, “Introduction to Transformers w/ Andrej Karpathy” — recorded in 2023, published through Stanford Online

Checked on August 21, 2026. The view and bookmark counts were read on that day and move afterward. I did not watch the video end to end; I relied on the chapter list Karpathy published alongside it. I could not establish with certainty which exact recording the “Delete everything, keep Graph” reposts used: those reuploads run about an hour while the Stanford CS25 lecture runs longer, so an edited cut is the likely explanation. The eight years at OpenAI and Tesla only add up if you count Karpathy’s second stint at OpenAI.

This piece also appeared in Dutch: Vorige week, zei het draadje. De lecture is anderhalf jaar oud..