When is Andrej Karpathy going to look at a chicken nugget and tweet that it helped him solve AGI, which in turn inspires 6 random devs to create GitHub projects giving us actual AGI?

Posted by Porespellar@reddit | LocalLLaMA | View on Reddit | 55 comments

Karpathy appreciation post. Seriously tho, he’s done this like a bunch of times lately. Every time he sneezes on the subway we get a bunch of developers becoming inspired by his ideas and turning them into viable AI-related Gitub projects that actually do really amazing things. This guy is on a roll lately.

He is one of the greatest minds in AI and we are very fortunate that he occasionally lurks on this sub. Andrej, if you’re reading this, Thanks for all the cool stuff you’ve put out into the world and thank you for inspiring others to do the same.

In case anyone needs a reminder, look into:

- Second Brain
- AutoResearch
- LLM-Wiki
- nanoGPT
- AgentHub
- LLMcouncil
- GPT-2
- Autopilot (Tesla)
- “vibecoding” (he coined the term)

I’m sure I’m missing a bunch of other of his accompaniments, projects, or ones he’s inspired, so please add if you know some others.

[-]

oe_throwaway_1@reddit

His 3.5hr youtube video on how LLMs are trained is one of the best things on the internet.

[-]

tekprodfx16@reddit

Happy cake day, do you have a link to the video? Never seen it

[-]

tekprodfx16@reddit

Im guessing it’s this one https://youtu.be/7xTGNNLPyMI?si=-CVWm3i30e6qS50L

[-]

PACmaneatsbloons@reddit

[-]

UnitedWraps@reddit

That’s the rick roll video link

[-]

tekprodfx16@reddit

Sneaky mofos, thanks!

[-]

oe_throwaway_1@reddit

Thanks, that's the one.

[-]

cristoper@reddit

His single-page Python implementation of a gpt-2-ish transformer is also pretty elegant:

https://karpathy.github.io/2026/02/12/microgpt/

[-]

FullstackSensei@reddit

LLM zero to hero

[-]

Silver-Champion-4846@reddit

I keep hearing about his accomplishment. If only I had good hardware or even the money to test with cloud platforms I could even use agents themselves to help me learn stuff.

[-]

dankfrankreynolds@reddit

Deepseek v4 is cheap AND 75% off the rest of this month

[-]

Silver-Champion-4846@reddit

Cheap!=usable for me, I can't even pay 1c

[-]

Stepfunction@reddit

We already know that all that's needed to solve AGI is a vibe-coded neural hierarchical spiking memory harness.

[-]

Silver-Champion-4846@reddit

Did you stuff every buzzword you thought could fit into this?

[-]

Stepfunction@reddit

That's the joke.

[-]

Silver-Champion-4846@reddit

AHA!

[-]

EagleNait@reddit

Gotta be quantum tho

[-]

jazir55@reddit

Needs more blockchain

[-]

qudat@reddit

Nah just a cortico-basal ganglia-thalamo-cortical loop

[-]

Time_Cat_5212@reddit

I have this revolutionary idea for a harness that uses a rigid agentic pipeline to seed injections directly into the backend. No memory required ;)

[-]

SmartCustard9944@reddit

Dirty bot

[-]

apeapebanana@reddit

gotta fit heretic-i1-mtp-gguf in there somewhere

[-]

Stepfunction@reddit

Only if it's opus-distilled

[-]

shockwaverc13@reddit

QAT trained for bitnet

[-]

CasualtyOfCausality@reddit

I dunno. While mostly accurate, without neuromorphic hardware, I'm pretty sure it is dead in the water.

[-]

techlatest_net@reddit

Super useful. Having exact configs and context lengths for the 5060 Ti saves a ton of trial-and-error. Love that it's focused on reproducible results instead of just hype numbers. Will definitely reference this when I tweak my own setup.

[-]

Agitated_Space_672@reddit

You wrote llmcouncil, did you mean llm-consortium? That's my karpathy inspired project from December 2024 https://x.com/xundecidability/status/1870861603111145667?s=20

[-]

onil_gova@reddit

Let's not forget that Karpathy acted as a human baseline for the ImageNet competition in 2014, jokingly referred to as the "reference human" for the test.

[-]

as_ninja6@reddit

Also hand picked each weight values for a neural net

[-]

CorpusculantCortex@reddit

August 27, 2043

[-]

FastHotEmu@reddit

He may tweet about it, but Georgi Gerganov will implement it overnight by hand on a napkin

[-]

KnowLimits@reddit

17 years after Schmidhuber already published it of course.

[-]

kyr0x0@reddit

Sure. I guess some old arab mathematician also solved it 2000 years ago, drawing matrices into sand

[-]

axiomaticdistortion@reddit

With all due respect, people were already doing what he made popular many times over. His deed is spreading the message and not being the creator most of the times.

[-]

jeffwadsworth@reddit

Dude this post made me genuinely laugh out loud.

[-]

Mr_Finious@reddit

Between the cargo-culting zealots, sloppy ai bots and general astroturfing… this sub needs some serious moderation.

[-]

Porespellar@reddit (OP)

LOL ok. Which one am I? And what dafuq is a cargo-culting zealot?

[-]

Mr_Finious@reddit

You are none my friend. It’s just what this sub has become. At least you seem to be a human. (Who knows anymore)

[-]

Porespellar@reddit (OP)

Just giving you a hard time man. I appreciate the sentiment. I feel like I’m always on high alert for emdashes. Friggin bots are ruining the sub.

[-]

Mr_Finious@reddit

Yeah. Between that and the astroturfing on the Chinese model subs, it kinda feels like an era is over on her.

[-]

Porespellar@reddit (OP)

Just looked that up, interesting, had never heard of that before seems similar fake-it-til-you-make-it or like a Dunning Kruger type thing.

[-]

simotune@reddit

Karpathy discourse is basically an R&D allocation mechanism now. One offhand post and half the open-source roadmap reprioritizes itself.

[-]

cristoper@reddit

His single-page Python implementation of a gpt-2-ish transformer is also pretty elegant:

https://karpathy.github.io/2026/02/12/microgpt/

Unfortunately if chicken nuggets were the key, believe me, I would have built AGI already.

[-]

silverud@reddit

It's obviously not just about the nuggets, it's about the relationship between the nuggets and the sauce you dip them in. Only there will you find the secret of AGI.

[-]

DontCallMeLarry@reddit

Brotherman probably doesn't have to work another minute for the rest of his life, and puts out nothing but useful content and good vibes for free - this is something we can all aspire to.

[-]

More-Curious816@reddit

don't forget the marketing gifts. Jensen Huang recently delivered the world's first NVIDIA DGX Station GB300 to Andrej Karpathy, with a handwritten note from Jensen that said "You were with me every step of the way."

[-]

CircularSeasoning@reddit

In my mind, the blog post below was modern-day LLMs' "Hello World" moment:

https://karpathy.github.io/2015/05/21/rnn-effectiveness/

[-]

conjuncts@reddit

That's a remarkably early use of the word "hallucination" in the context of language models.

[-]

CircularSeasoning@reddit

Interesting point. Also, near where he mentions hallucinations (twice), he notices that they can (attempt to) make textual diagrams, too. That's something people were still being surprised about much later with LLMs, as if it were an exclusively LLM thing.

I read the blog post in about 2017 and it inspired me to immediately rent a GPU to train my own LSTM-based char-level RNN on random texts from Project Gutenberg. That was a lot of fun. At the time I never expected it would all lead to... this... at least not so soon in my lifetime. Granted, this was about a decade ago now. How time flies when you're waiting for your PC to learn how to write Shakespearean poetry!

[-]

Porespellar@reddit (OP)

Right up there with the “Attention is All You Need” paper.

[-]