To help explain the weirdness of LLM Tokenization I thought it could be amusing to translate every token to a unique emoji. This is a lot closer to truth – each token is basically its own little hieroglyph and the LLM has to learn (from scratch) what it all means based on
@karpathy
-
Insufficient Training Data for Spelling Tasks in LLMs
By
–
I think there’s not enough training data naturally on the internet of spelling tasks compared to the difficulty of the task for the LLM, due to how text is chopped up into sequences of text chunks (tokens), all of which are unique / distinct. I have a whole video on Tokenization.
-
Jagged: Radar Charts and Uneven AI Capability Distributions
By
–
For me the etymology of "jagged" is the radar charts people use for evals, where each angle is an eval, and the LLM sweeps out an area. In this diagram, imagining a lot more of different kind of capabilities along each direction, capability would look more spiky / jagged.
-
LLMs Imitate Training Data Format During Post-Training Stage
By
–
It does that because all of its training data in the last, post-training stage are of the form [question -> authoritative sounding solution], where the solutions are written by humans. The LLMs just imitate the form/style of that training data.
-
Jagged Intelligence: LLMs Excel and Fail Unpredictably
By
–
Jagged Intelligence The word I came up with to describe the (strange, unintuitive) fact that state of the art LLMs can both perform extremely impressive tasks (e.g. solve complex math problems) while simultaneously struggle with some very dumb problems. E.g. example from two
-
Investing in Distributed Creator Intelligence Over Concentrated Resources
By
–
I'd be a lot more inclined to invest $10M into 2000 creators. The distributed intelligence and creativity of the crowd feels underutilized.
-
LLMs Reaching LHC-Level Complexity in Infrastructure and Development
By
–
LLMs as an artifact are trending to the complexity of something like the LHC. This is clear when you look at the datacenter computronium build out but it's a lot more than that – a large chunk is digital and much harder to see/appreciate, it's just a bunch of people on a laptop.
-
Karpathy Plans Multiple Llama 3.1 Fine-tuning Projects
By
–
Yep I’d like to do many Llama 3.1 finetunes, coming up.
