One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like "switching to speech recognition sorry for any typos...". Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.

This is interesting as a first large diffusion-based LLM. Most of the LLMs you've been seeing are ~clones as far as the core modeling approach goes. They're all trained "autoregressively", i.e. predicting tokens from left to right. Diffusion is different - it doesn't go left to right, but all at once. You start with noise and gradually denoise into a token stream. Most of the image / video generation AI tools actually work this way and use Diffusion, not Autoregression. It's only text (and sometimes audio!) that have resisted. So it's been a bit of a mystery to me and many others why, for some reason, text prefers Autoregression, but images/videos prefer Diffusion. This turns out to be a fairly deep rabbit hole that has to do with the distribution of information and noise and our own perception of them, in these domains. If you look close enough, a lot of interesting connections emerge between the two as well. All that to say that this model has the potential to be different, and possibly showcase new, unique psychology, or new strengths and weaknesses. I encourage people to try it out!

In 2019, OpenAI announced GPT-2 with this post: openai.com/index/better-lang… Today (~5 years later) you can train your own for ~$672, running on one 8XH100 GPU node for 24 hours. Our latest llm.c post gives the walkthrough in some detail: github.com/karpathy/llm.c/di… Incredibly, the costs have come down dramatically over the last 5 years due to improvements in compute hardware (H100 GPUs), software (CUDA, cuBLAS, cuDNN, FlashAttention) and data quality (e.g. the FineWeb-Edu dataset). For this exercise, the algorithm was kept fixed and follows the GPT-2/3 papers. Because llm.c is a direct implementation of GPT training in C/CUDA, the requirements are minimal - there is no need for conda environments, Python interpreters, pip installs, etc. You spin up a cloud GPU node (e.g. on Lambda), optionally install NVIDIA cuDNN, NCCL/MPI, download the .bin data shards, compile and run, and you're stepping in minutes. You then wait 24 hours and enjoy samples about English-speaking Unicorns in the Andes. For me, this is a very nice checkpoint to get to because the entire llm.c project started with me thinking about reproducing GPT-2 for an educational video, getting stuck with some PyTorch things, then rage quitting to just write the whole thing from scratch in C/CUDA. That set me on a longer journey than I anticipated, but it was quite fun, I learned more CUDA, I made friends along the way, and llm.c is really nice now. It's ~5,000 lines of code, it compiles and steps very fast so there is very little waiting around, it has constant memory footprint, it trains in mixed precision, distributed across multi-node with NNCL, it is bitwise deterministic, and hovers around ~50% MFU. So it's quite cute. llm.c couldn't have gotten here without a great group of devs who assembled from the internet, and helped get things to this point, especially ademeure, ngc92, @gordic_aleksa, and rosslwheeler. And thank you to @LambdaAPI for the GPU cycles support. There's still a lot of work left to do. I'm still not 100% happy with the current runs - the evals should be better, the training should be more stable especially at larger model sizes for longer runs. There's a lot of interesting new directions too: fp8 (imminent!), inference, finetuning, multimodal (VQVAE etc.), more modern architectures (Llama/Gemma). The goal of llm.c remains to have a simple, minimal, clean training stack for a full-featured LLM agent, in direct C/CUDA, and companion educational materials to bring many people up to speed in this awesome field. Eye candy: my much longer 400B token GPT-2 run (up from 33B tokens), which went great until 330B (reaching 61% HellaSwag, way above GPT-2 and GPT-3 of this size) and then exploded shortly after this plot, which I am looking into now :)