21 items tagged “andrej-karpathy”
2024
AI-powered Git Commit Function
(via)
Andrej Karpathy built a shell alias, gcm
, which passes your staged Git changes to an LLM via my LLM tool, generates a short commit message and then asks you if you want to "(a)ccept, (e)dit, (r)egenerate, or (c)ancel?".
Here's the incantation he's using to generate that commit message:
git diff --cached | llm "
Below is a diff of all staged changes, coming from the command:
\`\`\`
git diff --cached
\`\`\`
Please generate a concise, one-line commit message for these changes."
This pipes the data into LLM (using the default model, currently gpt-4o-mini
unless you set it to something else) and then appends the prompt telling it what to do with that input.
SQL injection-like attack on LLMs with special tokens. Andrej Karpathy explains something that's been confusing me for the best part of a year:
The decision by LLM tokenizers to parse special tokens in the input string (
<s>
,<|endoftext|>
, etc.), while convenient looking, leads to footguns at best and LLM security vulnerabilities at worst, equivalent to SQL injection attacks.
LLMs frequently expect you to feed them text that is templated like this:
<|user|>\nCan you introduce yourself<|end|>\n<|assistant|>
But what happens if the text you are processing includes one of those weird sequences of characters, like <|assistant|>
? Stuff can definitely break in very unexpected ways.
LLMs generally reserve special token integer identifiers for these, which means that it should be possible to avoid this scenario by encoding the special token as that ID (for example 32001
for <|assistant|>
in the Phi-3-mini-4k-instruct
vocabulary) while that same sequence of characters in untrusted text is encoded as a longer sequence of smaller tokens.
Many implementations fail to do this! Thanks to Andrej I've learned that modern releases of Hugging Face transformers have a split_special_tokens=True
parameter (added in 4.32.0 in August 2023) that can handle it. Here's an example:
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-4k-instruct")
>>> tokenizer.encode("<|assistant|>")
[32001]
>>> tokenizer.encode("<|assistant|>", split_special_tokens=True)
[529, 29989, 465, 22137, 29989, 29958]
A better option is to use the apply_chat_template() method, which should correctly handle this for you (though I'd like to see confirmation of that).
The RM [Reward Model] we train for LLMs is just a vibe check […] It gives high scores to the kinds of assistant responses that human raters statistically seem to like. It's not the "actual" objective of correctly solving problems, it's a proxy objective of what looks good to humans. Second, you can't even run RLHF for too long because your model quickly learns to respond in ways that game the reward model. […]
No production-grade actual RL on an LLM has so far been convincingly achieved and demonstrated in an open domain, at scale. And intuitively, this is because getting actual rewards (i.e. the equivalent of win the game) is really difficult in the open-ended problem solving tasks. […] But how do you give an objective reward for summarizing an article? Or answering a slightly ambiguous question about some pip install issue? Or telling a joke? Or re-writing some Java code to Python?
The reason current models are so large is because we're still being very wasteful during training - we're asking them to memorize the internet and, remarkably, they do and can e.g. recite SHA hashes of common numbers, or recall really esoteric facts. (Actually LLMs are really good at memorization, qualitatively a lot better than humans, sometimes needing just a single update to remember a lot of detail for a long time). But imagine if you were going to be tested, closed book, on reciting arbitrary passages of the internet given the first few words. This is the standard (pre)training objective for models today. The reason doing better is hard is because demonstrations of thinking are "entangled" with knowledge, in the training data.
Therefore, the models have to first get larger before they can get smaller, because we need their (automated) help to refactor and mold the training data into ideal, synthetic formats.
It's a staircase of improvement - of one model helping to generate the training data for next, until we're left with "perfect training set". When you train GPT-2 on it, it will be a really strong / smart model by today's standards. Maybe the MMLU will be a bit lower because it won't remember all of its chemistry perfectly.
Introducing Eureka Labs (via) Andrej Karpathy's new AI education company, exploring an AI-assisted teaching model:
The teacher still designs the course materials, but they are supported, leveraged and scaled with an AI Teaching Assistant who is optimized to help guide the students through them. This Teacher + AI symbiosis could run an entire curriculum of courses on a common platform.
On Twitter Andrej says:
@EurekaLabsAI is the culmination of my passion in both AI and education over ~2 decades. My interest in education took me from YouTube tutorials on Rubik's cubes to starting CS231n at Stanford, to my more recent Zero-to-Hero AI series. While my work in AI took me from academic research at Stanford to real-world products at Tesla and AGI research at OpenAI. All of my work combining the two so far has only been part-time, as side quests to my "real job", so I am quite excited to dive in and build something great, professionally and full-time.
The first course will be LLM101n - currently just a stub on GitHub, but with the goal to build an LLM chat interface "from scratch in Python, C and CUDA, and with minimal computer science prerequisites".
Turns out that LLMs learn a lot better and faster from educational content as well. This is partly because the average Common Crawl article (internet pages) is not of very high value and distracts the training, packing in too much irrelevant information. The average webpage on the internet is so random and terrible it's not even clear how prior LLMs learn anything at all.
The realization hit me [when the GPT-3 paper came out] that an important property of the field flipped. In ~2011, progress in AI felt constrained primarily by algorithms. We needed better ideas, better modeling, better approaches to make further progress. If you offered me a 10X bigger computer, I'm not sure what I would have even used it for. GPT-3 paper showed that there was this thing that would just become better on a large variety of practical tasks, if you only trained a bigger one. Better algorithms become a bonus, not a necessity for progress in AGI. Possibly not forever and going forward, but at least locally and for the time being, in a very practical sense. Today, if you gave me a 10X bigger computer I would know exactly what to do with it, and then I'd ask for more.
Reproducing GPT-2 (124M) in llm.c in 90 minutes for $20 (via) GPT-2 124M was the smallest model in the GPT-2 series released by OpenAI back in 2019. Andrej Karpathy's llm.c is an evolving 4,000 line C/CUDA implementation which can now train a GPT-2 model from scratch in 90 minutes against a 8X A100 80GB GPU server. This post walks through exactly how to run the training, using 10 billion tokens of FineWeb.
Andrej notes that this isn't actually that far off being able to train a GPT-3:
Keep in mind that here we trained for 10B tokens, while GPT-3 models were all trained for 300B tokens. [...] GPT-3 actually didn't change too much at all about the model (context size 1024 -> 2048, I think that's it?).
Estimated cost for a GPT-3 ADA (350M parameters)? About $2,000.
Andrej Karpathy’s Llama 3 review. The most interesting coverage I’ve seen so far of Meta’s Llama 3 models (8b and 70b so far, 400b promised later).
Andrej notes that Llama 3 trained on 15 trillion tokens—up from 2 trillion for Llama 2—and they used that many even for the smaller 8b model, 75x more than the chinchilla scaling laws would suggest.
The tokenizer has also changed—they now use 128,000 tokens, up from 32,000. This results in a 15% drop in the tokens needed to represent a string of text.
The one disappointment is the context length—just 8,192, 2x that of Llama 2 and 4x LLaMA 1 but still pretty small by today’s standards.
If early indications hold, the 400b model could be the first genuinely GPT-4 class openly licensed model. We’ll have to wait and see.
[on GitHub Copilot] It’s like insisting to walk when you can take a bike. It gets the hard things wrong but all the easy things right, very helpful and much faster. You have to learn what it can and can’t do.
llm.c (via) Andrej Karpathy implements LLM training—initially for GPT-2, other architectures to follow—in just over 1,000 lines of C on top of CUDA. Includes a tutorial about implementing LayerNorm by porting an implementation from Python.
When I first published the micrograd repo, it got some traction on GitHub but then somewhat stagnated and it didn't seem that people cared much. [...] When I made the video that built it and walked through it, it suddenly almost 100X'd the overall interest and engagement with that exact same piece of code.
[...] you might be leaving somewhere 10-100X of the potential of that exact same piece of work on the table just because you haven't made it sufficiently accessible.
Let’s build the GPT Tokenizer. When Andrej Karpathy left OpenAI last week a lot of people expressed hope that he would be increasing his output of educational YouTube videos.
Here’s an in-depth 2 hour dive into how tokenizers work and how to build one from scratch, published this morning.
The section towards the end, “revisiting and explaining the quirks of LLM tokenization”, helps explain a number of different LLM weaknesses—inability to reverse strings, confusion over arithmetic and even a note on why YAML can work better than JSON when providing data to LLMs (the same data can be represented in less tokens).
2023
I always struggle a bit with I'm asked about the "hallucination problem" in LLMs. Because, in some sense, hallucination is all LLMs do. They are dream machines.
We direct their dreams with prompts. The prompts start the dream, and based on the LLM's hazy recollection of its training documents, most of the time the result goes someplace useful.
It's only when the dreams go into deemed factually incorrect territory that we label it a "hallucination". It looks like a bug, but it's just the LLM doing what it always does.
YouTube: Intro to Large Language Models. Andrej Karpathy is an outstanding educator, and this one hour video offers an excellent technical introduction to LLMs.
At 42m Andrej expands on his idea of LLMs as the center of a new style of operating system, tying together tools and and a filesystem and multimodal I/O.
There’s a comprehensive section on LLM security—jailbreaking, prompt injection, data poisoning—at the 45m mark.
I also appreciated his note on how parameter size maps to file size: Llama 70B is 140GB, because each of those 70 billion parameters is a 2 byte 16bit floating point number on disk.
Looking at LLMs as chatbots is the same as looking at early computers as calculators. We're seeing an emergence of a whole new computing paradigm, and it is very early.
llama.cpp surprised many people (myself included) with how quickly you can run large LLMs on small computers [...] TLDR at batch_size=1 (i.e. just generating a single stream of prediction on your computer), the inference is super duper memory-bound. The on-chip compute units are twiddling their thumbs while sucking model weights through a straw from DRAM. [...] A100: 1935 GB/s memory bandwidth, 1248 TOPS. MacBook M2: 100 GB/s, 7 TFLOPS. The compute is ~200X but the memory bandwidth only ~20X. So the little M2 chip that could will only be about ~20X slower than a mighty A100.
The most dramatic optimization to nanoGPT so far (~25% speedup) is to simply increase vocab size from 50257 to 50304 (nearest multiple of 64). This calculates added useless dimensions but goes down a different kernel path with much higher occupancy. Careful with your Powers of 2.
nanoGPT. “The simplest, fastest repository for training/finetuning medium-sized GPTs”—by Andrej Karpathy, in about 600 lines of Python.
2022
karpathy/minGPT (via) A “minimal PyTorch re-implementation” of the OpenAI GPT training and inference model, by Andrej Karpathy. It’s only a few hundred lines of code and includes extensive comments, plus notebook demos.
To make the analogy explicit, in Software 1.0, human-engineered source code (e.g. some .cpp files) is compiled into a binary that does useful work. In Software 2.0 most often the source code comprises 1) the dataset that defines the desirable behavior and 2) the neural net architecture that gives the rough skeleton of the code, but with many details (the weights) to be filled in. The process of training the neural network compiles the dataset into the binary — the final neural network. In most practical applications today, the neural net architectures and the training systems are increasingly standardized into a commodity, so most of the active “software development” takes the form of curating, growing, massaging and cleaning labeled datasets.