This is a list of all the youtube videos which have contributed to my current interest in machine learning. Some of them go back years like CGP Grey, Computerphile, Robert Miles and 3Blue1Brown whereas the rest are ~1 year but these are channels I keep coming back to over and over again because they’re extremely interesting and don’t feel like a class.
Probably one of my all time favourite youtube channels of all time. I watched his video Humans need not apply back in 2014 and it was the reason why I wanted to go into computer programming. His explanations on How Machines Learn which was released 6 years ago (renamed to “How AIs, like ChatGPT, Learn” in 2023) is the best beginner video that exists on the topic and I think everyone should watch. These two videos you should send to your parents when they’re asking about “All this AI stuff”.
Computerphile, Numberphile and Sixty Symbols created by Brady Haran are also what got me interested in STEM around 10 years ago.
The videos from Mike Pound and Robert Miles are great, and were my introduction to data analysis, ML, and AI.
#TimsForLife
Robert Miles is a AI safety researcher and his content is excellent. His Computerphile videos on GPT-2 and Attention is all you need papers influenced me into studying ML/AI along with my CS degree. He covered a lot of the original papers that OpenAI was releasing for the last few years which is how I found out about OpenAI.
3Blue1Brown has the best math educational content that exists. Grant is the reason why I passed any of my math courses in university.
Grant’s manim graphing library for python (which also has a community fork) is also pretty important for STEM communication, and I’m planning on using it if I get to the data visualisation phase.
Probably one of the most important people in the field at the moment. His zero to hero playlist takes you through concrete examples and is really democratising language models through education: ttps://karpathy.ai/zero-to-hero.html
Anrej's blogpost on software 2.0 also holds up to reality pretty well and I think something similar is going to be where all software engineering is headed.
I ported his project llm.c to Go https://github.com/joshcarp/llm.go and this is when everything really started to click for me. There’s also a history of this occurring with his project llama2.c being forked into llama.cpp, which is currently the fastest and most optimised open source LLM runtime which many, many other projects rely on including ollama, llamafile, and localllm and many, many more.
The current work he’s doing on creating a highly optimised library in https://github.com/karpathy/llm.c/ is fascinating, and recently the project was able to get to 29.3% on the Hellaswag benchmark with $20 of training, where the GPT-3 small model (124M) got 33%.