Series: My AI Journey, Part 1

This series follows my journey of understanding AI, Machine Learning, LLMs, and the many various models, patterns, companies, APIs, people, concepts, movers, and news that shapes that understanding.

Reading Notes:

• Andrej Karpathy's Software 2.0
• What is ChatGPT doing and why does it work?

Reading List:

A16z offers the "AI Canon": https://a16z.com/ai-canon/

What's a "vector" in the context of AI? https://www.pinecone.io/learn/

Prompt engineering guide: https://www.promptingguide.ai/

OpenAI Cookbook: https://github.com/openai/openai-cookbook/tree/main

Chain-of-thought: https://arxiv.org/abs/2201.11903

Sparks of AGI: https://arxiv.org/pdf/2303.12712

A survey of LLMs: https://arxiv.org/pdf/2303.18223v4

Chinchilla's implications: https://www.lesswrong.com/posts/6Fpvch8RR29qLEWNH/chinchilla-s-wild-implications

AI for FSD at Tesla: https://www.youtube.com/watch?v=hx7BXih7zx8

Predictive learning: https://www.youtube.com/watch?v=Ount2Y4qxQo&t=1072s

Reinforcement Learning: https://www.youtube.com/watch?v=hhiLw5Q_UFg

Reinforcement Learning from Human Feedback (RLHF): https://huyenchip.com/2023/05/02/rlhf.html

Illustrated Stable Diffusion: https://jalammar.github.io/illustrated-stable-diffusion/

Built GPT: https://www.youtube.com/watch?v=kCc8FmEb1nY

Annotated Transformer: https://nlp.seas.harvard.edu/annotated-transformer/

Stanford NLP: https://www.youtube.com/playlist?list=PLoROMvodv4rOSH4v6133s9LFPRHjEmbmJ

Stanford ML: https://www.youtube.com/playlist?list=PLoROMvodv4rMiGQp3WXShtMGgzqpfVfbU

Convolutional Neural Nets: https://cs231n.github.io/

Backpropagation, Neural Nets: https://www.youtube.com/watch?v=i94OvYb6noo

Word2Vec: https://towardsdatascience.com/word2vec-explained-49c52b4ccb71

Practical Deep Learning for Coders: https://course.fast.ai/Lessons/lesson1.html

Wolfram Alpha Neural Net Repo: https://resources.wolframcloud.com/NeuralNetRepository

Building LLM applications for production: https://huyenchip.com/2023/04/11/llm-engineering.html

Criteria

To narrow the scope, at this point I need an understanding that allows me to confidently architect a user-facing application that leverages an LLM to accomplish some task.

Approach

To start, I believe the correct approach is to build a "skill map" or "conceptual framework", and challenge that framework on a regular basis with all the new information.

Key Questions

• Why doesn't Jarvis exist yet? What's stopping that?
• How would I build such an AI with the technology we have today?
• For an application using an LLM, how can you keep user data private?
• Given that LLMs hallucinate, how do you avoid hallucination?
• How do you teach your LLM to get smarter over time?

Key Vocabulary

ML

• VC dimension, over-fitting, under-fitting
• logistic regression, kernel trick, boosting
• SVM, Bellman equation, decision tree
• naive Bayesian model, autoregressive model

DL

• Adam, softmax, residual connections
• relu, dropout, CLIP
• ViT, transposed convolution layer
• SGD, batchnorm, tokenizer, VAE
• LSTM, GRU, GPT, GAN
• Transformer

Math

• Hessian, entropy, mutual information
• Jacobian, gradient, Bayes' law
• eigen-decomposition, svd