Fun AI Projects

Math, Algorithm & ML

Efficient ML
AI for math
- Kaggle competitions
  - Kaggle AIMO
    - My submissions
  - AI Mathematical Olympiad - Progress Prize 2
- FAIR math theorem proving
- Terence Tao's talks
  - Terence Tao at IMO 2024: AI and Mathematics
    - Key words: machine learning, LLM, formal proof assistants
    - my doc
- Terence Tao's YouTube videos showing how to use LLM (Github Copilot, Claude, GPT-o4) to formalize mathematical proofs in LEAN
  - https://www.youtube.com/@TerenceTao27
Learning Generative AI
- Google's Introduction to Generative AI (very general and high level introduction, no tech details)
- Google's Model Garden on Vertex AI: a single place to discover, customize, and deploy a wide variety of models from Google and Google partners
Learning Language Models
- Tokenization and Build GPT2 by Andrej Karpathy
  - My Practice: Fine tune GPT2 with The Simpsons
- Hugging Face's NLP Course
  - Resources
    - an introductory deep learning course: fast.ai’s Practical Deep Learning for Coders
    - programs developed by DeepLearning.AI
    - DeepLearning.AI’s Natural Language Processing Specialization, which covers a wide range of traditional NLP models like naive Bayes and LSTMs that are well worth knowing about
- What is Temperature in NLP?
- Andrej's sharing: How I use LLMs
  - Excalidraw file for this video
  - Start a new conversation whenever you need, since irrlevant information in the context window is distracting (so reducing the response quality) and decreasing the responding speed
  - Pay attention to what model you're using (model providers have their own pricing tiers)
    - GPT, Claude, Gemini, Grok, Perplexity
    - Use reasoning (thinking) models when the tasks are complicated (in particular, math and coding problems)
  - Tool use
    - Internest search (Perplexity is good at this)
    - Deep research (a combination of internet search and thinking)
    - Upload papers or books to LLM to help you read
      - A missing tool in the market: highlight a paragraph and then ask questions about it. Currently we have to copy and paste the paragraph, back and forth several times
    - Python interpreter (i.e. computer programes)
    - GPT 4o's Advanced Data Analysis (making plot, extrapolating)
    - Claude's Artifacts (e.g. make conceptual diagram of paper, chapter, book, source code, etc)
    - Cursor: Composer
  - Modalities
    - Use voice as input, it's faster than typing. Audio input(SuperWhisper, wisperFlow, MacWhisper, etc.). Audio output (usually within app)
    - Native audio input/output (Grok has more freedom than ChatGPT)
    - Podcast generation and interaction (NotebookLM)
    - Image input (see, OCR, ask about...)
    - Image output (DALL-E)
    - Video input (Advanced voice + video)
    - Video output (Sora, many many others)
  - Quality of Life features
    - ChatGPT Memory feature (save the text and always prepend it to the beginning of the prompt)
    - Custom instructions
    - Custom GPTs
Learning diffusion, flow/score matching model
AI tools
- Google's Illuminate which transforms text content, for example papers, into engaging audio discussions, e.g.
  - Attention is all you need is converted to a 4 min audio discussion
  - FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness is converted to a 6 min audio discussion
GPU and CUDA
For fun
- Lee Sedol vs. AlphaGo: The Human Mind’s Last Stand Against AI (a YouTube video)
- AlphaGo - The Movie

back