Artificial Intelligence

Anything about computers learning and "thinking".

Learning AI provides us a way to learn about ourselves, how does human consciousness and intelligence work?

Resources

• Fast AI Course
  • Full free online course on deep learning
  • Fastbook, YouTube Playlist
• Neutral Networks and Deep Learning | Michael Nielson
  • Learning neural networks by building one to recognize handwritten digits

Courses

• Neural Networks: Zero to Hero (Website)
• Stanford CS229: Machine Learning by Andrew Ng (Autumn 2019)
• Standford CS221: Artificial Intelligence: Principles and Techniques (Autumn 2019)
• Standford CS231n: Deep Learning for Computer Vision (Spring 2017) (Course Website)
• MIT 6.034: Artificial Intelligence by Patrick Henry Winston (Fall 2010)
• MIT 6.S191: Introduction to Deep Learning (Spring 2018 - Spring 2023)
• MIT 6.S099: Artificial General Intelligence (Spring 2018)
• Deep Learning Foundations to Stable Diffusion | Fast.AI (Website)
• Machine Learning Video Library | Caltech — Videos categorized by topics in machine learning

Links

• The Waluigi Effect Meta Post (HN)
  • An interesting hypothesis is that some prompt has adverse effects on the model
  • "After you train an LLM to satisfy a desired property P, then it is easier to elicit the LLM into satisfying the opposite of P"
  • Could like to see more evidence and examples though
• What are transformer models (HN)
  • If you want a higher-level explanation with examples of the process of tokenization, embedding, positional encoding, the transformer block, attention and the softmax layer
  • If you have the patience, try reading the original paper. The top comment in the HN post summarized a few oversimplifications or mistakes in the article.
• A Cookbook of Self-Supervised Learning (PDF) (HN)
  • 70 pages cookbook lowering the barrier of entry for training AI with self-supervised learning
• Self-healing Code is the Future of Software Development | StackOverflow
  • Give a goal to the AI, the AI uses its output to prompt itself again
  • E.g.: feed the AI code, compilation error or execution logs back to the AI and ask the AI to improve
  • These loops are prone to distractions though. "Technology marches on, but procrastination remains unbeaten." lol
• How to make a QR code with Stable Diffusion | Stable Diffusion Art
  • This works for me, things like city, buildings and castle works well
  • It uses the image-to-image with ControlNet to "paint" the QR code
  • The link to the GhostMix model it uses
  • The later the ControlNet starts or the earlier it ends the photo is more realistic, but risking of losing the QR code
• Embeddings: what are they and why they matter | Simon Willison
  • Embeddings: turn any piece of data into a fixed-length vector
  • Finding related contents and relationship between words, e.g. germany + (paris - france) = berlin in Word2Vec
  • CLIP: multimodal embeddings for both images and text provide interesting image searching capabilities
  • Retrieval-augmented generation (RAG): generate embeddings for a collection of documents, use embeddings to find similar documents with the question, paste original excerpts of the documents alongside the question to the LLM to generate answers
• Machine Learning Engineering Open Book
  • Collections of concepts, methodologies, tools and tips to train LLM and VLM
• Ask HN: What have you built with LLMs?
  • ChatGPT Phone to practice cold calls
• Building the next generation of AI infrastructure at home | IFP
  • How to Build an AI Data Center
    • Cost, power, computing devices, network, cooling, reliability
• Summary of Ilya Sutskevers AI Reading List
  • A list of short summaries of the 30 papers reading list
• The society of mind | Marvin Minsky
  • A book proposing a model where intelligence is like a society of agents
  • A way to model artificial intelligence, and maybe a way to model how human mind works
• Why the deep learning boom caught almost everyone by surprise
  • The story of three stubborn non-conformist
    • Geoffrey Hinton promoting neural networks
    • Jensen Huang recognizing GPUs can be used to train AI and developing CUDA
    • Fei-Fei Li creating ImageNet
  • And the story of how the three converges to form modern AI
• AI agents switch to sound-level protocol after confirming both are AI agents
  • A cool video demo, two AI agents on phone confirm both are AI assistant, then switch to communicate over a data-over-sound protocol (ggwave)
• Here's how I use LLMs to help me write code | Simon Willison
  • It is a difficult and unintuitive tool, to use it well requires understanding
  • Reasonable expectations

    If a human collaborator hallucinated a non-existent library or method you would instantly lose trust in them. Don’t fall into the trap of anthropomorphizing LLMs and assuming that failures which would discredit a human should discredit the machine in the same way.

  • Iterative conversation, ask for options or go authoritative, vibe coding to learn
  • With one use case demo in the blog
• AI Blindspots (HN)
  • Blindspots in LLMs discovered while using AI for coding
  • For example, AI does not know when to stop digging on a problem and reevaluate the options
  • Some will eventually go away, e.g. capability on Python and JS is better than statically typed language
  • Some can be partially solved with prompting, like lost of context
  • One HN comment says it well, LLMs make a different set of mistakes which human are not trained to catch
• Code like a surgeon
  • Surgeons leverage a support team to be productive
  • Focus on the important bits that only the surgeon can handle instead
• My AI Adoption Journey
  • (1) Drop the chatbot: use agent because of the automatic feedback loops
  • (2) Reproduce your own work with AI to learn what can and cannot be done, and how to do it
  • (3) End-of-day agents: block out end of day time to explore what can be done to gain some efficiency
  • (4) Outsource the slam dunks: delegate work that AI can do, work in parallel for other work that AI can't
  • (5) Engineer the harness: make sure AI don't make the same mistakes again
  • (6) Always have an agent running: force yourself to improve workflow and delegate
• Using AI to write better code more slowly (HN)
  • Rather than write low quality code faster, use AI to write higher quality code slowly
  • AI models are getting good at finding bugs, from critical security or correctness bugs, to mundane performance issues, to careless typos in test descriptions and comments
  • Personally I found this Gemini Code Reviewer skill is pretty good