User:Hiàn/AI

We've seen an explosion in the use of large language models since 2022. I think simplewiki is uniquely positioned to benefit from a thoughtful approach to the use of generative AI. Unfortunately, the project has seen some difficulty adjusting to the current climate (note the RfD backlog).

Here's what I think:

• First, we're in the unique position of a neutral and simple explanatory resource for AI/ML concepts. Interest in both fields have exploded, and a push to improve the quality and coverage of these articles is increasingly necessary. Our work in article-space unfortunately often lacks much usefulness to potential readers.
• Second, we should codify WP:LLM as a policy or guideline. This discussion appears to have stalled since its original introduction. There are some edge cases that should be clarified more explicitly (relating to less benign use (see this presentation) or communication), and some extra background information we could potentially pull from enwiki for.
• Finally, LLM-in-the-loop tooling is a promising avenue for improving editor productivity and article quality. I do not envision using LLMs as means for blindly generating text. But they could function to act as a smart linter and a second opinion for uncontroversial edits, possibly with fine-tuned small (<10B params) LLMs running locally.

Article-space

• Core machine learning topics
  • Supervised learning
  • Unsupervised learning
  • Reinforcement learning
  • Natural language processing
  • Computer vision
  • Classification
  • Linear regression
  • Contrastive learning
  • Error function (MSE, CE/BCE, NLL)
  • Metrics -- accuracy, precision, recall, support
• Core deep learning topics
  • Background topics
    • Gradient descent
    • Stochastic gradient descent, Adam
    • Tensor (at enwiki, this is a separate page at en:Tensor (machine learning
    • Backpropagation
    • Partial differentiation and automatic differentiation
    • Convolution
  • Basics
    • Artificial neural network
    • Activation function (softmax, ReLU)
    • Transfer learning
    • Regularisation, batch normalisation
  • Architectures
    • Multilayer perceptron
    • Convolutional neural network
    • Autoencoder
    • Recurrent neural network, LSTM, GRU
    • Graph neural network
    • Transformers, attention
    • Large language model
    • Generative adversarial network
    • Vector embedding
• Core artificial intelligence topics
  • Distinction betw. symbolic, sub-symbolic AI
  • Search, informed search (A*)
  • Alpha-beta pruning, game solvers
  • KR&R
• Tooling
  • NumPy, PyTorch, TensorFlow
  • sk-learn
  • CUDA