Saltar al contenido ForoHackers

Generative AI: start here

IA generativa
#345 · Ophuscado · · Denunciar

Generative AI is models that create new content, such as text, images, audio, video or code, by learning the distribution of the data they were trained on and sampling from it. It speeds up drafting, design and prototyping. The same ability produces plausible falsehoods and raises questions of provenance, so how you check, label and use the output matters as much as how it is made.

Key concepts

  • Generative modelling: learn the probability of data, then sample new examples from it.
  • Model families: autoregressive models, variational autoencoders, adversarial networks and diffusion models, each with different quality, speed and control.
  • Conditioning: steering output with a prompt, a class, an image or a style, often through guidance.
  • Latent space: a compressed representation where similar content sits close, so you can interpolate and edit.
  • Training data: coverage, duplicates, licences and bias in the data show up in the output.

Vocabulary

  • Sampling: drawing an output from the model's distribution, with settings that control randomness.
  • Diffusion: generating by gradually removing noise from a random start.
  • Mode collapse: a generator that produces only a narrow set of outputs.
  • Prompt engineering: shaping the input to get the output you want.
  • Synthetic data: generated examples used to train or test other systems.

What to study, in order

  1. Probability: distributions, likelihood and sampling from a simple one in code.
  2. Train a small autoencoder on an image set and look at its latent space.
  3. Implement a minimal autoregressive generator for characters or pixels.
  4. Study how diffusion works and train a tiny one on a toy dataset.
  5. Evaluate outputs: sample quality, diversity, memorisation checks, and a human review with clear criteria.

Related tags

#deep-learning #llms #ai-evaluation #ai-safety #computer-graphics

Inicia sesión o únete para responder.