What I think is true of these 4, somewhat related things.

Model compression

• Make models smaller or faster
• Why do we care about smaller models?
  • Less space
  • Less computation for inference
  • can deploy these on the edge
  • Latency
• Maybe we don’t care about all the juice the larger model has

Model quantisation

• A method of model compression
• Use less bits in the model parameters
• 32 -> 16 or something

Model distillation

• A method of model compression
• Transfer knowledge from a larger “Teacher” model to a smaller “student” one
• Distillation can also be used beyond compression. For example, to transfer capabilities between model architectures.

Model fine tuning

• Adapt a pre trained model to do even better on your task
• Can avoid fine tuning by just prompt engineering