"No one is harder on a talented person than the person themselves" - Linda Wilkinson ; "Trust your guts and don't follow the herd" ; "Validate direction not destination" ;

June 03, 2023

Promises and Lies of ChatGPT - understanding how it works



Key Notes

Basics
  • ChatGPT is the idea of n-gram models
  • Given n-1 words guess nth word likely to be
  • Distribution is learnt from sequence
  • People tried in small values of n
  • Sample from distribution of words
  • More likely words more often

With large data
  • Any N, Words next word
  • Frequency, Conditional probability
  • Generate words if the first word given
  • More likely words + Patterns
Large sentences/meanings
  • Abstract sequences
  • Different answers every time
  • Every sequence may be different generated distributions but a similar context is possible
  • Chatgpt = something well written
Why it works?
  • We believe in what seems realistic
  • Connect to human experience
  • Fact is different from possibility
  • Plausible or probable or reasonable answers 
Similarity to humans
  • Humans are not always factual
  • It can be perception based
  • People can be finalized in civil society
  • Machines can suggest without knowing the consequences
  • Automation still may have a bias
  • Being close to the truth we are impressed
Predictive modeling
Train / predict
Conditional modeling
  • Can create bias in information
  • Discriminate learning learns a conditional model
  • Classifier then finds dogs vs generates dogs both different
Generative distribution - Joint distribution
  • The prior distribution of reasonable images
  • Teacher = Generative model
  • Learning generative model is costlier
The human brain works by on-demand stitching
  • chatgpt does something similar
  • All learning is compression
  • All learning is lossy compression
  • jpeg lossy - approximating
  • Representation of compressed details
  • Significant footprint available to train systems
Good writing for all
  • Picaso style pics
  • Shakespeare style writing
  • Racial profiling not required
  • Character and form are not connected
  • Generalizations help for survival
  • AI as creator / editor
Badly written with original thought is human writing
  • Harder to write original creative ways
  • Original vs Derivative thinking
  • Bad handwriting vs Good content
  • Bad package vs Good product
  • We have one scale good or bad
  • LLM learns from human language
  • Most likely completion given soceity is
  • Social Enginner on Data
Is this a good representation of all ethnicity ? 


How it for fine tuned ?
  • RHLF
  • Show results
  • asks someone their likes
  • Thumbs up / down to change distribution
  • Re-learning it
  • Collectively offensive content on web vs making a decent prompt engine

  • Align to human values
  • Concentration campus, Genocide - Human values
  • Retrain for cultural norms
  • False positive
  • Different narrative, different takers

  • Make LLM overwrite conditional network through prompts
  • Adverserial learning prompts
  • How to put knobs how it behaves well

AI systems to work with
  • Basically put people to think about problem
  • With enough eye balls every downside can be shallow bug
  • We need more eyeballs to decide
  • ChatGPT will not generate grammatically incorrect sentence
  • Core problem of intelligent behavior - planning, diagnosis, reasoning


Keep Exploring!!!

No comments: