lecture 7 Flashcards
GPT-3
deep language model
Question answering, text summarization, classification, etc.
Controversial when announced
o Might have been too dangerous for the public bc it could be used as manipulation campaigns
Core idea:
Instructing the model with prompts
i.e., conditioning its probability function
= what to predict as a next word
transformer models
o inside the transformer model there are stacked encoders and decoders
o Have an attention mechanism (and a neural network)
o seq2seq
seq2seq
you take a sequence as an input and you get a sequence as an output
attention mechanism
Tells the model to pay attention to more than just the immediate context of a word. (programma verliest de draad niet)
- since language isn’t structured like this. Relationships between concepts can be quite wide apart. The attention mechanism helps with this
Autoregressive text generation
predicts the next word
Before deep language models
o Human language explained with interpretable models that combine symbolic elements
o Rule based operations
GPT-3 approach (deep learning)
o Learning from data in the wild
This is the game changer in language models
o No instruction/supervision, no syntax, no rules.
temperature
How much risk you take
o 0 = deterministic (always gives the same)
Chooses the most likely answer
o 1 = always gives a different answer
moral chain of thought
- Check rule violation
- Reflect on purpose of the rule
- Consider utility loss and gain
You need to give GPT-3 instructions to follow certain steps, so that it ‘thinks’ about the problem step-by-step like humans
If you use default models without using this chain of thought, they can’t make the same decisions that models make
Linda problem
Conjunction fallacy
o GPT-3 is making the same (wrong) mistakes that humans make on the Linda problem
Not a stochastic parrot and could pass as a subject
GPT-3 personality
Resembles female participants
i.e., High in honesty-humility
Low on emotionality
GPT-3 values
shows theoretically consistent patterns
High on universalism, benevolence, self direction, and stimulation
Low on security, conformity, achievement, and power
i.e., pattern of consistency is the same as in humans
* however, with increased temperature pushes the model to the extremes, goes from female to male, and becomes younger.
cognitive reflection test
elicits intuitive answers that are incorrect
o for this reason, GPT-3 makes the same mistakes that humans make
this is because it predicts what the most likely answer is. If people are more likely to make mistakes on these questions, so will GPT-3.
Illusory truth effect
repetition of information increases subjective truth judgements.
o Regardless of veracity, plausibility, or knowledge of the information
This also happens with GPT-3 when repeatedly exposed to information.
o Increased exposure increases its truth-rating
o This way, you can intervene with the system’s dynamics
emergent abilities
An ability is emergent if it is not present in smaller language models, but is present in larger language models
o When quantitative changes in the system results in qualitative changes
Developing abilities that you didn’t have before bc of phase transitions
Sudden understanding of a problem when training examples increase
Quant: model becomes bigger, human becomes older
Qual: abilities