13 - Epilogue Flashcards
What are LLMs?
LLMs are Large Language Models that predict the next word in a sequence based on prior words.
What is theory of mind?
Theory of mind is a cognitive ability that allows humans to make inferences about someone else’s beliefs or state of mind using external behavioral cues.
How does ChatGPT demonstrate theory of mind?
ChatGPT can infer that Alice will experience a headache after using the wrong glasses, based on the context provided.
What does the training of an LLM involve?
It involves predicting the next word in a sequence and adjusting parameters to minimize the loss between predicted and actual words.
What is word embedding?
Word embedding is the process of converting words into vectors embedded in high-dimensional space.
What is the function that LLMs approximate during training?
LLMs approximate a conditional probability distribution for the next word given a sequence of input words.
What does the softmax function do in LLMs?
The softmax function converts output vectors into probabilities that sum to 1.
What is generative AI?
Generative AI learns a probability distribution over data and samples from it to produce outputs.
What is emergent behavior in LLMs?
Emergent behavior refers to capabilities that arise in larger models that smaller models do not exhibit.
What issues were prevalent in AI before LLMs?
Concerns about bias and discrimination in AI systems were prevalent.
What was a notable incident of bias in AI?
Google Photos mistakenly tagged African Americans as gorillas, highlighting issues of racial bias.
What is a concern regarding data representation in ML?
Incomplete data can lead to biased predictions, particularly if certain groups are underrepresented.
What is the difference between correlation and causation in ML?
ML systems can mistakenly assume correlation implies causation, leading to erroneous predictions.
What must ML engineers do to avoid bias?
They must ensure training data is diverse and representative and may need to explicitly de-bias the data.
True or False: LLMs are capable of reasoning.
This is debated; some see LLMs as sophisticated pattern matchers, while others see glimmers of reasoning ability.
What is a significant advantage of LLMs for programmers?
LLMs can generate code based on natural language descriptions of problems.
Fill in the blank: LLMs are trained on a corpus of _______.
[training text]
What does backpropagation do in LLM training?
Backpropagation adjusts the network’s parameters to minimize prediction loss.
What is the size of GPT-3 compared to GPT-2?
GPT-3 has 175 billion parameters, while GPT-2 has 1.5 billion.
What is the role of the neural network in an LLM?
It acts as a function approximator for predicting the next word in a sequence.
What does the term ‘stochastic parrots’ refer to?
It refers to the notion that LLMs may simply repeat patterns without true understanding.
What is a potential danger of LLMs?
They can perpetuate and amplify societal biases present in the training data.
What is one way bias can enter machine learning?
Through the use of incomplete or unrepresentative training data.
What is the significance of the Simons Institute workshop in 2023?
It highlighted ongoing discussions and research on the implications of LLMs.
What is bias in the context of AI algorithms?
Bias refers to the inaccuracies in the algorithm’s predictions, influenced by how data is interpreted and the questions posed to it.
What is a significant concern regarding LLMs as discussed in the text?
Concerns include AI being biased, toxic, or dangerous.
What example is provided to illustrate bias in AI?
The interaction where GPT-4 identifies the nurse as pregnant in a sentence, showcasing a sexist interpretation.
What is RLHF?
Reinforcement Learning using Human Feedback, a technique to fine-tune AI models.
What is the implication of certainty in AI predictions according to researchers Celeste Kidd and Abeba Birhane?
AIs that make predictions with certainty, regardless of factuality, risk altering human cognitive beliefs.
How do humans form beliefs based on data?
Humans sample a small subset of available data and form beliefs with high certainty, making them stubborn to revise.
What is the problem of credit assignment in neural networks?
The challenge of determining how to adjust the weights of connections in a network when it makes an error.
What is backpropagation?
An algorithm used to train artificial neural networks by adjusting weights based on the error of predictions.
What did Daniel Yamins discover while working on his machine vision project?
The architecture that worked best for recognizing objects was a convolution neural network (CNN).
What was the focus of Yamins’ research in relation to CNNs?
To see if a CNN could predict biological neural responses to novel images.
What did the researchers find when they compared CNN activity to monkey brain activity?
The CNN predicted the behaviors of brain areas corresponding to the layers of the network.
What was a significant finding from DiCarlo’s lab regarding AlexNet?
AlexNet was used to model the ventral visual stream of macaques, correlating artificial neuron activity with monkey neural sites.
What is the ventral visual stream responsible for?
Recognizing people, places, and things.
What are LLMs beginning to hint at in terms of human cognition?
Hints of theory of mind and complex pattern matching.
What ongoing debate in cognitive science is mentioned regarding language acquisition?
Whether aspects of human language depend on innate abilities or can be learned through exposure.
True or False: LLMs can learn syntax and grammar from statistical patterns in human written language.
True.
What is a key difference between biological neurons and artificial neurons?
Biological neurons spike, while artificial neurons do not.
What is the energy consumption comparison between LLMs and human brains?
LLMs consume about 1,664 watts, while human brains use 20 to 50 watts.
What is a potential requirement for achieving human-like general intelligence in AI?
Whether disembodied AIs can develop such intelligence or if they need to be embodied.
Fill in the blank: The backpropagation algorithm is used to train _______.
artificial neural networks.
What does the phrase ‘form follows function’ imply in the context of CNNs and brain activity?
The structure of the CNN corresponds to the functions of specific brain areas.
What is one key challenge in AI that needs to be addressed alongside its development?
Concerns about bias and the impact of AI predictions on human cognition.
