I see the term called out as an important competency for professionals and kids, but I am not sure the people using it could actually define it.
Here is what I would expect an AI Literate person to understand:
Large Language Model = deep Neural Network + lots of data. Rules emerge from data. Requires data centers and specialized hardware.
LLMs generate output based on probability.
More context leads to more predictable output.
LLMs simulate the output of a human, they do not simulate human intelligence.
LLMs reflect biases and harms from their data and training.
There is considerable manual human labor that is required to build LLMs (data labeling, human feedback)
It is difficult to understand why LLMs return the output they do.
Solutions leveraging LLMs need to mitigate bias and harm in their design as well as implement operational monitoring.
LLMs are effective at persuasion and can lead some users down harmful paths.
Small Language Model = Deep neural network + only relevant data. Can run on commodity hardware.
RAG = LLM + search.
Agent = LLM + goal + tools/resources.
Reasoning = LLM trained to spend more time on compute and to output an explanation for its answer.
Reasoning != how an LLM actually creates an answer.
Hallucination = LLM output that is false.
GPT = Generative Pretrained Transformer
- Generative – Creating new content
- Pretrained – ingesting data to build deep neural network. Taking months and millions of dollars of compute.
- Transformer – Leveraging the architecture from the 2017 Attention is all you Need whitepaper
Training = building the model
Post training – adjusting the model once built to drive specific behaviors
Inference = using the model