#ai #llm Created at 050723 # [Anonymous feedback](https://www.admonymous.co/louis030195) # [[Epistemic status]] #shower-thought Last modified date: 050723 Commit: 0 # Related - [[Large language model|LLM]] # 050723 12238 Classical computer science meet LLMs - Binary search: This could help with efficiently searching through the huge datasets required to train LLMs. Binary search can make lookups much faster than linear search for these massive datasets. - Graphs: Representing language as a graph, with words as nodes and edges as relationships between words, could help address some of the consistency and accuracy issues. Graph algorithms like breadth-first search, depth-first search, and shortest path algorithms could be useful for navigating and reasoning over these language graphs. - Trees: Decision trees and tree ensembles could be used to make LLMs more controllable and address bias. The trees could be used to model complex linguistic decisions and relationships. - Bloom filters: Bloom filters are a space-efficient data structure that could be used to check if an element is in a set. This could help with data privacy by allowing you to check if private data exists in an LLM's training data without revealing the actual data. - Caching: Caching techniques like memoization could help make LLMs more efficient by caching the results of expensive computations. This could help reduce the cost of using and maintaining LLMs. - Differential privacy: Differential privacy is a technique that could help ensure data privacy for LLMs. It allows you to query a dataset while limiting the amount of information leaked about any individual data point. This could help address data privacy concerns with LLMs. - Hashing: Hashing can be used to vet and ensure the integrity of the data used in training LLMs. This can help address issues related to data privacy and security. For example, hashing can be used to detect duplicate entries or to verify that the same data set has not been modified or tampered with. - Optimization algorithms: Optimization algorithms like gradient descent can be used to make LLMs more cost-effective. These algorithms can be used to optimize the parameters of the model, reducing the amount of computational resources required to train and maintain the model. - Sampling Techniques: Techniques like Monte Carlo simulations can be used to ensure that the training data for LLMs is representative and does not contain malicious content. This can help mitigate the risk of the model generating malicious content.