Study by Summary
Black Friday 1 year Manning subscription and ACloudGuru handson AWS, Azure, GCP
Summarize the Manning per chapter summaries of all my purchased books.
End of each Chapter
https://www.manning.com/books/causal-ai
13.4 Summary Large language models are powerful AI models that generate text and other modalities and achieve high performance across a variety of benchmarks. LLMs have proven use cases for supporting causal analysis. LLMs can help build a causal DAG. The LLMs can provide code implementations of that DAG and analysis built upon that DAG using common libraries. LLMs can leverage common and expert knowledge about causal relations and mechanisms. The causal frame problem is the challenge of selecting the causal variables relevant to a given problem and excluding the irrelevant. Cutting-edge LLMs emulate how humans set the causal frame, which is useful for applications such as building DAGs and root cause analysis. LLMs can help us understand nuanced causal concepts and how to contextualize them within our domain of interest. LLMs can help us put causal queries into formal terms. However, LLMs are prone to hallucination, convincing yet incorrect responses to our queries. At their core, LLMs are a probabilistic machine learning models that model a joint probability distribution on sequences of tokens. The attention mechanism enables the LLM to learn higher-level representations that make cutting-edge LLMs so powerful. However, just because it learns a higher-level representation, doesn’t mean it learns a causal representation. Even if it did work in some special cases, it would be hard for the user to verify that it is working. Alternatively, we can build our own causal LLM by using a transformer architecture on a causal DAG scaffold. This allows us to work with an LLM, while admitting causal operations, such as a do-operator. Use the causal hierarchy theory as your North Star in the exploration of how to combine causality with large language models and multi-modal models, as well as exploring how well these models can learn causal world models on their own.