Most effective approaches to LLMs
Users have four options for incorporating their data into LLMs, with each option requiring differing degrees of complexity and computational resources.
Train. The most challenging and expensive of these options is to train LLMs from scratch, which requires an immense volume of training data and computing power to find the most optimal configuration for a novel neural network.
Fine tune. Given these enormous costs, a more common approach is to take a model that has been pretrained for natural language processing and fine tune its weights using proprietary data and corporate knowledge. Although this approach will produce a customized model tailored to fit the needs of its developer, the computing resources required for model retraining can still be an outsized expense for most large companies.
RAG. Retrieval-augmented generation (RAG) is a newer, cost-effective technique to embed more information into a single prompt. RAG works by transcribing the text from a document into a numerical representation (via word embedding) and then aggregating all resulting vectors into a database. This vectorized database contains semantic relationships across the document. Users can then search the database for pieces of information relevant to a given prompt. By augmenting prompts using RAG, users provide LLMs with greater context so they can yield custom responses without the need for users to directly tune the model’s weights.
Prompt engineering. The most straightforward and inexpensive way to use LLMs is prompt engineering, where individual users explore and carefully craft prompts that yield the most useful responses.
AI opportunities and limitations
LLMs recently gained the ability to interpret and analyze CSV files. By examining specific datasets, these models can provide customized insights and generate tailored code snippets for data analysis or visualization, enabling a more contextual and data-specific approach. For example, based on the structure, type, and trends within a CSV file, an LLM can suggest relevant statistical analyses, identify appropriate data cleaning methods, or even predict potential outcomes. This level of customization is pivotal for making more-informed decisions and simplifying the initial stages of data exploration.
However, it is important to distinguish this LLM approach from more Automated Machine Learning (AutoML). AutoML focuses on automating the end-to-end process of applying machine learning to real-world problems. AutoML systems can select the best model, apply hyperparameter tuning, and complete feature selections, all with minimal human intervention. In contrast, LLM-generated code for data analysis is more about guiding and assisting in specific tasks rather than automating the entire process. A hybrid approach combining the two could be highly beneficial: LLMs could assist in the initial data understanding and preprocessing, while AutoML could take over for model selection and optimization. This synergy could lead to more efficient, accurate, and accessible machine learning workflows, making advanced data analysis more approachable for non-experts and more efficient for experienced practitioners.
Generative AI technologies are also likely to undergo a drastic convergence. Currently, very distinct neural networks are used for image generation, text generation, audio generation, and self-driving cars. Future technology will likely unify these applications as we move towards general AI capabilities. General AI would function similarly to our own brains, with one technology accomplishing nearly all tasks. And while LLMs do not reason, at least not to the degree of humans, computational reasoning is another key advancement that may be on the horizon.
All indications point to 2024 being another year of rapid growth for generative AI. Generative AI will be as impactful as the internet revolution, and it is important for insurance professionals to embrace these emerging technologies in this new era.