From the course: Becoming a Good Data Science Customer

Large language models (LLM)

From the course: Becoming a Good Data Science Customer

Start my 1-month free trial Buy for my team

Large language models (LLM)

“

- [Narrator] The world of NLP is a rapidly changing area of data science. Large language models, known as LLMs, such as GPT, LLaMA, and PaLM 2 have radically changed the world of NLP. These models fall under the category of generative AI because of their ability to generate new content. Intuitively, they work by generating the most likely token or phrase to come next. Just as you might fill in the blank of, "I'm going to the store to buy some blank," with a word like milk, LLMs leverage their vast training corpus to understand the probabilistic relationships between words, allowing them to predict or generate the most likely sequence of words that follow a prompt. Using simple prompt engineering, users can now perform effective sentiment analysis, extracting meaning from large sets of text and other NLP tasks with minimum effort. The most effective way to use LLMs is to combine them with high quality domain specific data sets. Data is the most valuable asset, not any specific model. With advances in open source models, the barriers to deploying a large language model have dropped. But only you have access to your organization's proprietary data. An out of the box LLM is like a new graduate hire, a semi-competent generalist. To be truly effective, they need to be trained on the specific competencies required to succeed at your company. They need to become a specialist. Likewise, the best way to optimize the performance of an LLM is to train it or fine tune it on your company's data. While these tools are extremely powerful, approach with caution. LLMs are known to hallucinate, can be expensive to deploy, and may be susceptible to cyber attacks. Work closely with your security team before exposing your data to an LLM. Multimodal LLMs can analyze text, video, and speech, and are becoming more available. LLMs can be used to draft computer code as well as sit underneath many applications, now including Copilot that is used for programming and search engines. The capabilities of LLMs are expanding rapidly and the potential for applications continues to grow. For companies that are considering using LLMs, I advise that you start with the problem you're looking to solve and not with the tool you want to deploy. That said, if your data science team is using LLMs, then a few questions that can be asked include, what analysis did you perform before implementing the LLM? Did you use generative AI? If so, were you using an open source app or an API to use your prompt engineering? If you used an LLM, what security measures did you take to ensure that the data fed into the model would be safeguarded? Please try to get comfortable with the use of LLMs, even as a user of their products like Gemini and ChatGPT. The progress that has been made recently has been astounding and more advances are coming soon.

Contents