Digits

It's becoming clear there are quickly emerging sets of use cases that lend themselves well to LLMs (in their current state). In our survey about using "LLMs in Production" participants revealed some common themes. That's not to say these are the only use cases with LLMs. With the emergence of higher-quality agents, I think we will start to see new areas of exploration soon. How reliable those agents are is still up for questioning. 🤷♂️ Here are some of the main use cases we identified from the survey: • Niche Chatbots • Customer Support • Coding • Data Enrichment • Product Recommendations Let's go deeper on "Niche Chatbots" and "Data Enrichment". Talk with your data projects seem to be the Hello World of working with LLMs. There's plenty of support and tutorials out there to create your own spin. Whether it's asking questions about a corpus of audio data (like the MLOps Community podcast 😉) or it's asking questions about your company's internal knowledge hub, the applications lend themselves well to getting your feet wet in the AI space. Two particularly creative "talk to your data" use cases from the survey were: 💎 Helping managers get up to speed on different projects quickly. Give managers the ability to ask the status of projects, budget, team, blockers, etc (I am honestly surprised Jira hasn't added this functionality yet). 💎 Coaching subject matter experts. ie asking accountants or lawyers questions at different stages of their workflow. The tables have turned. The chatbot is actually prompting the human to make sure they thought about all vectors of the problem. (Hannes has some great material on how they do this at Digits with accountants). Also, I know ^^ this one isn't exactly a chat with your data example but it kinda is... I think you can make the logical conclusion and see how I got from A to B. As for data enrichment. If LLMs are doing all the work labeling the data, it raises a few issues yes. But it also has the potential to bring down the cost greatly. I am not convinced a model can do better than a high-quality labeling service like Surge AI... but I mean if the base model is good enough... alpaca proved its viability for that use case. So like always, it matters at what stage and for what end game you are using the LLM. Do you need the high-quality human-labeled data or is some LLM stuff good enough?