11-10, 11:00–11:45 (America/New_York), Music Box (5th floor)
With the advent of more and more large language models, the concept of Prompt Engineering is becoming increasingly relevant. It may not be long now until we see “experience with prompt engineering” as a requirement in job adverts. Prompt Engineering, in short, refers to the practice of changing the text prompt you give to a model to get the best results out of it.
In this talk we analyze what prompt engineering is, how to do it, and how much of a difference it can really make. We do this by comparing different prompts given to multiple large language models (and maybe some vision transformers) and compare the results. We’ll see that by changing key elements of the prompt, the results can be very different. We’ll use results from a research study Mantis has conducted on over 6 different language models with 5 different types of NLP tasks (classification, qa, ner, similarity and summarization).
In this talk you will learn:
- How to interact with large language models
- How to construct a prompt
- How a smaller model 11b can get close to a 150b model’s results just by using better prompts
- How prompts can make or break a model’s response
Previous knowledge expected
I've been a developer for over a decade now. I have extensive working experience in different areas of IT, and for the last 4 years I've transitioned to Machine Learning, out of passion for the field. I've taken certifications in Machine Learning and competed on Kaggle. I've specialized in Natural Language Processing (NLP), working on multiple projects in which I've learned how to combine the different techniques and models available today, and how to tackle the different problems in the field. I've spent the last 3 years building Intelligent AI assistants by developing NLP and ML approaches for the interface between the human and the device, and using regression methods. My main motivation is making an impact and creating wonderful and innovative products that help others and deliver the best value.