Maria Vectomova, Lead Machine Learning Engineer & Başak Eskili, Machine Learning Engineer at Ahold Delhaize, talked about the challenges of doing enterprise ML.

Standardizing the MLOps work process isn't about the ML tools that you use in building your MLOps framework. It's all about creating the golden task of how to do things. A huge chunk of the model development tasks carried out by data scientists are repetitive in nature. Their production process technically involves running a script with a set of tools like Databrick, Kubernetes, SageMaker, etc. Irrespective of the tooling combination, data scientists can make use of their reusable workflow and don't have to think about service users or service principles.

It is important that data scientist can integrate their workflows with the central team seamlessly. The number of tools in designing ML products is not as important as a set of tools with the right functionalities for the ML solution.

Video || Spotify || Apple

Simba Khadder, Founder & CEO of Featureform, shared the origin and working principle of the Featureform platform.

Featureform's recommender system is made up of a generation step that suggests recommendations and a ranking step that points out the most appropriate recommendation at a particular time. The process generally involves feeding embeddings into this deep neural network.

Essentially, embedding is generated by representing every user of similar attributes as clusters in a vector space. This creates a holistic view of users in 3D space that explains a particular user implicitly.

YouTube

Adam Nolte, CTO, and Co-founder at Autoblocks, talked about ensuring accuracy and quality in LLM-driven products.

Autoblocks is a company that provides enterprise-grade features to build, deploy, and monitor LLMs at scale.

An LLM product refers to a product or service powered or significantly enhanced by a large language model (LLM) like GPT-4 from OpenAI. Large Language Models are advanced artificial intelligence systems that excel at understanding and generating human-like text based on the context and input provided.

Incorporating human feedback in LLMs is an efficient way of knowing how well the model is doing in production. The user's satisfaction with the model's performance can be measured using basic "thumbs up, thumbs down" questions.

To understand human behavior beyond the "thumbs up, thumbs down", more product analytics-type questions are needed that enable the developers to think deeply about;

- The failure modes that can occur in the product
- The most likely cause of a negative user experience
- The positive outcomes
- Adding tooling to measure the human outcomes
- Using the information to improve the LLM-driven features

This blog was written by Niels Bantilan

Fundamentally, prompt engineering is about getting the model to do what you want at inference time by providing enough context, instruction and examples without changing the underlying weights. Fine-tuning, on the other hand, is about doing the same thing, but by directly updating the model parameters using a dataset that captures the distribution of tasks you want it to accomplish.

This blog describes prompt engineering and fine-tuning in more detail. It gives you a practical sense of how they are different, and provides you with a few heuristics that will help you begin your fine-tuning journey.