"Machine learning models are like a language for computers, and embeddings are the words." I often use this analogy to describe machine learning (ML) models and their corresponding embeddings, whether in a one-on-one conversation or during a keynote presentation. I've been working in and around machine learning (primarily in the subfield of computer vision) for nearly a decade since I graduated. This analogy is a succinct way to explain the importance of embeddings to a broader audience.
Just as the human brain has a dedicated region for memories and emotions (the hippocampus), machines also need a permanent solution for storing and indexing a machine's words and thoughts. This area is where the vector database comes in - it's a hippocampus for machines. It is exactly what we're building at Zilliz.
Zilliz's mission wouldn't have resonated with me if I didn't understand the power of embeddings. In this blog post, I'll take you through my journey in this space and show you how I ended up at Zilliz.
A quick primer on embeddings
For decades, we've tried to "teach" computers to understand the world as we do. Early attempts at creating chatbots, for example, were centered around recognizing keywords and phrases, creating the illusion of general understanding. Until recently, we've made incredible progress toward a general-purpose chatbot with LLMs such as Claude, Bard, and ChatGPT, spawning the possibility of intelligent "agents" that can plan and execute complex tasks. At their core, these chatbots are specialized neural networks - ML models trained with one of the many flavors of stochastic gradient descent. If you are unfamiliar with neural networks, you can think of them as large computer models that use successive "layers" to build powerful representations that are impossible with handcrafted algorithms.
At the core of all ML models are concepts known as embeddings, which are high-dimensional vectors. They provide an abstract but compelling way to represent input data in the model. These embeddings have unique properties, but I won't cover them in this post. If you're interested in learning more, you can read my post on vector search, which covers most of the basics.
My (many) run-ins with vector search
My first encounter with the potential of embeddings was in 2014 when I began working with neural networks at Yahoo. At the time, machine learning was still the "wild west," and no tools were available for neural networks. Container-based orchestration platforms just started gaining popularity (Docker's first version was released in 2013). It was an incredibly exciting time to be involved in computer vision and machine learning.
As part of our efforts to enhance Yahoo platforms and services with machine learning capabilities, we became an early adopter of vector search. This decision resulted in a months-long project to bring large-scale semantic search to Flickr, a product Yahoo owned at that time. Although I was not directly involved in this project, I followed it closely. An early iteration of this vector search was incorporated into Vespa, another player in the vector database space.
Around the same time, many large corporations started to realize the potential of vector search, especially for applications in computer vision, such as image recognition. Although I appreciated the power of neural networks and embedding representations, I itched to work more in hardware since I studied Electrical Engineering in college. As a result, I left that field and founded an indoor localization company based in Shanghai. Over the next 2-3 years, I learned a big lesson - starting a hardware company is challenging.
In 2019, we pivoted to using machine learning and streaming data from inertial measurement units (IMUs). Embeddings played a crucial role in our success, and we secured contracts to deploy our solution for various customers, including many Fortune 500 companies. I continued working on the startup for two more years until we finally broke even. At that point, I decided it was time for a new adventure.
Fast-forward to 2021, and I had the opportunity to sit down with Charles (our CEO) and Robert (Head of Product). Over the years, embeddings and vector search have become central to my identity and success within the broader ML space. So, Zilliz's mission immediately resonated with me. After reviewing the broader vector database space, I was surprised that the market lacks an affordable and scalable vector search solution despite it being a well-known and powerful tool. The capability for vector databases to serve as the de-facto storage solution in the age of AI/ML was clear to me. So, I took the leap here.
Democratizing enterprise AI infrastructure
At Zilliz, we have encountered some challenges when adopting scalable, cloud-native vector databases. As powerful as embeddings may be, ML engineers (myself included) often underestimate the importance of infrastructure and tooling. We are builders at heart and almost always prefer to spend time wrangling the training data or perfecting the model architecture rather than worrying about how to deploy an application to production. In the realm of embeddings, Zilliz is firmly situated at the center of a broader effort to democratize enterprise AI infrastructure.
Welcome to join us!
If this message resonates with you, please let us know! We have a variety of openings in GTM, product, and engineering. If you are also passionate about recommender systems, semantic search, and making computers more "human" in general, feel free to check out our careers page.
- A quick primer on embeddings
- My (many) run-ins with vector search
- Democratizing enterprise AI infrastructure
- Welcome to join us!
Content
Start Free, Scale Easily
Try the fully-managed vector database built for your GenAI applications.
Try Zilliz Cloud for Free