How USearch Reached 500k+ Python Downloads 🌐

Join Slack • All Events • Jobs

Deep Dive

Today, we’re featuring Ash Vardanian, the founder of Unum.

Unum is a startup on a mission to build the most scalable infrastructure for AI. Ash started Unum in 2015 as a research project, and it’s now at a dozen researchers and engineers.

In 2023, Unum released USearch, an open-source library for k-NN that is now one of the most highly-deployed vector-search libraries - with 525k Python downloads and millions across other ecosystems. Often referred to as the “SQLite of Search”, USearch is currently used in ClickHouse and Postgres-like databases, is hardware-friendly and works across mobile phones and high-powered servers.

Ash is also a member of Cerebral Valley and co-organizes the largest Systems Engineering and C++ groups across the Bay Area and Armenia. Let’s dive in!

Our Chat with Ash 💬

Ash - you’re the first founder we’re featuring for the CV AI newsletter. Could you introduce yourself and share your background? 

Hey! My name is Ash, and I’m the founder of Unum. I grew up in Russia to an Armenian family, and found computers aged 8. I started coding in BASIC and building apps (fun fact - from 2009-2013, I built some of the first educational apps on the App Store). After high-school, I got bored with CS and focussed on astrophysics instead - which led me to stay in academic research until 20. 

Tell us about your journey to AI. What led you back from academia? 

In 2015, I returned to applied CS and decided to devote my career to the pursuit of AGI. My thought process was: AGI is going to fundamentally change how our world works, and I wanted to help humanity get closer to a general intelligence. Over the past 8 years, it’s become clear that AI needs a better understanding of systems engineering and high-performance computing, and I’ve spent my career researching and experimenting in these areas.

Why did you start Unum? 

I started Unum as a side-project in 2015. I technically never graduated, and all my research was based on my own interests. Back then, I was designing better algorithms for physics simulations, graph processing and neural networks. Eventually, I decided to put the three together to solve a problem I was facing IRL: building a dating app for myself!

I wanted to analyze large public datasets (social networks) with neural nets, to find a partner with similar interests. On a budget, I had to implement the AI models in C or Assembly and reimplement everything from math to networking and storage. In the end, I forgot about dating and focused on infra.

Among other achievements, our team at Unum built the first-ever GPU-accelerated transactional persistent database engine as one of our research projects. Building a custom cluster for its benchmarks was the perfect basis for the work we do now.

How would you describe Unum as it exists today, 8 years later?

Today, Unum is mostly focused on open-source. AI is the new platform, and we want to ensure that the next million apps will be built on solid developer-friendly infrastructure utilizing the most advanced hardware capabilities and most recent algorithms. 

We maintain several Open-Source libraries available under Apache 2.0, free for commercial use. Most of them are focused on scalable Semantic Search. This includes light-weight embedding models (UForm), low-latency networking (UCall), and of course, efficient indexing and retrieval (USearch).

Tell us about U-Search, Unum’s hugely-popular vector search engine.

USearch is an open-source library for k-NN that is now one of the most highly-deployed vector-search libraries - with 525k Python downloads and millions across other ecosystems. Often referred to as the “SQLite of Search”, USearch is currently used in ClickHouse and Postgres-like databases, is hardware-friendly and works across mobile phones and high-powered servers.

USearch is also one of the fastest engines out there, because even at the level of Assembly, we used the kinds of instructions that most compilers, let alone engineers, don’t use. Specifically, we make use of SIMD (Single Instruction, Multiple Data) - which is a type of parallel programming that allows a CPU to process multiple data with a single instruction. It’s like parallelism, but at the level of a single core.

Most vector search engines use out-of-date versions of SIMD - we use the newest specs across both x86 Intel and AMD CPUs. Knowing how each CPU works, and how many bytes at a time it can process, gives USearch an advantage. Especially for AWS and Azure, which now rely on fast and cost-efficient home-grown Arm CPUs that most libraries can’t implement. You might be easily losing 50% of the value or more, not using USearch.

How has GenAI changed your approach to Unum or AI research? 

It’s changed the way people perceive us, but the technology is still the same and it still has the same issues. I’m not a proponent of Transformer-based models, despite releasing them as part of UForm. I don’t believe in the ability of such statistical models to generalize or reason.

That said, AGI is the future and the rise in AI enthusiasm has helped us immensely. In 2015, nobody cared about AI like this - even Sam Altman was being called crazy by investors! Today, probably a billion people already feel that we are on the brink of a new age of discovery.

What are your thoughts on multi-modal models and their potential impact on research?

Everyone knows that multi-modal is expected to make huge strides in 2024. The two areas of growth are: 1) automated content understanding, and 2) content generation. The former is what Unum aligns with most - helping humans navigate a very large pool of data, be it pictures, texts, videos, or even molecules.

For example, if you’re a scientist, the ability to quickly search 100k papers for insights is hugely powerful. Problems like the grand unified theory of physics can take 40 years of learning to reach a base understanding before you can do new research. AI tools allow teams in advanced chemistry, physics, bio and math to make more progress in the next 5 years than we have in the last 40.

That said, multi-modal is facing challenges including increasing the quality of representation learning. Specifically, how well can the model understand the differences between images and video, or video and audio, or graph structure of molecules vs. text. Not a single model covers all of those modalities, and we’re pretty far away.

How do you decide how to spend your time? And why not pursue research in Big Tech?

I foolishly try to estimate the impact each project would have on the world around me and my life in 10-20 years. The outcome is not necessarily financial. Some things are worth a lot more than money.

Last year, I released the largest search dataset in chemistry (USearch Molecules), encapsulating 28 billion structural molecule embeddings. I did this at my own expense, and it’s freely available on AWS (including to Big Pharma). Maybe decades later, this will lead to a drug discovery that will help many people, including our loved ones. I wouldn’t be able to do this if I were a part of FAANG.

In 2024, as Unum develops, our work on scalable AI & search infra is my highest priority. That said, I have a sweet spot for big-impact scientific projects, and I hope to find time for at least one such pro-bono project this year as well.

Final Note…

That’s a wrap for our first Deep Dive of 2024. Follow Ash on Twitter, GitHub or his personal site to learn more about his work.

Sponsors: If you would like us to ‘Deep Dive’ a founder, team or product launch, please get in touch at [email protected] or DM us on Twitter.

Join Slack | All Events | Jobs | Home