AiThority Interview with Tal Shaked, Moloco’s Chief Machine Learning Fellow

Please tell us a little bit about your journey and what inspired you to start at Moloco. 

I’m very excited to join Moloco and realize our mission of empowering businesses to grow by making operational machine learning (ML) more accessible. Over the last couple years I’ve interacted with 100+ ML companies and Moloco stood out because we have one of the best approaches to leveraging advances in ML to create real value for hundreds of businesses across the world.  

I became interested in ML as soon as I started studying computer science in 1999. When I finished my undergraduate studies in 2002 from the University of Arizona, I told my advisor that I wanted to study AI because I was fascinated by combining the reasoning capabilities of humans with the compute power of machines. However, one year into my PhD program at the University of Washington, I gave up on AI research around planning.  

I had trouble getting excited about research problems in planning, such as having a Mars rover extract a sample of dirt, figuring out the best plan for 2 trucks to deliver 4 packages in 3 locations, or finding the optimal plan for elevators to pick up and drop off people. Instead, I was excited about how ML could learn from all the data in the world by extracting information from the web. I dropped out of the PhD program in 2004 to join Google because I saw a unique opportunity to work with some of the world’s best researchers and engineers to transform how people navigate the web.  

As it turned out, I ended up spending the next 3 years banging my head against the wall in web search, trying to convince people that ML for ranking was better than manually tuned functions.

In 2007 I co-founded Sibyl which was a project that initially focused on building the world’s most accurate, large-scale logistic regression system for modeling problems in ads. That effort got stuck in 2009, at which point Sibyl was at risk of getting shut down. We desperately looked for Sibyl customers across all of Google and finally found one in YouTube. A team of two engineers had built an innovative new product feature called YouTube Featured Videos, and they had no time to optimize the ranking, so they were happy to get the Sibyl team to help.  

I spent the next 9 months working very closely with a junior engineer from YouTube who was entrepreneurial, humble, curiousand very interested in learning as much as possible about ML. This led to the first Sibyl model launch in 2010, about 10 years after I first became interested in ML. After the success of that launch, the Sibyl team got a little more runway, which enabled Sibyl to grow and eventually become the most widely deployed ML system at Google in 2015.  

That junior engineer from YouTube moved to Android in 2010, and then left Google in 2013 to found a mobile advertising company. That engineer was Ikkjin Ahn, Moloco’s CEO and co-founder.  I’m honored to have the opportunity to work with him again.  

Tell us about the enterprise-level AI tech developments that you are currently focusing on and how it changed in the last 3 years? How did the pandemic change the landscape? 

While some people use the terms AI and ML interchangeably, I consider them to mean very different things. Here is an example definition from Columbia Engineering that captures the differences: “… artificial intelligence refers to the general ability of computers to emulate human thought and perform tasks in real-world environments, while machine learning refers to the technologies and algorithms that enable systems to identify patterns, make decisions, and improve themselves through experience and data”.  

I consider ML to be the future of software, and ML engineering (as a discipline) to be a superset of software engineering (more details here). With that perspective, the overall ML space is huge and growing faster than any person or company can keep up with. The ML technologies needed to create generative AI capabilities have been under development and already in use for at least a decade or more, depending on how one counts. It’s exciting to see more engagement with the general public through accessible APIs, which is creating a sense of urgency at the big tech companies, startups, and VCs.  

Personally, I’m still wondering what many companies will do to adapt to the different mindset and leadership that is needed to transform from a traditional company to an ML-first company like Moloco.

“ML-first” refers to thinking about products and software systems from the ground up with large data, scalable compute and ML in mind, which enables new product surfaces and approaches to existing problems. Many companies I speak with are still thinking about ML as something that gets bolted on to existing products and software, which can significantly limit what is possible. Similarly, leaders often treat ML as just more software, as opposed to new ways of building software and approaching problems. I’m focusing on how organizations and leaders are changing to adapt to what ML is enabling, and I think we still have a long way to go. For example, how many companies have a leader with the title of “Chief ML Officer”? (Not many.)

 What kind of IT infrastructure does a company need to fully benefit from AI and machine learning capabilities? What are the major programming languages that are driving these developments?

If we build off of the assumption that “ML is the future of software” and “ML engineering as a discipline is a superset of software engineering”, then companies need a scalable and flexible compute infrastructure to build everything on top of.  

If we make another assumption– which is that operational ML will create the most value in the long-term– then a great development environment and the appropriate talent that can build scalable software systems will be critical. Since the core ML code is a tiny fraction of the overall software systems and products that enable AI capabilities, I’d fall back to thinking about what languages people use for general software development. 

Tell us more about your recent recognition and how you plan to expand on your knowledge for AI industry.

This blog post does an excellent job summarizing some of my contributions. I’m proud of co-founding Sibyl in 2007, and then evolving it into TFX and “TFX is now [referring to 2020 when the article was published] the most widely used, general purpose E2E ML platform at Alphabet, including Google.” 

The industry is still searching for the best way to organize teams and projects to enable ML to transform their businesses. Together with other leaders, I helped create a structure that worked well at Google, which had roughly four pillars that were tightly integrated:

1.     ML infrastructure– Providing common tools and frameworks for many teams to leverage that enabled them to integrate ML into their software and products

2.     Product engagements–Working with many product teams to distill the common ML infrastructure needed such that all teams could more efficiently leverage ML

3.     Applied research–Doing research within the context of product engagements and common ML infrastructure to shorten the path from research to experimentation to production and ultimately business value

4.     Education–Providing educational materials so that everyone in the company could learn enough about ML to be more effective in their roles 

I plan to take what I learned at Google across research and product teams, and make ML more accessible to many companies via Moloco. This will center around operational ML and more end-to-end solutions that incorporate business logic and other domain-specific capabilities into products and services. 

What is the opportunity for organizations when it comes to utilizing AI and machine learning in their operations? How do you make a difference to the existing Big Data and IoT industries? 

There are many opportunities for companies to better leverage ML, and I suspect operational ML will be at the core of much of that value creation. The biggest tech companies like Google, Amazon and Meta have already been doing this for at least 10-20 years. They still have a long way to go, and many other companies are still near the beginning of this journey. 

The biggest way to make an impact is by being ML-first, which includes looking at software systems and products from the ground up, getting the right leaders in place, and rethinking how to organize teams to better enable a development environment that is conducive to creating ML-powered products. 

ChatGPT and Google’s Bard have taken the tech industry by storm. Could you let us know how generative AIs could transform low code DevOps? 

I think directions like GitHub copilot are great and could really accelerate software development with a human-in-the-loop approach. It would be even cooler if we could finetune such models on company specific code-bases. To balance our enthusiasm in this space, we should recall how self-driving cars have evolved over the last two decades and specifically that the last-mile problems with ML can be quite challenging. That being said, human-in-the-loop scenarios such as software developers assisted by code-gen tools seem easier.

Thank you, Tal! That was fun and we hope to see you back on AiThority.com soon.