Machine learning in ads: Insights from a chess grandmaster

Image courtesy of Damiano Lingauri.

Even with today’s technological advancements, the quest for the perfect advertising system remains elusive. The intersection of machine learning (ML) and advertising continues to evolve, but a flawless solution for matching the right ads with the right audience has yet to be achieved.

Moloco, a company offering ML-based mobile advertising solutions, is one such organisation working to improve this complex process. Specialising in performance advertising, Moloco seeks to equip businesses with more effective tools for reaching their target audiences.

In this exclusive interview, Frontier Enterprise talks to Tal Shaked, Chief Machine Learning Officer at Moloco. Shaked shares insights into the role of ML in advertising, drawing from his unique background as a chess grandmaster.

Chess and ML both involve strategic decision-making based on patterns and probabilities. How has your background as a chess grandmaster influenced your approach to solving the ‘matching problem’ in advertising through ML?

Each of us is uniquely wired from birth, a trait that extends to computers and algorithms as well. The diversity in our wiring and learning abilities is broader than what many of us may realise, making it difficult to fully appreciate how each person or computer learns and perceives the world. My personal journey with chess began at the age of seven, and due to my inherent wiring, I won the primary state championship within months, and the national championship a year later. Although I learned chess more quickly than others, the path to becoming a grandmaster was long and arduous. I had to engage in deliberate practice, focusing on my unique mistakes to get feedback and improve over time.

My experience in learning chess parallels how computers use ML to make better decisions over time. Like humans, computers have diverse learning algorithms which result in varied capabilities. Similar to my experiences in chess, ML algorithms learn from their unique errors, continuously optimising their decisions through exposure to new data.

Playing chess often feels like a mix of art and science. While it’s solvable with enough computation, certain moves can seem very creative and beautiful to humans. Likewise, ML involves complex scientific algorithms, but applying them effectively to solve interesting business challenges, such as in advertising, takes a lot of creativity.

Cloud computing and open source are said to level the playing field in the tech industry. How have they impacted the development and deployment of ML models in advertising, and what future opportunities do they offer?

Thanks to open-source software and cloud computing, many companies can now build operational ML systems, which were once exclusively available to Big Tech companies. As these technologies continue to improve, becoming more cost-effective and flexible, the playing field will become even more level for tech companies of all sizes.

However, developing operational ML systems tailored for specific domains like performance advertising still takes a lot of work. This complexity presents many opportunities for companies like Moloco to innovate and develop new, ML-driven products for a variety of businesses.

Talent is another crucial factor. Many individuals who were instrumental in building the first operational ML systems at Big Tech companies are leaving to join startups or launch their own companies. This trend is likely to accelerate innovation, especially in sectors where larger companies may be less agile.

Why is it important to build ML models with specific purposes and use cases from the outset? Can you share some examples of how this principle has been applied at Moloco, and its impact on your advertising strategies?

Tal Shaked, Chief Machine Learning Officer at Moloco. Image courtesy of Moloco.

There are many nuances to operationalising ML models. ML is not a silver bullet that can automatically create amazing product experiences on its own. Similarly, it takes much more than generic software to build great products, because building software solutions and using ML effectively requires a lot of effort and specialisation. Companies often have to build models from the ground up, tailored for specific use cases.

Another consideration is that many products powered by ML need to continuously learn and adapt to changes in user behaviour and the rest of the world. This often requires customisation to incorporate new data and real-time signals that are specific to the problem domain. 

In a practical application of this principle, Moloco collaborated with Singapore-based live-streaming application Bigo Live. The objective was to identify active, valuable users to aid Bigo Live’s global expansion. Utilising global inventories, the partnership facilitated Bigo Live’s growth into over 20 different geographies. Through ongoing campaign monitoring, diversification, and optimisation, the results surpassed Bigo Live’s initial goals by an average of 10-15%, although this varied by geography.

ML is said to be a superset of the future of software. How do you see its role evolving in the next decade, and what steps should be taken to nurture the next generation of ML engineers, especially in the context of the existing talent gap?

Here is my current perspective on how the industry should approach ML:

  1. ML is the future of software.
  2. ML engineering as a discipline, is a superset of software engineering.

ML will continue to evolve and play an increasingly larger role across all types of software for the foreseeable future. Furthermore, building ML-powered products will require expertise that goes well beyond traditional software engineering. This includes areas such as data science and analytics, and new ways of thinking about systems that interact with and learn from people and the rest of the world.

Take web search as an example. Two decades ago, I served as a teaching assistant for a class that taught how to build inverted indices for document retrieval. Today, we are talking about technologies like ChatGPT and vector databases, which are transforming how we think about information retrieval. These technologies are built on entirely new software stacks. As we continue to innovate and discover new ways to build ML-powered products, we will continue to rethink our end-to-end software and development processes to better enable those ML capabilities.

For those wanting to work in ML, my advice would be that nothing beats experience. While it’s important to learn ML fundamentals through books and courses, there is currently no standardised curriculum for designing and building ML-powered products. Many companies are in the experimental phase, learning how to do this as they go along. Real-world applications of ML are constantly evolving, so learning on the job is the most effective approach at this point.

The other piece of advice I’ll give is to not lose focus on the actual problem when going deep into ML technologies. Although ML can often be the best way to solve a specific problem, it is not a magical technology that solves all problems. This is especially true when the problem itself lacks a clear objective. In such cases, more problem exploration and experimentation are often needed before resorting to an ML solution.

Chess has often served as a benchmark in the development of AI. As a chess grandmaster, how do you view the relationship between chess and AI? What insights can be gained from AI systems that excel at chess, and how can these be applied to ML in other fields, such as advertising?

I’ve always been fascinated with chess-playing programs, and recall playing chess against computers in the 1980s shortly after I learned how to play. At that time, the chess programs were quite predictable, and I took pride in developing strategies to consistently outplay and win against computers.

It was exciting to watch the evolution of chess computers, culminating in Deep Blue’s defeat of the world champion, Garry Kasparov, in a 1997 match. Although this particular chess match had essentially no connection to ML, Deep Blue’s strength was primarily from a blend of brute force calculations and excellent heuristics, developed by domain experts. This was an impressive illustration of “machine intelligence” at that time. If we define “artificial intelligence” as the process of building computers and machines capable of reasoning, learning, and acting in a manner that would typically necessitate human intelligence, then it’s fair to categorise that chess match as a classic AI moment. However, it’s crucial to note that Deep Blue was still a long way from mastering chess, not to mention emulating the chess-playing capabilities of a human.

About 20 years after the Deep Blue match with Kasparov, AlphaZero (a computer program developed by DeepMind) employed reinforcement learning to teach itself chess. By playing games against itself, it rose to become one of the strongest chess-playing programs in the world. This achievement was a significant milestone in ML, creating a new type of machine intelligence and marking another advance in AI. Looking back, it appears that the development of an algorithm capable of superhuman chess performance was only a matter of time, especially when paired with a powerful-enough computer.

Leveraging increasingly sophisticated ML algorithms on powerful computers will naturally lead to new and more powerful forms of machine intelligence, which will power a broad array of future AI capabilities. For the past 20 years, operational ML has been enabling machine intelligence within the realm of performance advertising, but we are nowhere close to an ads system that behaves perfectly. Similar to chess-playing computers, I expect continued innovation in using ML to create new forms of machine intelligence that will be integral to high-value products and capabilities.