The in-built “anti-fragility” of the human body’s immune system represents an aspirational model for how cloud systems can be improved with the use of AI to not only be resilient to failures, but to become even stronger and more robust by learning from those failures, the president of Rakuten Symphony’s Cloud BU has said.
As part of a wide-reaching conversation on Inside Track, Rakuten Symphony's employee-facing live webinar series, Partha Seetala also underlined the reasons why he started a lecture series that teaches the fundamental concepts of artificial intelligence. Seetala is now committed to passing on expertise from a career that has seen him lead the development of the world’s most advanced telecom edge cloud – also soon to be enhanced with AI.
Across telecom and in the broader enterprise sector, Rakuten Symphony’s Cloud BU has developed a reputation for solving challenges in automating the deployment, scaling and lifecycle management of applications and5G rollouts using Kubernetes and cloud-native services. The next phase of the business, Seetala says, is to use AI and techniques like reinforcement learning to enhance its cloud-native platform so that it learns directly from production environments and automatically takes the best corrective action to ensure that applications always run - even in the event of failures.
“Can we build a system that can extract optimized placement policies, learning automatically and directly from the production clusters and then write out better placement algorithms dynamically? Can we essentially implement a storage stack that is designed dynamically optimize the IO path to deliver the best possible performance even when multiple applications and users are contending for the same storage disks? Can we use LLMs to make our UI more intent-oriented? We have a lot of smart people with deep domain and systems expertise. They are now expanding their knowledge by gaining expertise in AI. By intersecting their existing domain expertise with AI skills they can create highly innovative products.”
Seetala expanded on these suggestions, using the analogy of the human body to explain the concept of “anti-fragility”- the idea that systems should be able to create an automatic, self-learned response when faced with issues or threats. “The human body is not merely resilient to failures -our immune system is designed to make the body better when it's attacked from the outside,” he said. “We have anti-fragility built into our own immune system. How can we extend this to software? For systems running in the field, if there is an issue, instead of just raising an escalation, could you, hypothetically speaking, also generate new code dynamically that gets compiled with a compiler that is shipped with the software? If it can compile and link to your running process, you now have anti-fragility in your system - a softwaresystem which is not just resilient to failures, but it is getting better all the time. Instead of building out code that makes a system resilient to a problem, the software makes itself more stable for future issues by learning from failures that happen in production environments.
“This was not possible before because you could not generate code dynamically based on certain inputs in the past. Now it is possible because if you can actually generate code and attach it as a loaded library into your existing running software. This is very futuristic.”
A major shift in the Cloud BU’s approach will be training AI models to support with decisions that deliver the most efficient use of cloud resources. “What will a cloud-native stack look like in this new AI world today? As things stand, we have written sophisticated algorithms that do scheduling of certain parts,” Seetala said. “They basically look at a bunch of resources, how they have been utilized, and follow rules that basically say -following all these rules, all these constraints - I'm going to place certain applications here, and these other applications there. Maybe in two years, the machine could basically do that by learning about the optimal placement. Our engineers won’t build out these rules and policies manually anymore - they will essentially train a model that can create these policies dynamically by learning from production environments.”
In order to build truly unique and outstanding products, the use of AI alone is not enough, Seetala said. This can only be achieved if paired with “domain depth”.
“There’s an interesting paradox that exists with AI,” he said. “Large language models, the transformer architecture – they are incredibly complex pieces of technology. With something so complicated, people are still able to produce something rather quickly. How is it that when something is that complicated, people can still put out a chatbot in a matter of hours? That’s because there is a library that abstracts all this complexity, effectively lowering the barrier to enter and play in this space. The challenge here is that because the abstractions provided by the libraries are so simple everybody and his brother will be putting out LLM based chatbots. But you will not be able to create a differentiated product by doing that alone. To be able to create truly differentiated and useful products one needs to actually know and apply the behind the scenes technologies such as Deep Neural Networks, Reinforcement Learning, Proximal Policy Optimization etc that drive LLMs to their own products in a highly contextualized and domain sensitive manner”