Anusha, a Member of Technical Staff at Microsoft AI, is basically an AI sculptor, taking those raw “foundation models” and shaping them into something truly useful. She’s the one fine-tuning Copilot’s personality and skills, from data collection to algorithm tweaking. Her journey into AI started with a Siri internship (way back in 2013!), and she’s been hooked ever since. Now, she’s sharing insights into her work, the future of AI, and some advice for aspiring AI scientists.
How would you describe your job and what you do? I’m a Member of Technical Staff, AI (MTS-AI), and I work on what’s often called “post-training” of foundation models, for consumer Copilot. The best metaphor for it is sculpting: foundation models (LLMs) are like clay, and the possibilities for what you can create with them are truly endless. The post-training process I work on involves sculpting this clay and imbibing it with personality and capabilities to turn it into the Copilot that users experience. In the day-to-day, this involves a variety of steps: collecting data to demonstrate the behavior we’d like Copilot to have; creating evaluation benchmarks that measure model performance for the tasks we’d like Copilot to be good at; improving our post-training algorithms and recipe; and working with other engineering teams to ship our latest and greatest models to all Copilot users.
How did you first become interested in generative AI and machine learning and what motivated you to pursue it as a career? My interest in generative AI and machine learning began with an internship at Siri in 2013 while I was in college. It was my first industry experience with machine learning and natural language processing (NLP). Seeing how these algorithms were applied to real conversational AI products that deliver value to users was eye-opening. This experience inspired me to take more NLP and AI courses and pursue research opportunities as part of my coursework.
What educational background or training do you have that has helped shape your skills? I hold a Bachelor’s and a Master’s degree in Computer Science, with a specialization in AI during my Master’s. I was fortunate to work on several research projects in conversational AI and had the opportunity to publish papers at and attend major NLP conferences. Additionally, I’ve worked on building AI products at Meta, Microsoft, and Inflection AI, gaining valuable knowledge from each experience.
“One of the most exciting things for me has been that we’re developing technology simultaneously with users discovering its best use cases.”
Anusha Balakrishnan, Scientist at Microsoft AI
What are you currently working on? I work on post training for Microsoft Copilot, and I’m currently focused on improving Copilot’s overall personality and its ability to retrieve and use real-time information from Bing.
Can you share an overview of your role as a Member of Technical Staff, AI and how it specifically relates to Microsoft AI? As an MTS-AI working on post training, my primary goal is to push the performance of our current models by driving innovations on all aspects of the model post training process. This involves, in no particular order: running a lot of experiments to test new ideas; keeping up with new papers in the LLM space; working with feedback from our human annotators and users to understand how to improve our models; and working closely with product and platform teams to ensure the models we train are shipped fast and responsibly.
What role do user feedback and usability play in your process? How do you incorporate feedback to improve your work? User feedback is the strongest signal we can get to guide how we need to improve our models! It comes to us in many forms, from user research studies to app store reviews to thumbs down feedback in the Copilot app. One of the biggest areas we’re investing in right now is developing methods for improving our models based on all these different forms of user feedback. One very interesting aspect of this is how exactly we improve Copilot based on the feedback; we’re currently thinking both about how to improve Copilot for all users based on user feedback, as well as how to improve each user’s personal experience based on the feedback they provide.
“Hands-on experience has always been the best teacher.”
With the introduction of AI, what changes or shifts do you anticipate within the industry? The generative AI space has been moving rapidly, and the most significant shift we’re currently observing is towards agentive models that can perform actions such as searching the web or running code on your behalf, and reasoning or “thinking” models like OpenAI’s o1, which dedicate time to reasoning before answering any question. I think the intersection of these (models which spend time doing research using various tools before providing an answer) is going to fundamentally alter how people do their work in every industry, all the way from doing software engineering at Microsoft to conducting drug discovery research.
What excites you as an MTS-AI with the integration of AI? One of the most exciting things for me has been that we’re developing technology simultaneously with users discovering its best use cases, and the speed of innovation has been incredible to see. For my own work, it’s been fascinating to see how each model innovation reshapes the roadmap for what we’re building or how we’d go about building it; features or capabilities that seemed very complex just a year ago are now much easier to implement, and that means we need to adapt our approaches and expectations in real-time.
How do you stay updated on current tech trends and emerging practice? AI X (formerly Twitter) has been one of the best sources for me. Other than that, I keep up with interesting papers from the major labs/companies and those that get shared in my extended network (including the reading groups we have here at Microsoft AI!)
Looking back at your journey, what advice would you give to aspiring AI scientists who are just starting out on their own path? I would say the most important skill to hone is the ability to acquire new knowledge and quickly learn from new trends. This means taking courses and reading blog posts, technical reports, etc., to familiarize yourself with the terminology and the state of the art in AI models. For me personally, hands-on experience has always been the best teacher, and I’ve found that even half an hour spent on playing with the API for a new model or product can teach you a lot.
