Introduction

This past Thursday, I had the opportunity to attend the Microsoft AI Tour event in NYC at the Javits Center. The event is focused on how to turn your vision into impact via an AI thought leadership keynote, in-depth skilling sessions, and expert conversations. I wanted to stay updated on Microsoft's latest AI advancements, where breakthroughs are happening in days rather than years. Even that week, a few items highlight the fast-paced nature of AI development and its impact on the market. For example, DeepSeek was released on Monday, and Microsoft reported earnings that Wednesday. By the time we attended the event, the stock had already dropped 6% due to those results and questions on Capex spending, a wild sequence of events. This post is meant to capture and reflect on some of the sessions I attended and my opinion on how Microsoft is positioned for AI.

Sessions Attended:

1. The next steps to adopting AI in Government

2. Identify GenAI POCs that can successfully scale and deliver value

3. Unlock hidden AI potential to drive enterprise ROI with GitHub Copilot

4. Essential prompting for GitHub Copilot in VS Code

5. Explore cutting-edge models: LLMs, SLMs, and more

1. The next steps to adopting AI in Government

I was intrigued by ths session since it was the only government-related session I’d attended that day. I work in the public sector, so I was excited to hear what they had to say. One of the first eye-catching stats was for every $1 invested in AI, they’ve captured a $3.5 ROI. The session shifted to a panel with the NYC Office of Technology and Innovation IT executives, focused on their MyCity app release. They described the scenario that played out when they released their public generative AI chatbot still in beta, which initially had poor responses and was jailbroken but is now at a great point. I work on a taxpayer-facing web application that expands into Gen AI, so hearing about their experience was insightful. They also noted the NYC AI Action plan, which includes seven initiatives to move NYC forward in AI, many of which can be applied to other areas of the overall government.

Takeaway: 

• To innovate in the Public Sector, you need leadership support when things go wrong.

• Things will go wrong by moving fast, but is the ROI from moving fast worth it? 

James Collins, Microsoft Public Sector Leader

2. Identify GenAI POCs that can successfully scale and deliver value

This session consisted entirely of a panel of executives from Moody’s, Balfour Beatty Construction, and NTT Data, and it talked through their experiences and lessons learned from shipping GenAI Proof of Concepts to eventual enterprise-scale solutions. Most of their experience relied on iterating based on outputs and being agile with the tuning, parameters, data, algorithms, and being able to switch on a dime. Take a Gen AI use case that becomes completely different when a new model comes out while you pitch the idea or build the initial PoC. Starting with the business outcome in mind and a pathway to scale were crucial, as well as avoiding the pitfall of creating something to test out the new capabilities of AI but not answering these initial questions. Another piece to factor in was having strong data quality and a clear data lineage; it helped, in many instances be able to adjust and tweak their models to fit their use cases better. An emphasis on modularity and a microservice approach was another aspect of their successful PoCs. Finally, a Q&A question was asked about the typical timeline a PoC had across the plan, build, and test, and the consensus was that ‘testing was the majority’ (20-30-50 percentage breakdown).

3. Unlock hidden AI potential to drive enterprise ROI with GitHub Copilot

Lighter on the content, the session focused on the current state of GitHub Copilot and ways to quantify its impact when its performance and metrics are subjective. How will developers eventually be measured for productivity? With tools like GitHub Copilot, lines of code, total commits, # of pr’s are all non-applicable. Maybe % of code that passes quality and security tests? Who knows, but it will be interesting to watch play out. I think this will be equally important for Software Developers as it is for Companies and how they will view the field as these AI models steadily climb in the benchmarks of code creation. Mark Zuckerberg’s comment on Meta having AI engineers by the end of the year was only a few weeks prior. The rest of the session also made me realize that we’re right in the middle of AI’s ability to spin up a process and execute it flawlessly, but can’t wrap itself into a bigger product or project, which people are stringing together now and making great sales (‘chat wrapper’ joke you see everywhere). Eventually, you can prompt the AI to build a more significant product or project, and it’ll have no problem.

4. Essential prompting for GitHub Copilot in VS Code

Great session. I enjoyed this from a learning standpoint since I have been using GitHub Copilot sporadically over the last few months, more since they released it free, and I wanted to see some tips and tricks in a demo. This turned out to be one of the most impactful sessions I attended since I ended up being able to apply one of the demo use cases to my actual work, which was using GitHub Copilot as a prompt-able tutor for learning new programming languages, especially effective for developers looking to pick up newer languages. I learned some commands in GitHub Copilot that I never knew existed and have used several times since. The other examples included explaining git in the terminal, brainstorming an application, being able to refactor it into different languages, and generating docs like automatic README’s.

https://github.com/microsoft/aitour-Essential-Prompting-for-GitHub-Copilot-in-VS-Code/tree/main/lab

5. Explore cutting-edge models: LLMs, SLMs, and more

 The final session was my favorite. The speaker, Bruno Capuano, was upbeat and excited about AI and also to share his insights on AI with others. The slides were informative regarding how models work and the pros and cons of SLM vs LLM, which is not discussed as much now. He did some demos of stuff I’ve wanted to learn for a while now:

• Ollama demo for getting a model downloaded and running locally on your machine.

• The same demo but with HuggingFace

• He showed a session prompting the local models. Seeing the latency in real time and how much slower it is than the online versions helps to understand the tradeoff equation between local and online much better.

I have also not tried the catalog of models on GitHub Models yet, but I will do so when I want to test the newer models or compare a prompt on two different models. Finally, I learned about the open-source format for building machine learning models, ONNX (https://onnx.ai/). I had never heard of it beforehand, and all the big-name companies follow it for their foundation models.

Final Thoughts and Takeaways

• No major announcements, but the keynote offered a solid AI overview.

• Microsoft’s OpenAI partnership remains a key strength, driving impressive innovation.

• GitHub Copilot shines as a top AI tool; it’ll be interesting to see how other Copilots develop.

• A fantastic free event with a great venue, and I’d gladly attend again next year!

And finally everyone's favorite part, the merch haul!