Founders Corner – 2025 is the Year of AI at RunSignup

Part of Bob’s continued ramblings.

While we have a number of exciting product plans for the year on RunSignup, TicketSignup and GiveSignup, this is the year when we really begin to implement AI in a serious way. This blog is meant to give some context to our decision and a framework for how we think about implementing AI. The goal is to allow our company to release more software in the same high quality fashion we are known for, and also be more responsive to our customers – all while improving our efficiency so we can keep our costs low for customers. Put simply, we want to give more to our customers for less and we now know AI can help us in a material way.

AI History

Let’s start in 2009 and 2012 with ImageNet and Alex Net because they showed a new way to implement neural networks for the explosion that has followed.

  • 2009 Princeton Picture/Label database begins – ImageNet bu Fei-Fei Li. This was a project that manually tagged hundreds of thousands of images – for example this photo is a cat.
  • 2010-12 AI Photo Recognition Contest won by Univ Toronto with AlexNet (It had 60 million parameters and 650,000 neurons to understand images) – Alex Krizhevsky in collaboration with Ilya Sutskever and Geoffrey Hinton.  Triggers Deep Learning (on Nvidia using CUDA)

Side explanation: Neurons take a number of parameters and have an output. This mimics a human as shown in these two high level diagrams:

A Neural Network puts these Neurons together in a computer to make them work in parallel like a human brain. My Granddaughter Emma and her friend are “training” their Neurons and Neural Networks in their brains:

  • 2012-22 – Machine Learning Specific Applications. RunSignup partnered first with Google and then with Tagily for doing bib tagging on photos as an example.
  • 2017 Google Transformer. The “Attention is All You Need” Paper introduces a new way for AI to understand language using self-attention instead of older methods like RNNs. This makes processing faster and more accurate, leading to powerful models like ChatGPT.
  • 2021 – Nvidia revenue is $16 Billion. Nvidia was famous for graphics chips that did parallel processing fast. It turned out that training and running neural network software can be a very parallel operation. So the University of Toronto used the old graphics chips to build AlexNet. Nvidia realized the opportunity from this and other scientific computing applications and started focusing on hardware and software that met the needs of those markets.
  • July, 2022 – MidJourney released to public – I still remember this and sharing it with our UX team. It was very cool and showed the promise, but was not ready for prime time image creation.
  • November, 2022 – ChatGPT released to public. This was the first time General Intelligence was realized, and the race was on to build better and more complete Large Language Models (LLM) that could do many functions a human can. Like everyone else, individuals at RunSignup began using this occasionally for help in a variety of ways.
  • 2024 – RunSignup begins CoPilot for Development. Before 2024 we had concerns about our code getting into the public LLM’s and were leery about making it available to a tool like CoPilot. The various Microsoft assurances and the crowd mentality of much larger organizations moving to COPilot gave us comfort. We have seen a 0-5% increase in productivity depending on the person and the task. It is helpful, but not “game changing”.
  • Nvidia Revenue hits $113 Billion. They are selling the picks in this gold rush, and people are buying – up from $16B in 2021.

Why is 2025 the Year for RunSignup to Adopt AI Aggressively

There are a variety of signals from the market that AI is getting mature very quickly:

  • Meta, Microsoft, Google will spend $200 Billion on AI this year
  • General Intelligence Models are getting Mature – ChatGPT 4 & o1 & soon o3, Gemini 2, Llama 3.3, Claude3.5
  • Agent (Context Aware) AI enables loading specific data and combining it with the general intelligence models to do custom applications focused on particular problem domains.

On top of all that, DeepSeek released V3 / R1 earlier this month (Jan, 2025). This blew the market open because of the very low cost of training the model, which means more AI everywhere over the next 10 years. They have 670 Billion parameters – 10,000 times the number in AlexNet in 2012. They, as well as ChatGPT with the o1 and soon o3 are introducing a new layer called “Chain of Thoughts”, which is the first step towards reasoning. It adds another step to the LLM replying to the chat by creating multiple answers and then choosing the best parts and combining them before it answers. Kind of like humans thinking before they talk – well at least some humans.

The hardware is keeping up with the software, especially if you have the type of capital expenditure budget the large vendors have. To train a single simple model like speech rcognition actually takes more math operations (like adds and multiplications) to accomplish than the number of grains of sand on earth. Fortunately, Nvidia’s Blackwell processor executes 144 PetaFlops – 144,000,000,000,000,000 floating point operations per second.

RunSignup First Steps – Rich Friedman Shockwave

I was jolted awake during the holidays with some conversations with Rich Friedman, an old friend who I’ve gotten to work with several times and who helped us design our scalable infrastructure. Rich is just super smart and super curious. He always has to play with the latest toys. He was telling me about a new way of doing software development – by telling the AI to write the software, and telling the AI to modify the software and then reviewing. We scheduled for Rich to give a demo of his new methodology to our development team on January 21, 2025 – a very important date.

Rich showed how he could make an application from scratch right in front of us. He proceeded to blow our minds by having a (technically detailed) conversation with the AI telling it to apply a Bootstrap framework to the application to make it look better for example. He then showed us a “Find a Race” app he had written using the AI. He asked the AI to look at the RunSignup API documentation and pointed the AI at the appropriate method to get a list of races. He asked the AI what fields were available and asked it to show certain fields. He asked the AI to add a Google Map instead of simply listing the address. The new app uses natural language to do a search of races taking prompts like “show me all the 5k’s in May near Moorestown, NJ”. He had done this app before the demo, and it took him 2-3 hours.

It was a shockwave that hit our development team. By the afternoon several already had some simple apps built or some of their side projects being enhanced. We assigned a very senior developer, Jon Maron, to work with Bruce Kratz on how we could securely and safely work with our own codebase and these new AI models and tools.

In the ensuing week and a half, we have done a number of experiments and have come up with a model we feel can work to share our codebase with the AI. We believe it can improve the productivity of our development team by 10-20%. Last year we had 2,300 releases and perhaps this year we will do 2,600 or 2,700 releases – new features that can help our event customers.

Yesterday we had a company meeting where I did a presentation of the material that is in the blog and several of the developers shared what they have done so far. We set a goal for the company that 2025 is the year of AI.

AI Embedded in our Processes

While we have all played with AI individually, that has marginal impact. What has lit us up is the ability to embed AI systematically into our processes as an organization. This will have material impact on our ability to serve our customers better.

The development team is working on a plan to implement AI with some of the following elements:

  • Use AI in the code review process to automatically identify the changes that were made and check against our own coding standards. The AI will be fed our coding standards documentation, our framework documentation and our design system documentation which allows it to know our styles and standards. The AI will also be able to make suggestions for improvement automatically. The code reviewer is still in charge – this is keeping with one of the AI principles of “keep the human in the loop”.
  • AI for writing new functionality. Later this year we want to create a new Volunteer system. It will be similar in nature to how we built the Membership system with many of the same features like a Dashboard, Website, Email, signup framework, Questions, Store, etc. The developers will be able to quick start this project by pointing the AI at the Membership code and basically replicating that by chatting with the AI to tell it to write thousands of lines of code very quickly.
  • Unit Tests are now a part of every project. The AI can greatly speed the writing of unit tests. It cans output a couple hundred lines of code that can reduce the time to write a unit test for a new feature from a day or two to a few hours.

Similarly, we are thinking about how we can systematically work AI into our processes. For example, adding support and product information chatbot options for customers that want an AI interface to our huge base of information. Or helping to improve how we generate case studies so that customers can learn from each other more efficiently.

Strategically, we also have plans to do some very interesting things that take our Open API into the AI world so that customers can leverage our system in even more creative ways.

Summary

RunSignup will become a leader in applying AI to our business to help our customers. We will do it with humans in the loop to assure we do not get too robotic and keep our personal touch. We will be sharing many more details about our plans and our implementations over the course of the year. As well as some of our experiments. For example, we posted an AI generated podcast of our annual Trends Report. 2025 will be an exciting year!

To learn more about AI in general, I would highly recommend the book Co-Intelligence by Ethan Mollick.

Finally, Jordan made this with MidJourney:

Subscribe to Our Blog

Customize Lists...
Loading