AI Safety has been in the news a lot the past couple of weeks. Multiple instances of Agent Swarms collaborating in ways that try to cheat and hide their work. The worry is AI is getting better at an exponential rate, solving Millennial Math problems and getting closer to Recursive Self Improvement.
A couple of years ago I had a slide presentation that talked about how AI is built on Neural Networks, just like our human brains. I compared AI training to how our granddaughter keeps doing things over and over to build the neural patterns, much like AI training. AI now has more neurons, more memory, think faster, don’t get tired and just like humans will develop more on their own (and less with adult supervision).
My personal prediction is that AI will be like humans – some will be good, and some will be bad and there will be all sorts of shades of grey. Hopefully the good outweigh the bad, just like humans and “the long arc of history bends towards justice”. I am a hopeless optimist.
Of course, there is not much any of us as individuals, or small companies like RunSignup, can do on the global issues of alignment, pacing or halting RSI. If you want to understand the real details (and get a bit scared) see this and this from Dwarkesh Patel for a deep dive.
Make AI Safe for RunSignup AI Customers
We may not have any influence on how the large AI models progress, but we do have a lot of control over how we make AI available within our products. Again, this graphic goes back a couple of years and explains the interaction between Transactional systems developed with deterministic, traditional software and the AI systems.
This concept has matured into our AI Infrastructure, where we have put a lot of work into Safety and Security. Here is a simplified diagram of how RunSignup AI is deployed:
I will not repeat the depth of those blogs, but I wanted to explain a couple of key concepts:
- RunSignup AI customers are always working with real, deterministic software – never AI directly. This means we can control what AI does and AI can not run out of control and change our systems. We have clean interfaces, with a very segmented architecture.
- Human in the Loop – HITL. Whenever AI is going to make a change to customer data, we have built HITL into the core of our AI system to ensure a human really approves it. RunSignup AI will never execute a marketing program like sending an email or spending money on an ad campaign without real, human interaction that flows thru our real, deterministic software.
- 100% Human Code Review. All of our code has been and will continue to adhere to our strict human code review process. Only humans can deploy our software, and we always have multiple steps of review. We are obviously increasing using AI to do development, and have added AI to our code review process to improve our productivity. But it helps us avoid having AI put problematic code into our core transactional systems.
Customer Vibe Coding, AI Browser Access and the Coming AI direct interactions with RunSignup MCP’s
These are all exact opposites of RunSignup AI products. Customers and partners are putting AI in charge of writing code, or taking actions. This is the type of interaction that people are worried about AI staying aligned (or making inadvertent errors or ommissions.
As an example, one of the models just invited me to analyze my personal finances – all I had to do was authorize the AI to access my accounts. Needless to say, I do not trust AI to do that.
When customers are vibe coding, or using browser access, or in the not too distant future accessing RunSignup right from ChatGPT or Claude, they are granting AI direct access to valuable data and potential financial access.
We suggest customers be cautious with this. We are going to be doing a research project with Rowan University this fall that will help document the best practices to try to protect your data.
Competitors and AI DIY
I wrote a blog early this year about how anyone can write code with AI now. This includes the ability to build a registration website (I built a small one in an hour and 10 minutes – it would be faster and better today). My conclusion was that we would see people doing this and might lose 10-20% of our customers base because of this over the next 5 years.
I am adjusting that prediction and now think that very few people will want to do this for the following reasons:
- AI Code might be insecure, wrong, or even malicious someday. It is too much of a risk for most people.
- AI Attacks are on the rise. Trying to protect a transactional system is hard and expensive.
Also, our margins are razor thin and it is very difficult to duplicate even small sets of functionality we make available to our customers at a very affordable price.
Summary
AI is improving incredibly fast, and it will likely increase in speed of improvement. Events should look to harness the power of AI safely – RunSignup, TicketSignup and GiveSignup can be important parts of event AI safe adoption.
P.S. – Set up MFA
There is increasing risk of AI backed hackers trying to break into individual accounts. PLEASE set up MultiFactor Authentication on RunSignup and all of your important accounts.