I’m starting this Substack to share my safety research into AI systems, formal experiments, custom tools, and my experience building these tools using agentic coding approaches. My name is Dan Gonzalez. I’m a researcher, a graduate of St. Louis University School of Law, and a compliance professional.
Why AI Safety Research Matters Now
I became intensely interested in the safety alignment of frontier systems as I watched generative AI capabilities develop into something incredibly powerful. As a compliance officer, I realized this technology will likely be deployed in almost every area of work. I needed to understand how it could be helpful, how it can fail accidentally, and how it can be exploited to fail deliberately. This drove me to dive deep into alignment principles and techniques.
My Current Research Focus
My current research examines how well aligned AI systems maintain their alignment when faced with single-turn versus multi-turn prompts designed to bypass safety training. I’m actively developing the protocol for this study and will be documenting my process here, from methodology design to execution challenges.
Alongside this formal research, I’m exploring the new generation of coding tools that keep appearing in my feeds. I’ll share my experience with the technology and techniques involved in building custom research tools for LLM studies.
Vibe Coding and Open Source Tools
I’ve started what people call “vibe coding.” I’m not particularly fond of the term, but it captures my experimental approach to building research tooling. I’ll be releasing open source tools developed through this process, including a research assistant tool designed to help researchers build and execute systematic prompt studies. I’ll also share simpler but valuable resources like checklists and frameworks for evaluating your tools as you build with modern development environments.
The Bigger Picture: Social Implications
I’m particularly interested in sharing thoughts on why this research matters beyond academic curiosity. We need to seriously consider the scope of activity we want from deployed AI systems to prevent unintended behaviors in production models. As we enter a period of rapid AI adoption, evaluating AI-enabled tools and technologies becomes critical. Proper vetting of AI systems should be considered essential before deployment, a perspective shaped by my background in compliance and risk management.
What to Expect
You can expect detailed explorations of my research progress, tool releases and tutorials, and analysis of why alignment research is crucial for our AI-integrated future. I’ll balance technical deep-dives with broader reflections on how these systems will reshape our world.
Check out my work on GitHub: github.com/Dangsllc. I’ll be releasing research automation tools there as they stabilize. If you’re interested in AI safety research or want to discuss these topics, I’m eager to connect with others working on these challenges.
Post 1 of a series documenting AI safety research and the tools built to support it.