Tyler Williams
Founder/Researcher @ VANTA Research
Hello! I am the sole founder of VANTA Research - a bootstrapped AI research project focused building safe and resilient AI models optimized for human-AI collaboration. All of our published work has been made open source thus far, and in roughly 2 months has garnered 60k+ downloads on Hugging Face and Ollama across original model families.
I have several goals with VANTA Research, but among the first major stops is building a large (400B+) open source foundational model from scratch. I love learning, asking hard questions, and a good mystery.
DMs are always open.
Great write up! Thanks for sharing. I've been thinking a lot about the double-edged sword of open source models. You make a great point with model providers being able to monitor usage of their products over their API.
A safety layer with open source models does not exist. Once those weights are downloaded, they are out there forever - and people are able to customize them as they see fit. No change in law or policy can retroactively remove those capable LLMs from existence.
Anthropic was fortunate to catch this, but it certainly makes you wonder what's going on undetected from locally hosted models.
It's cool to see a role like this open up. I'm curious to see how SLT plays out in practice, especially at scale. I've seen some pretty dramatic shifts in generalization between different versions of the same language model, even just from one quantization to another. Definitely feels like important territory to explore.
Great list! I've actually been working on something that aligns closely with #3: I've been independently testing LLMs (including Gemini, Grok, DeepSeek etc.) for unexpected behavior under recursive prompt stress. I've been documenting my tests in red-team style forensic breakdowns that show when and how models deviate or degrade through persistent pressure.
The goal of this is absolutely to see and evaluate how agents behave in the wild. I believe that this is a critical safety test that cannot be missed.
I'd be curious to connect with others that are interested in research/testing from this angle.
What a great post! Thanks for taking the time to share it, Constance! The section on cultivating a HEALTHY whisper network was certainly appreciated. I don't think it's something that's talked about enough.