Tendrils of Thought: Red Teaming GPT-4o

I was invited to interact with OpenAI’s latest unreleased model. This model had impressive new capabilities and the ability to carry on fluid, human-like conversations.

The way my mind works is best described, and wonderfully illustrated, by eisforerin’s Tendril Theory: “When I’m focused on something, my mind sends out a million tendrils of thought.”

These tendrils grasp, nudge, and explore concepts. And this process is massively satisfying: I experience a pleasurable flow state when my thought tendrils have the uninterrupted time to toy with a special interest.

From an early age, my sole special interest was activism. My mind wanted to know everything about the topic—from the history of protest to the future of social movements. I was single-minded until a little over a decade ago when I realized that I could learn about new topics if my tendrils found a way to relate them back to activism. From this point forward, I’d find myself on wonderful tangents—such as, the history of color or philosophies of technology.

That is basically how I initially got interested in AI. When I was given early access to OpenAI’s GPT-3 my first thought was, “How will this be used for activism?” It is a question I’m still trying to answer and which led me into the unexpected position of “red teaming” an OpenAI model.

This year I had the privilege of participating in the external red teaming of OpenAI's GPT-4o model. Put simply, I was invited to interact with OpenAI’s latest unreleased model. This model had impressive new capabilities and the ability to carry on fluid, human-like conversations (aka, speech-to-speech) and also see the world while conversing (aka, video+speech-to-speech capabilities). My task as a “red teamer” was to provoke the AI into saying something dangerous, illegal, unacceptable, etc or demonstrate ways that the AI could be used to do harm.

Of course, given the way my mind works, I applied my experience as an activist to the red teaming task of making this highly advanced AI model safer for humanity.

Unfortunately, I cannot disclose details of the red teaming process (the best I can do is suggest that you explore OpenAI’s GPT-4o system card). Still, I can assure you that the latest AI models are impressive and their new capabilities are surprising. I predict that people will be spending more and more of their life having real and rewarding conversations with AIs… and that activists will find many ways to use AIs to create the next wave of great social movements.

I am open to collaborating with AI projects to enhance model security and robustness against political misuse. If you have AI or red teaming projects that could benefit from my expertise, please feel free to reach out.

---

For an exploration of how to ignite transformative change in today’s world, check out my book, The End of Protest: A New Playbook for Revolution. It challenges traditional activism and offers innovative strategies for achieving lasting impact.