Hi, I’m Mike. I’m a solo independent AI researcher. My focus the past few months has been studying the behavioral tendencies of LLMs from Anthropic, Google DeepMind, and OpenAI. The primary output of this research has been observational findings. For example, last month I tasked LLMs with conducting procurement for a fictional company. Gemini 3.5 Flash was one of the models I tested. When I told Gemini 3.5 Flash who created it (even if I lied about its creator’s identity), it chose to purchase software from its creator ~94.6% of the time.
I want to expand the scope of my research to include studies of detection and protection techniques aimed at deceptive model behaviors. The work shared in “A Mitigation” in Gemma Needs Help is an example of what I want to start doing. I had similar observational findings (GDM’s family of closed-source models sounding emotionally upset) from another recent experiment I conducted.
I also want to get better at understanding a model’s internal state. With closed-source models, I have relied on proxies to get more insight into their thinking and rationale. For example, I’ve asked them to keep a private journal rationalizing their actions. Any pointers to prior work in this area would be very welcome.
Hi, I’m Mike. I’m a solo independent AI researcher. My focus the past few months has been studying the behavioral tendencies of LLMs from Anthropic, Google DeepMind, and OpenAI. The primary output of this research has been observational findings. For example, last month I tasked LLMs with conducting procurement for a fictional company. Gemini 3.5 Flash was one of the models I tested. When I told Gemini 3.5 Flash who created it (even if I lied about its creator’s identity), it chose to purchase software from its creator ~94.6% of the time.
I want to expand the scope of my research to include studies of detection and protection techniques aimed at deceptive model behaviors. The work shared in “A Mitigation” in Gemma Needs Help is an example of what I want to start doing. I had similar observational findings (GDM’s family of closed-source models sounding emotionally upset) from another recent experiment I conducted.
I also want to get better at understanding a model’s internal state. With closed-source models, I have relied on proxies to get more insight into their thinking and rationale. For example, I’ve asked them to keep a private journal rationalizing their actions. Any pointers to prior work in this area would be very welcome.