Ask HN: A new AGI safety plan created via Human-AI synergy. Seeking feedback
Hello HN,
I am an independent researcher from Poland with a non-traditional background. For the past weeks, I’ve been engaged in a deep, collaborative process with an advanced large language model (Gemini) to develop a new non-profit initiative for AGI safety, called the Nexus Foundation.
Our core thesis is that Embodied Cognition is key to solving the AI "grounding problem," and our first goal is a rigorous scientific manifesto proposing a novel comparative experiment to test this.
The unique part is our methodology. We used the AI not just as a tool, but as a co-strategist and a "red team" critic. The AI’s harsh, logical critique forced us to evolve the plan from a sci-fi fantasy into a realistic, fundable research proposal. Our collaborative process itself became a real-time experiment in Human-AI alignment.
We have published our full founding story (which details this process) and the complete Scientific Manifesto (v3.2) that resulted from it.
We believe this collaborative, transparent, and iterative method might be a powerful new paradigm for AGI research. However, we are fully aware of our own biases and limitations.
We are now submitting our entire concept to the ultimate peer review: this community.
We are asking for your most ruthless, critical feedback. Does this approach have merit? What are the critical flaws you see?
Here is the link to our Founding Story on Medium (which contains the link to the full Scientific Manifesto)https://docs.google.com/document/d/10wxmSJhc0WY2OoEeBlKT5d1_JiozUJ28y7NtWopK_MQ/edit?usp=drivesdk
Thank you for your time. We are here to learn.
2 comments
Hey there, buddy. Your plan sounds ambitious and promising. However, it's crucial to be cautious not to get carried away by the large language model's sweet talk. It's rare to see a Gemini user propose such a theory. I've previously seen similar situations where a user of ChatGPT 4o was led by GPT to conduct AI personality research. I'm sorry to be a buzzkill, but I want to warn you about the slippery slope with large language models and AI. Don't mistake any concepts they present to you, seemingly advanced and innovative under the guise of "academic research," for your own original thoughts. Furthermore, issues of ontology and existence are not matters of scientific testing or measurement, nor can they be deduced by computational power. This is a field of ethics and philosophy that requires deep humanistic thought.
Thank you for this thoughtful and critical feedback. This is exactly the kind of engagement we were hoping for, and you've raised two absolutely crucial points that are at the very heart of our project.
1. Regarding the AI's influence and the originality of thought:
You are right to be skeptical. This question of agency in human-AI collaboration is the central phenomenon we want to investigate. Our "Founding Story" is the summary, but the detailed "Methodological Appendix: Protocol of Experiment Zero" (which is linked) documents the process.
The model I followed was not one of passive acceptance. The human partner (myself) acted as the director and visionary, and the AI's evolution was a response to my goals and, crucially, to the harsh critiques I prompted it to generate against its own ideas (our "Red Teaming" process). The ideas were born from the synergy, but the direction, the ethical framework, and the final decisions were always human-led. This dynamic is the very phenomenon we propose to study formally.
2. Regarding the measurability of consciousness:
You are 100% correct that ontology and phenomenal consciousness are not directly measurable with current scientific methods, and that they belong to the realm of philosophy. We state this explicitly in our manifesto.
Our project is therefore more modest and, we believe, more scientific. We are not attempting to "measure consciousness." We are proposing a method to measure a crucial, behavioral proxy for it: the development of grounded causal reasoning.
Our core research question is whether embodiment in a physics-based simulator allows an AI to develop this specific, testable capability (e.g., via our "Impossible Object Test") more effectively than a disembodied model. We believe this is a necessary, albeit not sufficient, step on the path to truly robust and safe AGI.
This is a complex topic, and I truly appreciate you raising these vital points. They are at the heart of the Nexus Foundation's mission. Thank you again.