In a hypothetical co-amp system, you can ask it to write your program. But in addition to this, you get a trace of what was thought and why. This is not the same as a narrative story; GPT-4 does not have access to its own internal cognition. It itself doesn't know why it makes the decisions it makes. Is it possible to do science this way? Yes, I think so. If you were like a hive mind species or like an alien species where every time you reproduce the children inherit all the memories then yeah, I think you could probably do crazy leaps of logic that are completely make no sense to human and cannot be taught.
Read the full transcript here.
Does AI pose a near-term existential risk? Why might existential risks from AI manifest sooner rather than later? Can't we just turn off any AI that gets out of control? Exactly how much do we understand about what's going on inside neural networks? What is AutoGPT? How feasible is it to build an AI system that's exactly as intelligent as a human but no smarter? What is the "CoEm" AI safety proposal? What steps can the average person take to help mitigate risks from AI?
Connor Leahy is CEO and co-founder of Conjecture, an AI alignment company focused on making AI systems boundable and corrigible. Connor founded and led EleutherAI, the largest online community dedicated to LLMs, which acted as a gateway for people interested in ML to upskill and learn about alignment. With capabilities increasing at breakneck speed, and our ability to control AI systems lagging far behind, Connor moved on from the volunteer, open-source Eleuther model to a full-time, closed-source model working to solve alignment via Conjecture.
Staff
Music
Affiliates