The quote being discussed is about nuclear risk.
@jeremyphoward
-
Nuclear Experts and AI Risk Assessment Expertise
By
–
Nonetheless, even under your proposal, the best person to understand analogies to nuclear risk are nuclear experts. And that is who I quoted. You're a marvelous DL researcher and I've learned so much from you — but perhaps there are also people you can learn from?
-
Academic Research on AI Risk Management and Safety
By
–
Actually many academics have spent their lives dedicated to exactly this question. Whether or not it's "normal risk management" it's nonetheless a well trodden path. You are ignoring years of research and study.
-
Model Transparency and Training Data Release Critical for AI Accountability
By
–
They really need to release their model and full transcripts to see what happened, but I'd be extremely shocked if the model described anything not already in the training data.
-
Model Fine-tuning and Data Memorization in AI Systems
By
–
I'm not wrong on methodology – but perhaps I did overstate this. The model was fine-tuned on papers describing exactly the thing they then asked questions about. In my experience models remember fine-tuning data precisely, and regurgitate it on demand.
-
Sparse Packed FP64 Models: Efficiency and Safety Trade-offs
By
–
Sparse packed fp64 models might become the new hotness… So few FLOPS — must be safe! 😀
-
Open Source LLMs and Licensing on Kaggle Competitions
By
–
I think btlm is cool too — but stablelm seems to work well too afaict. Where on kaggle would CC BY-SA be a problem? Are there some comps that require that solutions must be usable in closed source software?
-
Public Chemistry Resources Access Without Surveillance Implications
By
–
The textbooks used by chemistry students, along with the papers they're based on, are publicly available, and you don't go on a list.
-
Virologists Are Not Soldiers: Knowledge Access Ethics
By
–
Virologists are not soldiers, and people learning virology (regardless of whether by reading textbooks or chatting with an LLM) are not toddlers.
-
AI Model Training on Academic Papers for Student Education
By
–
No, it sounds like you're wrong on the facts of what the model did. It was *literally* trained on academic papers, and *only* returned information already in those papers. The virology students are encouraged to directly find, read, and discuss papers such as these.