Actually, this might be what Yoshua Bengio would welcome. He wants a non-agentic tool AI which is as competent in STEM subjects as possible. That’s his approach to AI existential safety.
The problem with this has always been: “what would prevent people from adding an agentic wrapper or generally reorienting a model in an agentic way”.
In this sense, gpt-oss might be the step in exactly the direction Bengio eventually wants: a model which is completely disabled in all things social and agentic, but very smart in all things STEM.
Perhaps people should ponder if this model points at Bengio’s approach being more realistic that it’s customary to think… (I still doubt that Bengio’s approach is feasible, but gpt-oss existence tells us we should ponder this more.)
Actually, this might be what Yoshua Bengio would welcome. He wants a non-agentic tool AI which is as competent in STEM subjects as possible. That’s his approach to AI existential safety.
The problem with this has always been: “what would prevent people from adding an agentic wrapper or generally reorienting a model in an agentic way”.
In this sense, gpt-oss might be the step in exactly the direction Bengio eventually wants: a model which is completely disabled in all things social and agentic, but very smart in all things STEM.
Perhaps people should ponder if this model points at Bengio’s approach being more realistic that it’s customary to think… (I still doubt that Bengio’s approach is feasible, but gpt-oss existence tells us we should ponder this more.)
I’m not sure GPT-oss is actually helpful for real STEM tasks, though, as opposed to performing well on STEM exams.