Google DeepMind researcher Andreas Kirsch contends that national or local AI rules alone cannot stop misuse of capable models, so hardware makers would need to throttle performance for unauthorized inference runs while also restricting hosting and inference services.
@xeophon @sdmat123 No local regulation is sufficient. No hosting, no providers, no inference. Harder to control local inference but could work together with HW providers to limit performance for uncontrolled inference etc
@sdmat123 @BlackHC I am sure that Chinese companies are stoked to work together with Anthropic to establish robust safety strategies, mitigations and safeguards!
Kirsch offered no technical details on how performance limits or DRM-style controls might work in practice, leaving their feasibility unaddressed in the thread.
Replies pointed out that shifting workloads to data centers in countries with weaker oversight might evade hardware limits entirely, with no global coordination plan discussed.
Users criticize the researcher's claim that local AI rules are insufficient because it seems to endorse banning open source models as anti-freedom.
No Digg Deeper questions have been answered for this story yet.
@BlackHC @sdmat123 Sounds good, what should bad actors do against that? Just… use DCs in another country with less regulation?
@xeophon @sdmat123 No local regulation is sufficient. No hosting, no providers, no inference. Harder to control local inference but could work together with HW providers to limit performance for uncontrolled inference etc
@BlackHC @sdmat123 Leaving even aside the business interests of orgs like MSFT, NVIDIA, the whole startup ecosystem which all want/need open models to flourish.
Banning open source models is anti-freedom and authoritarian
@BlackHC @xeophon "uncontrolled inference" - you want states to mandate DRM on computation?