Excellent comments Nathan. I wonder what your view of Z.ai's native AI Safety capabilites is. They and Alibaba seem to have the best teams, but all Chinese AI labs are short of compute, so one wonders how much compute is committed to safety testing, particularly of more complex agentuc deployments with orchestration, got cyber related tasks for example....sounds like Z here us taking an Anthropic-like as approach on pre-release testing, without yet a government "voluntary" mandate...
I largely agree. My read is they dedicate WAY less compute, as they're way less philosophical about AGI, but we don't have good data here. Getting information on this sort of stuff public would be a good target for international cooperation. It doesn't really impact anyone but helps monitor the ecosystem.
Also, the "two week lag" to weights after announcement is feeling similar across orgs and maybe is a domestic policy (or unwritten rule)
Excellent comments Nathan. I wonder what your view of Z.ai's native AI Safety capabilites is. They and Alibaba seem to have the best teams, but all Chinese AI labs are short of compute, so one wonders how much compute is committed to safety testing, particularly of more complex agentuc deployments with orchestration, got cyber related tasks for example....sounds like Z here us taking an Anthropic-like as approach on pre-release testing, without yet a government "voluntary" mandate...
I largely agree. My read is they dedicate WAY less compute, as they're way less philosophical about AGI, but we don't have good data here. Getting information on this sort of stuff public would be a good target for international cooperation. It doesn't really impact anyone but helps monitor the ecosystem.
Also, the "two week lag" to weights after announcement is feeling similar across orgs and maybe is a domestic policy (or unwritten rule)