- Openai’s material manager warns future AI models need material death switches in real time
- Richard Ho highlights networking, memory and power challenges in the scaling infrastructure
- He calls for references, observability and cross partnerships to deal with reliability and confidence
An OPENAI senior executive warned that future IA infrastructure will require material safety features, including Kill switches.
Richard Ho, responsible for the company’s equipment, made these remarks during his speech at the top of the infra in Santa Clara.
“It must be integrated into the hardware,” said HO. “Today, a lot of safety work is in the software. It assumes that your hardware is secure. It supposes that your hardware will do the right thing. It assumes that you can strive the equipment. I do not say that we cannot strive this equipment, but I tell you that these things are unpleasant, the models are really devoted, and so as I want to be able to is ensuring you. “
Safety measures in terms of silicon
Ho argued that the growth of the generative AI obliges a rethink of the architecture of the system and described how future agents will be with a long lifespan, interacting in the background even when a user is not actively committed.
This change requires an infrastructure rich in memory and low latency to manage sessions and continuous communication between several agents.
The networking, said HO, becomes a bottleneck. “We will have to have real -time tools in them – which means that these agents communicate with each other. Some could look at a tool, some may do a website search. Others think, and others need to talk to each other. ”
HO has described several material challenges that must be met, including the limits of large bandwidth memory, the need to integrate 2.5D and 3D chips, advances in optics and extreme energy requirements that could reach 1 megawatt per rack.
The proposed OPENAI security measures include real -time killing switches integrated into AI clusters, telemetry to detect abnormal behavior signs and secure execution paths in processors and accelerators.
Ho has wrapped things by saying: “We do not have good benchmarks for architectures and equipment devoted to agents, and I think it is important to know about latency walls and latency tails, what is efficiency and power and things like that. We must have good observability as a hardware functionality, not as a debugging tool, but constantly built and constantly monitoring our equipment. “
“The network is a really important thing, and while we are heading towards the optics, it is not clear that the reliability of the network is there today. We must get there with enough tests of these optical test beacons and these other communication test beacons that show that we really have reliability.”
Via The next platform
