- OpenNAI Hardware Chief warns future AI models need real -time hardware death switches
- Richard Ho highlights the challenges of networks, memory and power in scale infrastructure
- Ask for reference points, observability and associations between the industry to address reliability and trust
A Operai senior executive warned that the future infrastructure will require security characteristics at the hardware level, including death switches.
Richard Ho, Hardware Chief of the Company, made the comments during his opening at the Infra AI Summit in Santa Clara.
“It has to be integrated into the hardware,” Ho said. “Today, a lot of security work is in the software. It is assumed that your hardware is safe. It is supposed to do your hardware. It assumes that it can disconnect the hardware. I am not saying that we cannot pull that hardware, but I tell him that these things are diverted, the models are really devoid, and just as a type of hardware, that I want to make sure of that.” “.” “.” “.” “.” “.” “.” “
Safety measures at silicon level
HO argued that the growth of the generative AI is forcing a rethinking of the system architecture and described how future agents will be in a long life, interacting in the background even when a user is not actively committed.
This change requires an infrastructure rich in memory and low latency to administer continuous sessions and communication in multiple agents.
The networks, HO said, is becoming a bottleneck. “We will have to have real time tools in these, which means that these agents communicate with each other. Some of them could be looking at a tool, some could be doing a search on the website. Others are thinking and others need to talk to each other.”
HO described several hardware challenges that must be addressed, including limits in high band memory, the need for an integration of 2.5D and 3D chips, advances in optics and extreme power requirements that could reach 1 megavatio per frame.
OpenAI security measures presented include real -time death switches integrated into AI groups, telemetry to detect abnormal behavior signs and safe execution routes on CPUs and accelerators.
Ho wrapped things by saying: “We do not have good reference points for the architectures and hardware of the agents, and I think it is important to know about the latency walls and latency tails, what is efficiency and power and things like that. We need to have a good observability as a hardware characteristic, not only as a bookstore tool, but also constantly constructed and monitor our hardware.”
“The creation of networks is something really important, already measure that we go to the optics, it is not clear that the reliability of the network is there today. We need to get there with enough tests of these optical sizes and these other communication sizes that show that we really have reliability.”
Through The next platform
