00:00Can NVIDIA's kill switch save us from AI going rogue?
00:04Lately, we saw a string of AI scares,
00:07with AI agents stepping beyond their assigned guardrails.
00:10Now, US chip giant NVIDIA has launched a new safety architecture
00:14to monitor and quarantine AI agents.
00:17So, how does it actually work?
00:19What does it mean for users?
00:20What's the real reason behind NVIDIA's move?
00:23And why is Pope Leo getting involved?
00:28First of all, NVIDIA did not introduce a kill switch,
00:32but a software sandbox and a watchdog for AI agents.
00:36If you're using an AI agent, it can plan tasks, use software,
00:39and take action on its own.
00:41That comes with the risk of AI agents going beyond the limits they are given.
00:45To prevent this, NVIDIA developed a safety platform
00:48designed to monitor, restrict, and isolate AI agents.
00:51It has two elements.
00:53OpenShell keeps AI agents in isolated, restricted environments,
00:57so-called sandboxes, with tight limits to what they can do.
01:01Developers can set rules for which files, networks, and tools the agent can access.
01:05The system then monitors and enforces those rules while the agent is running.
01:10Sentry acts like a watchdog.
01:12It monitors AI agents' behavior, ready to quarantine them if they cross a line,
01:17and it runs on separate NVIDIA hardware.
01:19What's new about this?
01:20The safety mechanisms sit outside the AI agent, so it cannot simply reason around them.
01:26NVIDIA absolutely has the right idea here,
01:29which I think can be succinctly put as don't rely on the AI to police itself.
01:34Instead, let's put the limits outside of it where it can't reach them.
01:39NVIDIA says Sentry can quarantine rogue AI within milliseconds.
01:44So it's not a kill switch, but more like a combination of walls and a watchdog.
01:49Just how confident is NVIDIA with its new platform?
01:52Well, they say it could have prevented incidents like the hugging face attack.
01:56But it's still too new for extensive, independent testing, so we'll see.
02:01Anyhow.
02:01What does it mean for users?
02:04Look, if it's widely adopted, it could become a common standard and one I think that the ecosystem desperately needs.
02:11For users, safer agents could mean letting AI handle more sensitive tasks,
02:15from managing files to making bookings or shopping with your credit card.
02:19With reliable safeguards in place, AI agents could eventually help manage everything from your finances to healthcare appointments.
02:27Countries and companies could build their own AI systems while maintaining greater control over security, infrastructure, and data.
02:36So, why is NVIDIA doing this?
02:41NVIDIA dominates the hardware infrastructure behind today's AI boom.
02:45Calls for tougher regulation or even a pause in AI development are not in NVIDIA's interest.
02:51Safer AI would be adapted by more companies, which would also be good for NVIDIA.
02:56The more AI systems companies use, the more demand there is for NVIDIA's chips and computing infrastructure.
03:02Guess who's skeptic about NVIDIA's move?
03:04The Pope.
03:05He pointed to what he sees as a contradiction.
03:08The people building AI safety tools are often arguing against stricter regulation.
03:12He's the same one, however, that says there should be no limits placed and no government regulation.
03:18He refers to NVIDIA CEO Jensen Huang, who has been outspoken about his view.
03:23He thinks that AI safety is mainly an engineering problem and not one that requires further regulation.
03:29NVIDIA's answer is clear.
03:31Build stronger technical safeguards, not more regulation.
03:35Whether that approach is enough is exactly what critics are questioning.
03:38Regulators help set the environment and the safeguards, but we also need independent, ideally even third-party auditors and assurers
03:47to make sure that the claims being made by organizations like NVIDIA, which absolutely could be true, are backed up
03:55by reliable and robust evidence that the public can trust.
03:58What do you think?
03:59Can we cage AI?
04:00Why don't we cage define?
04:02Why don't we cry?
04:04Let's cut-'44-04-04-04-04-04-04-04-04-04-04-04q-11-1, where can we call AI?
