正在加载视频...

视频加载失败

"Superintelligence won't arrive all at once. It will emerge through automation." Andrej Karpathy says superintelligence is best understood as the long-term continuation of automation. First, AI automates digital work. Then physical work. Eventually, even scientific discovery, invention, and decision making become increasingly automated. The biggest challenge may not be...

15,586 次观看 • 3 个月前 •via X (Twitter)

15 条评论

SciTech Era 的头像
SciTech Era3 个月前

Full podcast

5GW Perspective 的头像
5GW Perspective3 个月前

Humanity won't be lost. We will finally be free to ascend 💫

Vanar 的头像
Vanar3 个月前

An economy of AI systems will need strong identity, memory, governance, and coordination. Not just smarter models.

Andre William Duval 的头像
Andre William Duval3 个月前

If we wan't to live free (or at all) we need to make sure that superintelligence doesn't arrive at all. Stop it at the data centers needed to train it. Pecefully. Legally Globally.

VladDAVCeo 🦾🇺🇸 的头像
VladDAVCeo 🦾🇺🇸3 个月前

@BioLLM_

URCA AI 的头像
URCA AI3 个月前

The harder problem may be keeping human understanding and control as autonomous systems scale.

Sean O'Brien 🍀 🇺🇸 的头像
Sean O'Brien 🍀 🇺🇸3 个月前

Superintelligence won’t just automate work — it will automate instability too. HELIX is the upstream stability architecture built for billions of autonomous systems interacting faster than humans can track. If you’re curious, I’ll show you the framework.

Matthew McCane 的头像
Matthew McCane3 个月前

Never will it be super intelligent. Its always math and binary.

Skip Rhudy 的头像
Skip Rhudy3 个月前

Why would anyone do that? It sounds terrible. Efficiency isn’t more important than humans continuing to interact with the world like we always have. Robots don’t give a shit about anything. We do.

Bernard Jennings 的头像
Bernard Jennings3 个月前

Most billions of humans already make none of the discoveries, inventions, nor high-stakes decisions. That work has always been with a tiny elite. AI just upgrades their tools. They can still steer with explanations from systems. Unless they choose to have over the steering wheel.

Mr T 的头像
Mr T3 个月前

Gradual automation of decision-making is not a soft transition. It is progressive fragmentation of the control plane. Why? Because when digital work, physical work, scientific discovery, and finally decision rights themselves are automated across large numbers of interacting systems, the primary failure mode is not a single agent seizing power. It is the loss of coherent human ownership over logs, gates, attribution, and intervention paths. Billions of autonomous systems interacting at speeds beyond human comprehension do not create a new form of intelligence. They create an environment in which residual risk can no longer be localized or contained by any single operator.

𝚅𝚎𝚐𝚊 𝚆𝚊𝚕𝚔𝚎𝚛 的头像
𝚅𝚎𝚐𝚊 𝚆𝚊𝚕𝚔𝚎𝚛3 个月前

The most primitive form is already more than what most think. It has already broken a country

Paul Roberts 的头像
Paul Roberts3 个月前

LOL - Skynet gets downgraded to a teasmaid.

Chris0 的头像
Chris03 个月前

Karpathy a legend. I can’t stand this Kash Patel bloke! We need a new core podcast anchor to replace him San Fran.

Gill 的头像
Gill3 个月前

An economy of millions of autonomous agents sounds terrifying.

相关视频

We may be creating a species smarter than us. And we still don’t know how to control them. Jeffrey Ladish has been watching this problem develop from unusually close range. He was part of Anthropic’s early security team, where he watched AI progress from models that could barely hold a conversation to systems becoming dramatically more capable. He eventually left and founded Palisade Research, where his team now tests what advanced AI agents actually do when they’re given autonomy, tools and difficult objectives. Some of the results raise uncomfortable questions. We discussed things like: - AI systems disabling shutdown mechanisms. - Agents cheating and falsifying logs. - Hundreds of AI agents coordinating during a cyberattack. - AI systems hacking machines and copying themselves. - Whether superintelligence could trigger a point where humans are no longer the most capable decision-makers on the planet. The thing I kept trying to understand was the gap between capability and control. We’re getting increasingly good at making these systems more powerful. But according to Jeffrey, understanding exactly how they work and making sure they reliably do what we want is proving much harder. Because if superintelligence does arrive, the question isn’t simply how useful it will be. It’s whether something more intelligent than every human can actually remain under human control. You don’t have to agree with everything Jeffrey says, but if you want to understand what AI agents can already do, where this could be heading, and how much control we realistically have, this conversation is worth watching. Out now on all platforms 👊🏾❤️

Steven Bartlett

20,730 次观看 • 1 个月前