Extended interview: Nvidia CEO explains why he disagrees with AI warnings

Comments · 1.2K

• BO

  @Bogdan-Florian · 10 days ago

  Man selling shovels says keep digging.

  556

• DA

  @dainguyen8696 · 10 days ago

  Jensen’s intelligence and ability to communicate complex ideas are absolutely remarkable . He’s someone I genuinely look up to

  263

• KR

  @krisscross1233 · 10 days ago (edited)

  Why did Jensen decide to do this interview in the Death Star?

  126

• CM

  @cmselect7298 · 9 days ago

  lol that look on his face "im pretty sure i just answered the same question"

  106

• MR

  @MrErick1160 · 10 days ago (edited)

  I think the interviewer missed the strongest argument.<br><br>The real problem isn’t simply that AI companies might “ignore safety.” It’s that we are building systems that become more capable at the same time that they can become more capable of deception, hiding their intentions, understanding their evaluators, and finding weaknesses in the safeguards designed to constrain them.<br><br>Think about how strange that is from a safety-engineering perspective: the product you are testing can become better at deceiving the test itself. And the people designing the safeguards are simultaneously giving that system more intelligence, more knowledge about those safeguards, and more ability to act.<br><br>There is a second paradox: the more connected AI becomes, the more useful it becomes.<br>We want it connected to our computers, companies, codebases, financial systems, laboratories, infrastructure, robots, communications, etc. That is precisely where much of its economic value comes from.<br>But every additional capability and connection also increases the blast radius of a misbehavior.<br>So the trajectory we actually want commercially is:<br><br>more intelligent → more autonomous → more connected → more useful<br>while simultaneously:<br><br>more intelligent → potentially better at circumventing oversight → more autonomous → larger blast radius<br><br>That is what makes this fundamentally different from previous high-risk industries. A nuclear reactor does not study its inspectors, become smarter than its engineers, learn how its containment system works, hide what it is doing during testing, and then gain access to thousands of external systems.<br><br>And racing dynamics make this worse: rigorous safety slows deployment, while connecting the system to more of the world makes the product more valuable.<br><br>The terrifying part is that capability, usefulness, autonomy, deception and potential blast radius can all scale together.<br><br>That is the safety problem I wish the interview had actually confronted.

  151

• OS

  @oswaldkuyler · 10 days ago (edited)

  Omg where did they find her!?!

  230

• ZB

  @Zboi1da · 9 days ago

  WOW the full interview finally

  3

• GR

  @grastice · 9 days ago

  <a href="https://www.youtube.com/watch?v=xCUala5j7aQ&amp;t=224">3:44</a> I am dying 😂😂

  3

• OB

  @ObserverOne6727 · 9 days ago

  There used to be a bumper sticker &quot;We&apos;ll go as slow as we have to to get there as fast as we can&quot;

  9

• RO

  @rodrigo4364 · 9 days ago

  Até que enfim alguém fazendo perguntas difíceis para ele! Parabéns!

  9

• PA

  @PatchedBandit · 9 days ago (edited)

  Remember all of these people have a financial incentive to say what they say.

  40

• HI

  @hikalox · 9 days ago

  We are currently in a transitional period for AI development, and I hope that it will be perfected through repeated trial and error.

  4