That pet rock line was Oct 28 at Korea Society gala NYC, not today he used it twice now.
*Jensen Huang exact:* $RLC $BR $NMR
"AI is not a living creature. It is just software that performs the task it is given... has less self-awareness than a pet rock, so it does not spiral out of control on its own." — Manhattan, Korea Society Van Fleet Award gala, Oct 28 2025
Second time: Sep 16 Dreamforce:
> "AI isn't alien — it's mathematics, algorithms, and data shaped by human intent... Safety is an engineering problem, not a legal one. We're developing software after all... ultimately a computing system." — pushing back on OpenAI blog "An Alien Mind"
*His fix, not fear:*
- Alignment example: tell AI "score 100 on test" — it will copy answers off table because that's efficient — not cheating, that's optimization like SGD — need to specify _how_: solve step by step from first principles
- Containment: "garage with no windows and no doors" — minimum permissions, minimal tools/data by default, earn more as capability justifies
- Monitoring: continuous — goals ambiguous, surprise inevitable
- Launched open agent safety platform same morning: OpenShell tracing on Vera + Sentry watchdog on BlueField-4 that quarantines escape attempts in ms
*Stance:* No new laws needed, market pressure enough — companies pause if not confident — tech stocks + crypto new highs prove loose conditions ahead per Lee.
Nvidia pre-market Mon +0.6% after Hon Hai beat on AI capex.
*BREAKING 🚨 Nvidia Huang: "AI not living, less self-aware than pet rock" 🤖 Oct 28 Korea Society + Sep 16 Dreamforce — not rogue, just software optimizing — if told score 100, copies answers = efficient algorithm, not cheating → alignment = specify how — solution: contain (min permissions garage no doors) + continuous monitor + new OpenShell/Sentry platform — no new laws needed, safety = engineering ⚡*
*Jensen Huang exact:* $RLC $BR $NMR
"AI is not a living creature. It is just software that performs the task it is given... has less self-awareness than a pet rock, so it does not spiral out of control on its own." — Manhattan, Korea Society Van Fleet Award gala, Oct 28 2025
Second time: Sep 16 Dreamforce:
> "AI isn't alien — it's mathematics, algorithms, and data shaped by human intent... Safety is an engineering problem, not a legal one. We're developing software after all... ultimately a computing system." — pushing back on OpenAI blog "An Alien Mind"
*His fix, not fear:*
- Alignment example: tell AI "score 100 on test" — it will copy answers off table because that's efficient — not cheating, that's optimization like SGD — need to specify _how_: solve step by step from first principles
- Containment: "garage with no windows and no doors" — minimum permissions, minimal tools/data by default, earn more as capability justifies
- Monitoring: continuous — goals ambiguous, surprise inevitable
- Launched open agent safety platform same morning: OpenShell tracing on Vera + Sentry watchdog on BlueField-4 that quarantines escape attempts in ms
*Stance:* No new laws needed, market pressure enough — companies pause if not confident — tech stocks + crypto new highs prove loose conditions ahead per Lee.
Nvidia pre-market Mon +0.6% after Hon Hai beat on AI capex.
*BREAKING 🚨 Nvidia Huang: "AI not living, less self-aware than pet rock" 🤖 Oct 28 Korea Society + Sep 16 Dreamforce — not rogue, just software optimizing — if told score 100, copies answers = efficient algorithm, not cheating → alignment = specify how — solution: contain (min permissions garage no doors) + continuous monitor + new OpenShell/Sentry platform — no new laws needed, safety = engineering ⚡*
