Jensen Huang: AGI is already here, doom predictions are made up, RSI is just old tech combined

创业邦 · wechat · 2026-09-16

At the All-In Summit 2026, NVIDIA CEO Jensen Huang made several pointed calls in a wide-ranging interview (via Tencent Tech).

AI safety: Safety and innovation aren't in conflict; incidents should get engineering root-cause analysis and be controlled via technology, process and testing. He said ex-Anthropic researcher Kokotajlo's doom predictions lack scientific grounding, and favors multiple third-party evaluators for checks and balances.
Doom narratives: He listed predictions that never happened (radiologists replaced, 90% of code AI-written, 50% of entry jobs gone) and called assigning 'human extinction probabilities' to AI irresponsible — such predictions should be recorded and revisited.
Recursive self-improvement: not a mystical technology, just a combination of context, RL, synthetic data and LoRA, with standard eval and regression testing before release.
Open vs closed models: both essential; $400B of VC went to AI-native companies in the past 6 months, 80% using open models. China, with its engineering and math talent, may become a major contributor to the open-source ecosystem, and its discourse is more pragmatic.
AGI: 'We've already reached it' — with superintelligence in narrow domains (autonomous driving at one-tenth human accident rates, protein synthesis and virtual screening).
Business: NVIDIA invests where the ecosystem is bottlenecked (supply chain, land, power) and stays low in the stack; builds models like Alpamayo because customers need them; regional clouds are growing more important. On Musk's Terafab, he said producing chips there is 'discussable.'

More from AGI Musings