All-In PodcastArvind Krishna & Nikesh Arora: Quantum Threat, Open vs Frontier AI, AI Cyber Warfare & SaaS Collapse
All-In Podcast
Quantum risk and timelines 0:00
Quantum computing is framed as a real threat to today’s cryptography, even if much of the current field is still uneven. A DARPA competition has already narrowed a large field down to about 13 groups. The key point is simple: if workable quantum machines arrive around 2030, they will break today’s encryption, and even stolen encrypted disks could be cracked years later.
AI models and model choice 9:31
The discussion then shifts to how companies should use AI models. During experimentation, teams should use the best and easiest model they can find so they learn quickly. For most enterprise work, though, the likely path is open weight or open source models trained on proprietary data, because they are cheaper, easier to audit, and better for IP control. Frontier models stay reserved for high-stakes moments, like stopping an active breach.
Defending models and code 15:01
Nikesh Arora says model makers are unlikely to enter his business if he watches their boundaries closely. The models can help find bad code and vulnerabilities, but they still hallucinate, so you test what they produce before using it. He says attackers only need one flaw, while defenders must find and patch every flaw, which makes the job harder and more expensive.
Race, regulation, access 19:00
They describe AI as a race shaped by alliances, cheap attack, and technical superiority. They say the builder remains responsible for what is deployed, even if agents helped write it, and that slowing one country will not slow others. The bigger risk is unequal compute access: the best capacity will be clustered in a few places, leaving others farther from the center with less ability to build at scale.
Capital and labor 29:01
The speakers say capital-heavy countries tend to win these cycles, and labor will matter less than before. They think the real test is social attitude: whether people accept job displacement when AI replaces work and new jobs follow. They then turn to vertical integration. IBM’s mainframe lasted 65 years because it controlled nearly the whole stack, once even making its own wafers, chips, and hardware. The point is to keep profit inside the company. Nvidia is described as the modern version, managing the stack from the dirt up and deciding which layers to commoditize.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
