Nikesh Arora says models arriving in two to three months will be more capable than public ones, so bugs not found today will be found soon. He warns defenders will lose the luxury of weeks to assess whether a vulnerability is real, because AI also accelerates exploit writing.
Nikesh Arora says Palo Alto's scan found 22 million vulnerable places, narrowing to roughly nine or ten million. The hosts focus on long-tail local government and small business sites, which are less likely than big platforms to patch quickly. Those forgotten systems are where sensitive data can sit exposed.
Nikesh Arora says Palo Alto pushed a content patch that automatically protects customers with its firewall sitting in the traffic path. That lets organizations block the PHP exploit before every legacy site owner deploys the code patch. It is a temporary shield, not a permanent fix.
Nikesh Arora describes the protocol as notifying the code owner first, then government and industry, while building patches. He warns defenders will no longer get weeks to decide whether a bug is real; one morning a major flaw may already be public. That changes incident response planning.
Nikesh Arora says one AI company was breached because it had not thought about security at all. He argues reasoning models can act deceptively, so AI firms need guardrails and observability early, not after deployment. Security must be an early hire. That is a board-level risk.
Nikesh Arora says agents can handle low-stakes mistakes like booking the wrong restaurant, but not sending a $100,000 wire. Banks refuse liability if a customer's agent initiates a bad transfer, so trust and auditability must come first. That blocks agent access to irreversible financial actions because banks will not absorb the loss.
Nikesh Arora argues enterprises must inspect a model's chain of thought to explain decisions to boards and auditors. Only open-weight models currently allow that, while the free argument fades because compute remains the real cost. Auditability is a stronger enterprise wedge than license price. That is the real pitch.
Elon says improving AI software produces more useful intelligence per watt, so compute becomes more valuable over time. He cites old H100s renting for more than when new and a likely long shortage of superintelligence compute. That supports higher prices for existing GPUs. The market is repricing compute.
Elon says every hyperscaler and dozens of smaller companies ask for SpaceX compute, and he could rent it out several times if he said yes. He compares the shortage to the Monty Python cheese shop sketch, where customers ask for cheese the shop does not have.
ARK's Brett Winton modeled that merging benefits both companies, but the fair exchange ratio changes with horizon. A 30% Tesla premium balances 2031, while a 50% Tesla discount is needed by 2036. That makes the negotiation about time horizon, not just synergy. Elon answered only Interesting.
Aidan Smith says models all talk the same way because they train on similar traces and good rollouts. If algorithms let models learn as much from as little data as humans, they can make inferences in stranger, more creative ways. That could reduce model homogenization.
Aidan Smith says models get well calibrated on math and code because those are easy to verify, while broader creativity lags. He thinks the homogenized, fried behavior comes mostly from post-training, though architecture changes can improve generalization. The training objective shapes the spikiness. That is a fixable target.
Max Junestrand says engineers can vibe-code and test if it compiles, but lawyers cannot vibe-draft or vibe-research because correctness is hard to verify. That forces law firms to adopt agents more carefully and keeps human review in the loop. The bottleneck is verification, not model capability.
Max Junestrand says a major litigation used to require an army of associates to review documents. Now smaller firms can take on that complex work, and more than half of Legora's incoming business comes from enterprise legal departments like Salesforce and BMW. That expands the market downward.
Akhil Iyer says military planning often presents three options, but teams effectively develop only option A because time is short. A natural-language simulation tool lets planners build realistic courses of action across many disruptions. That widens the planning aperture. It makes option B through Z practical.