AI spending is not slowing down. That was the central verdict at last week's Big Technology AI Summit in San Francisco, where OpenAI President Greg Brockman, Anthropic Labs Lead Mike Krieger, Box CEO Aaron Levie, and former Meta security chief Alex Stamos all pointed in the same direction: costs are rising because capabilities are rising, and appetite is outrunning efficiency gains. Levie put a number on it: Box tasks that once consumed 5,000 to 20,000 tokens now routinely hit 1 million to 5 million tokens per agent execution. Moore's Law is not repealed, it's just being lapped.
The cybersecurity signal buried in the summit is worth the full read on its own. Stamos argued the real AI cyber inflection point was not any recent export control announcement. It was the quiet arrival of the Opus 4 and GPT-5 series models, which handed human engineers capabilities that Stamos compared to high school track athletes suddenly running Olympic times. He named Mythos as the current best public model for finding software vulnerabilities. That is a specific, falsifiable claim with serious implications for every security team operating today.
Levie also made a pointed observation about U.S. AI policy that reframes the entire regulatory conversation. He argued that the White House's current practice of reviewing and effectively approving individual frontier models is functionally identical to what the 2023 six-month AI pause petition was asking for, just arrived through a different door. The original piece connects these threads across the full summit, including a look at Fable and what agentic systems doing more with fewer round-trips actually looks like in production.
[READ ORIGINAL →]