EOs, founders, and leaders across the AI industry gathered at the HumanX event in San Francisco this week. The buzziest topics will surprise you.

The Deep View team was on the ground at the event all week, including senior reports Nat Rucio-Licht and Sabrina Ortiz, CEO/founder Faris Kojok, and yours truly.

Here were the major topics that perked up our ears:

• Slow down to speed up: AxonIQ CEO Jessica Reeves was on my HumanX panel, "What do developers need to ship AI faster," and she emphasized that the smartest enterprises are learning to slow down and build a stronger foundation of data and security so that they can speed up. This sentiment was echoed by much of the enterprise contingent this week.
• The compute crisis has downstream effects: It's no secret that the entire AI industry is massively compute-constrained. The compute shortage is reportedly the reason OpenAI shut down Sora. It's why Anthropic just signed a key deal with Google and Broadcom. But Jennifer Li of a16z also told me at HumanX that it's having a big impact on startups as well. If they want to do anything involving models, they'll often need to raise $50-$100 million to secure the capital to access meaningful amounts of compute. That's some serious startup inflation.
• Token costs are soaring: For months, I've been hearing enterprises complaining about the fact that inference/token costs are killing ROI on AI projects, and they need to find more sustainable alternatives (like this one I've written about). And it's not just affecting businesses. Howie Xu of Gen told me that he's constantly getting requests from his teenager to refill his token budget because he's using AI coding tools and agents and constantly burning through his budget. On the other hand, it’s never been a better time to be an inference provider.
• Agents are out of control: So many of the executives running around HumanX were focused on the same problem: How to implement better security, governance, and controls for AI agents. Now that the torrent of agent announcements and momentum has captured the zeitgeist for 2026, organizations are scrambling to contain the fallout. OpenClaw, Manus, and Hermes have made it so easy for workers to spin up their own agents that, along with all the potential benefits, they can introduce all kinds of issues. If you want more context on the issue, see my latest podcast interview with Guild AI CEO James Everingham.

Our Deeper View

One of the hottest terms going around the AI industry right now is "AI psychosis," as coined by Andrej Karpathy, to describe the mania among coders working with agents to do 10x the work they used to do on their own. However, I got a good gut check on that at HumanX, where people said it's popular to say you have AI psychosis right now. The reality is that most of these folks aren't actually working 18-hour days and getting massive productivity gains using a swarm of agents. Nevertheless, "agents" has nearly replaced "AI" itself as the hottest topic in the industry, and that fact was on full display this week at HumanX.