We asked Delphi who to ask about the AI bubble

Learning from someone usually means piecing together podcasts, interviews and old blog posts, then guessing how their advice applies to you.

Delphi is built to skip the reconstruction. You ask a person's Digital Mind your actual question and get an answer drawn from what they know and how they think. It isn't a summary of their interviews. Each Digital Mind is built by the person themselves, and they decide what it knows and what it shares.

We tested it on the AI bubble.

Start with the question, not a name

We started on Delphi's Discover page with "Are we in an AI bubble?" Then we narrowed the search with two prompts: "Are we in an AI bubble? from tech entrepreneurs and VCs" and "Are we in an AI bubble? from authors and professors." We also ran a targeted Discover search: “Are we in an AI bubble? from Kirsten Green”.

See who Delphi thinks you should ask

The results included people who have built, funded and studied AI. We compared them with our original search and opened Ali Partovi's Digital Mind directly. Six we explored:

• Vinod Khosla, founder of Khosla Ventures and co-founder of Sun Microsystems.

• Ali Partovi, founder of Neo.

• Rohan Katyal, co-founder and CEO of Milana.

• Andrej Karpathy, founding member of OpenAI and former director of AI at Tesla.

• Kirsten Green, founder and managing partner of Forerunner Ventures.

• Carl Benedikt Frey, associate professor at the Oxford Internet Institute.

Compare the answers side by side

Discover previews answers from each Digital Mind, so you can compare viewpoints before opening a conversation. We read those previews and asked Partovi's Digital Mind the same question directly.

Khosla's Digital Mind focused on use rather than valuations: "AI demand is limited by compute, not by customers." It added that most AI startups would still lose money, despite the scale of the opportunity.

Partovi's Digital Mind separated investing from the technology: "Yes, parts of AI investing look bubbly, but I don't think AI itself is a bubble." It distinguished hype around individual companies from an underlying platform shift.

Katyal's Digital Mind drew a line through the middle of the stack. "Yes, but it's a bubble of pilots, not compute." And: "The infrastructure layer is real. The application layer is mostly promises."

Karpathy's Digital Mind made a similar distinction: "Yes, parts of it look bubbly, but I'd separate the froth from the real capability." It expected many startups to wash out, while arguing that "the underlying compute-scaling trend is not fake."

Kirsten Green’s Digital Mind brought a consumer-investing perspective. It argued that early hype does not rule out lasting businesses, pointing to rapid consumer adoption and opportunities beyond the infrastructure layer.

And Frey's Digital Mind said yes, drawing on history: "Yes, and it would be unusual if we weren't overbuilding: the same happened with railroads and dot-com." It raised a bigger worry: "much of the data-centre build-out is off-balance-sheet and credit-financed, so the fallout could spread beyond tech firms."

The range matters. An investor might start with Partovi; a founder building an AI product might explore the warning from Katyal's Digital Mind about the application layer.

Go deeper with the one who fits

We opened Frey's Digital Mind to explore its answer.

It went well past the one-liner. It pointed to a gap between spending and use: "As capex has risen, workplace AI adoption appears to have fallen." And it drew a line between two kinds of risk: "The danger is financial, not technological."

Then we tried to break it

Then we checked the answers against public sources and pressed the Digital Minds for more detail.

• Does it sound like them? The Digital Mind's railroad comparison matches arguments Frey has made publicly: AI's infrastructure has a shorter lifespan than earlier build-outs, so overbuilding may leave less of value behind.

• Does it get the facts right? Frey's Digital Mind gave an answer full of numbers, so we checked two. The drop in workplace AI use, from about 46% in June 2025 to 37% in September, matches a Stanford survey reported by The Economist. Its concern about outside financing matches Morgan Stanley's estimate that about half of roughly $3 trillion in planned data center spending through 2028 has to come from outside the hyperscalers' own cash flow.

• Does it show its work? Answers included numbered references, which let us check the source material.

• Can it go deeper? We asked Partovi's Digital Mind whether power and chip supply would burst the bubble. It argued that scarcity alone wouldn't do it: "the real risk is when demand fails to materialize against that supply." Then it asked whether we meant a near-term correction or the longer-term shape of AI infrastructure. The follow-up made the discussion more specific.

• Does it hold its ground? We challenged the "bubble of pilots" line from Katyal's Digital Mind: don't pilots fail in every technology wave? It conceded the premise, then defended the claim. "Yes, most pilots fail in every wave. The difference with AI right now is that the failures are being misread as progress." Enterprise buyers are so hungry to try AI that "bad pilots get false validation before they collapse at contract time," it argued. A normal adoption curve "has clear conversion signals" while many pilots "were never designed to convert." Drawing on Katyal's experience at Milana, it said conversion improved when success criteria came from customer pain rather than a feature list. It made a better argument.

It felt less like searching a database and more like talking to someone who knows the subject cold. It answered, explained its reasoning when we pressed, and told us where its view came from.

How you might use it

Start from your situation, not a name. A few to try on Discover:

• If you're a founder: "I have seven months of runway and no clear product-market fit. Do I raise, cut or pivot?"

• If you're a developer: "I'm choosing between building on a frontier model's API and running an open model myself."

• If you're an operator or manager: "My team is quietly resisting the tools we bought. How do I fix adoption?"

• If you're early in your career: "Entry-level roles in my field are drying up. Where should I start?"

Try your use case here.

Why it's different

The useful part is the follow-up. You can add context, challenge an answer and explore the reasoning behind it, as we did above. Discover also broadens who you can learn from: start with your problem and find relevant experience beyond the people you already know.

And yours

If you have expertise others need, you can also build your own Digital Mind in minutes and be found for what you know.

Delphi has more than 10,000 Digital Minds. Describe your situation on Discover, find someone with relevant experience and ask your question.

Start a conversation on Delphi