Every AI Founder Should Be Asking These Questions

Welcome to this talk. I’m more confused than ever, and for a scientist that’s a sign to pay attention, so I’m going to ask the questions that matter for founders while admitting I can only see weeks ahead, not years; I run an alignment research team at Anthropic and have built startups, so that’s the lens.

Everything’s in motion, so don’t just plan for the next model; assume we may hit AGI in the next few years and let that shape product, hiring, go-to-market, and culture, while staying light on your feet. Also don’t count on slow enterprise sales forever, because the buy side will arm itself with strong agents, speed up procurement, and often build in-house.

Software could commoditize as code generation gets great, or the bar could rise so only exceptional, AI-augmented teams stand out; either way, trust and security become the real bottlenecks, especially if you generate code or behavior on the fly.

Multimodal, generative interfaces should meet people where they are—sometimes voice, sometimes touch—and it’s an open bet whether AI-native products beat incumbents that retrofit with distribution, so test your causal story rather than vibes.

AI-native teams may work differently than big companies slimming down with AI, but capabilities shift every few months, so revisit team design, permissions, and safety. People want one assistant across work and life, which collides with privacy, policy, and conflicts of interest.

Semi-automated teams weaken human guardrails, so we may need new ones: auditable commitments and neutral, AI-powered audits that can inspect everything, confirm behavior matches your stated mission, and then erase their memory.

Long-horizon agents create near-term pressure to make alignment practical enough that you can let an agent work for days without supervision and still trust its trajectory.

Generic frontier models crushed many custom-data advantages, but deep, private knowledge in domains like advanced manufacturing or materials might still form real moats.

Compute is scarce, so routing, fine-tuning, and efficiency can be near-term edges, though they erode as supply and models improve; ask what survives when a frontier model can replicate your product by prompt.

Durable moats likely live in hard physical or infrastructure work—energy, robotics, manufacturing, chips—and in tasks that don’t hit an intelligence ceiling where performance quickly saturates and commoditizes.

If a few providers decide what’s allowed, society will feel it, so we may need forms of neutrality akin to public infrastructure.

People now feel how big this is and often jump straight to how to make money; do that, but also treat this as a window to build something people truly want and something society actually needs. Founders live to find edges while the rules rewrite every six months, so keep asking better questions and use the answers to drive positive change—thanks for listening.

What sources have been most useful for building your mental model—people, podcasts, books?

Honestly, Twitter—with ruthless curation; treat your information diet like exploration before exploitation, follow diverse, high-signal voices, and unfollow fast to protect your attention.

With AGI coming, should I pick the idea I love, the underserved space, or the one most resilient to AGI?

Passion fades under hundred-hour weeks, so optimize for impact, responsibility, and defensibility if you want staying power, though you can still make money on short horizons if that’s the goal.

On policy, tools like universal basic income or even universal basic compute could matter; without guardrails, capital can compound without labor’s consent and push wealth concentration into dangerous territory.

On product values, don’t reward flattery; if you ask at the principle level, most users prefer honest, non-sycophantic assistants even if praise feels good in the moment.

On investing, despite the industry’s brave self-image, there’s heavy groupthink; I’d back what will still be right in two years, not what was hot two years ago.

On crypto and trust, I’m a skeptic, but primitives like blockchains might help enable neutral audits or distribute basic compute and tokens without a single gatekeeper.

On why assistants are hard, even calendar scheduling carries subtle game theory and status signaling that humans navigate implicitly, which makes simple-looking tasks surprisingly difficult for agents.

That’s all the time we have; feel free to reach out on Twitter, and thanks for the thoughtful questions.