tom

The cost of equivalent intelligence falls ~10x annually, so founders should not optimize for current token costs; within 6-12 months, AI will be cheaper than human engineers for any task, fundamentally changing startup economics.

Writing software is no longer a 'hard bit' — AI makes replication trivial. Durable startups must tackle genuinely hard problems: regulatory barriers (banking licenses), physics (space launch, nuclear), or complex B2B sales. Pure SaaS like Calendly is 'basically dead'.

A YC solo founder building autonomous self-improving agents hit a crisis of confidence at 6 weeks; co-founders provide essential emotional modulation (pulling you up when down, grounding you when manic) that AI cannot replace. Solo billion-dollar companies are theoretically possible but economically irrational.

While AI reduces capital needs for software, it enables founders to pursue harder problems (small nuclear reactors requiring $800M, regulated banks) that demand massive capital. VC will fund these ambitious bets, yielding societal breakthroughs like abundant energy and curing disease.

Most pivots are driven by founder sadness and sales rejection, not genuine hypothesis invalidation. Good pivots come from customers pulling you into a bigger adjacent problem (GoCardless: sports teams → business payments). Bad pivots abandon deep domain knowledge for shallow new ideas.

With AI making code generation near-free, distribution (getting customers, standing out) is now a key differentiator. Early-stage founders should do unscalable white-glove service; later-stage must invent novel distribution meta-games (like Brex buying all SF billboards).

Users will soon just type queries; the system will automatically route to the right intelligence level (smart model spawning cheaper sub-agents). Open-source routing models and big labs will handle this, making model selection invisible and driving costs down.

High agency = belief your actions produce impact. Develop it by picking projects that are hard but tractable, seeing output, then increasing difficulty (side project → startup → Zip2 → PayPal → Tesla → Mars). Co-founders accelerate this loop.