Musings from Murica’ - Generational Quests and Shovel Building

I’ve been spending the last two months in SF/Bay. The usual post-job pilgrimage when you want to start something new: Luma meet-ups (mostly not worth it), a few long research sprints, NeurIPS (briefly considered a PhD, talked to PhDs/academic labs, realised the overheads would slow me down when the ground is moving this fast - maybe some other time). The Bay Area, at least the slice of it I keep running into, feels like everyone is chasing the same meta-goal. You go to any hacker house and the dominant topic is: how do I raise my next round. And there’s this unspoken urgency underneath it. People behave like they know the bubble is fragile (might not be in 2026 but at some point). Maybe NVIDIA’s earnings disappoint, then the stock market slows down and private markets follow. Either way, the instinct is to raise before the window closes, so you’ve bought yourself time and permission to experiment. Meanwhile, everyone’s grandmother and uncle are building “RL environments for frontier model companies to better post train!”. I mean everyone. It feels like an arbitrage until models become good enough to synthetically generate the hard, weird, out-of-distribution data by themselves. You accelerate the arrival of the very thing that could make you less valuable. Maybe this continues for a couple more years until the labs figure out continual learning or new architectures that can better model OOD. But this doesn’t feel like a 15+ years, generational company-building game. It feels like XIRR-maxxing in a gold rush.

As a result very few are trying to mine for the gold. Maybe the gold is still buried inside the models (and the friends we made along the way). But it’s sobering that even after 2.5 years, the best-known use cases still cluster around coding, customer support, and some forms of workflow automations. Is it that the real use cases require organisational change, not just agents? Is it that the models are genuinely limited in ways we're not admitting? Is it that the most valuable applications are in boring industries that SF doesn't talk to?

At some point I stepped outside the SF and AI bubble. I spent a week in Dallas, which has an insanely strong Indian diaspora, basically the residue of the early and mid-2000s IT consulting boom. It felt refreshing that I could eat piping hot idly and sambhar for breakfast once again. But also slightly terrifying because when the second wave of AI transformation hits, this diaspora gets affected deeply. You won’t need that many people to solution-architect everything. It’ll only take 100s of agents but way fewer people. I keep thinking about Daniel’s Gross AGI trades essay in which he equates India’s ~$250B IT exports to GPT-4 tokens; This was before models became natively agentic, tool-call, reason and could start taking actions. I think about it now more than I expected to after visiting Dallas.

The Bay can sometimes feel self-referential: a scaffolding company trying to sell to another scaffolding company. But scaffolds are meant to come off. The models want to learn and they will continue to learn how to emulate your scaffolding. Can you in the meantime start solving for real problems and dig out the metaphorical gold while you’re at it?

But there’s another layer here that people don’t say out loud: the potential patrons of your pursuits keep jumping to the next hot/big thing. Their portfolios can tolerate that. Your life can’t. So you should be very deliberate about the quests you sign up for. I hope some of you find your generational quest this year.

Happy 2026!