@vgr

There’s this old heuristic in SV-flavored startup land that I got from pmarca that it doesn’t matter if a startup gets a lot of stuff wrong by normal business playbooks, so long as it gets one thing really really right. Unspecified in there is what that “one thing” is, but the presumption is that it’s whatever the unique unfair advantage of the idea is. Usually this is unique to a given (good) startup. Its “unfair advantage fingerprint.” But occasionally for a few years, there is a shared “floats all boats” unfair advantage variable driving a period of synchronized evolution. Right now, obviously, the one thing to get “really really right” regardless of startup idea is AI. But it’s useful to start with the unsynchronized version of the idea. The idea makes sense because a normal business is close to its constraint boundary on all fronts and has no dimension of extreme leverage that it can asymptotically drive out on to “pay” for sins on other dimensions. A startup is a mechanism that is far from the ETTO frontier (efficiency-thoroughness tradeoff) but is also ideally alone in its corner, so it doesn’t matter whether it chooses efficiency (gets ops optimized) or thoroughness (gets everything past good enough). But “one thing really really right” means it must be oriented correctly in the OODA sense, which means the heading is towards the frontier rather than away. Depending on the business and technology, actual progress may be slow and frustrating and take a decade or a breathless acceleration that is over in a make-or-break year. The capitalization structure has to match the progress rate possible (enterprise, web tech, hard tech, biotech, climate tech etc all have different rates). But the orientation has to stay “really really right” throughout. Often, this has a profane-sacred inversion or weird quality to it. Eg: Amazon’s early focus on free cash flow over margin or revenue. Or later focus on pure SOA (the infamous memo in Yegge platform rant). Or Tesla’s steadfast refusal to go Lidar. But back to synchronized periods. SaaS is an older example. For a few years it didn’t matter what the business idea was. You had to get the SaaSification structure right or die. In a synchronized era, everyone is trying to head towards the same foggy frontier and you get a handful of mimetic herds that share orientations and sink or swim together. Pivots are not individual reorientations in this phase but defections/joinings of different herds. Towards the endgame you have 1-2 herds that have figured it out and all survivors preferentially attach to those. After that it’s a land grab on the frontier. An emerging frontier is typically a rather small target even if it’s apparently very visible and dominating everyone’s attention. If your orientation is off, you’ll end up missing it. Everyone can see where the moon is in the sky, but actually navigating to it with a spacecraft involves very tricky orienteering. AI is like the full moon right now. You can already see companies making fatal mistakes. For eg, “loop engineering” is replacing “harness engineering” but is radically more expensive in tokens. Which means you can go bankrupt in a quarter if your financial controls are sloppy and your business doesn’t have the upside potential to benefit from it. An emerging promising (and token-cheap!) orientation is stigmergic AI/substrate engineering. At a more retail level, individuals in consumer-usage herds are getting things really really wrong or really really tbd. For eg I don’t know if my approach of pure vibe coding (never look at source) but with obsessive attention to raw data integrity is “really really right” but it’s at least “really, really tbd” (ie it is an unhedged opinionated style). But it’s obvious by now that any approach that involves reviewing every source file is very very wrong. Whatever the really, really right approach ends up being, it will involve only rarely looking at source. In writing it’s even clearer, but substack is not ready to hear it. Footnote: We’re in a rare era where “get AI really really right” is an economy-wide imperative, not just for SV-flavored startup-land. That’s a once-in-a-century type deal.

2 replies

2 recasts

8 reactions