The objection Tan takes head-on is that models improve so fast that harness and context work will be obsolete — just wait for the next release. His counter is that the observed direction is the opposite: “the better the models get, the more the differentiator moves to context. When everyone’s engine is a thousand horsepower, the race is won on the driver and the map.” The weights are everyone’s; the library is yours. And the compounding runs the right way, because “a better model makes your library worth more because a smarter reader extracts more from the same books” — every lab release is a free upgrade to a workforce you already own.
This is in direct tension with Scaffolding is tech debt against the next model — the bitter lesson applied to product building, and the tension is worth keeping rather than resolving. The discriminator is what the artifact is made of. Scaffolding is code written to patch a capability the model lacks — it is a bet against the next release and dies when the release lands. A library is accumulated private context — it bets on nothing about the model and appreciates when the reader gets smarter. Both can be true simultaneously, and the practical test is to ask of any artifact whether it would be deleted or made more valuable by a model twice as capable. Tan’s related objection-two answer is Retrieval is the primitive, not the product — being worth retrieving from is the product, and the asset he is describing is the one defined in Personal AGI is an asset you build, not a subscription you rent. Note also the tension with Deliberately overspending on inference buys lead time, not efficiency from his own earlier interview: there the lead is temporary and gets competed away, here the library’s advantage is claimed to widen.