What if the best move today is to wait?
Wait on the build, because the next model will absorb it. Do not wait on trust, on data, or on deciding what to make.

Wait on the build, because the next model will absorb it. Do not wait on trust, on data, or on deciding what to make.

Each new model does more with less: what took the last one its maximum effort, the next one does at default for a fraction of the cost. If that rate holds, anything that demands heavy effort today will take a fraction of it soon, and may come out better. Impact moves the same way: what would look disruptive today will be ordinary a release later. So is doing nothing now sometimes the stronger move?
Each release hits only the left layer. The loop runs through the right one: the seed regenerates the build, the build sells, and sales feed the seed.
The charts keep two kinds of comparison apart.
The pattern holds outside benchmarks. The first formalization of Fermat’s Last Theorem took 11 days with a purpose-built harness and human guidance. In a different-tier comparison, a team of Fable 5.2 agents did it in 30 hours with a proof seven times shorter, while Opus 5.5, the lower tier, in the same setup stalled at about 26%.
If the next model will do it with more impact, for a fraction of the effort, what does doing it now buy you?
Take any project that demands months of committed effort today. If capability keeps improving at the rate shown above, the same project costs a fraction of that effort a few releases from now. The result may also be better than anything refined by hand in the meantime.
Under that condition, effort committed today is not an investment. It is the price of arriving early, and whatever it builds is outdone by the next model starting from a short description.
Impact follows the same curve. The awe a build can earn is capped by the model it was built on, and that cap rises with every release. What reads as disruptive today is the default a release later, and the same idea built then can reach further than anything possible now.
That leaves an uncomfortable possibility. For work that is heavy on effort or built to impress, doing nothing now can beat doing it at all costs. An idea held until the right model exists may be worth more than the same idea forced through a weaker one.
The model does not arrive for everyone on the same day. Mythos went to a small set of vetted organizations first, and that head start lasts only until a comparable model is public. What does not arrive with it is knowing what to build, the data and context only you have, and people who trust you.
So the useful version of the question is narrower. Which part of today’s effort will the next model do for free, and which part will it never do for you?
Each lane trades effort against position. Which one wins depends on how fast the next model arrives and how much of the work it absorbs.
Follow the trend to its end. If building costs almost nothing, having built something proves almost nothing. Anyone can have the same thing the day the model ships. Awe goes the same way: once anyone can produce the impressive thing, it stops impressing.
Effort used to be the filter. It decided who finished, and finishing was the signal. When effort disappears, the filter moves to what stays scarce: choosing what deserves to exist, and getting someone to care.
So the deeper question is not when to build. When doing becomes free, what is still worth doing?
Now scale the question up to an industry. Every team gets the same capabilities, and whatever one builds with them, another rebuilds weeks later for almost nothing. What is left to compete on?
Only what a release cannot hand out. Trust, reputation, relationships and data accumulate at human speed, so no new model delivers them to a competitor overnight. Everything replicable drifts toward free, and the surplus moves to whoever owns the model or the customer.
Then there is the choice itself. A model answers; it does not start. In The only missing element keeping us different from AI I called this the endogenous trigger: acting with no prompt. When any answer is cheap, the scarce act is deciding, unasked, which question deserves one.
That may also settle the title. Wait on the build, because the next model will absorb it. Do not wait on trust, on data, or on deciding what to make, because no later model refunds that time.
All of this holds only as long as models keep improving at the rate the last releases show.
The core is the wait calculation plus vintage capital, with a Baumol-style bottleneck at the end. References are from recall and should be checked before citing.
| Rank | Concept | How it maps | Reference |
|---|
| 1 | Wait calculation | Later, faster launches overtake earlier ones. Applies only when the asset cannot be refitted in flight. | Kennedy 2006, JBIS 59:239–246 |
| 2 | Technological expectations | Expecting faster improvement delays adoption of the current technology. | Rosenberg 1976, Econ. J. 86 |
| 3 | Vintage capital | Productivity is embodied in the vintage. Maps the depreciating layer. | Solow 1960; Johansen 1959 |
| 4 | Real options | Wait only while option value exceeds foregone cash flow and learning. | McDonald and Siegel 1986, QJE 101(4) |
| 5 | Baumol bottleneck | The non-automated input sets the pace. Maps the demand bottleneck. | Aghion, Jones and Jones 2017, NBER w23928 |
| 6 | The Bitter Lesson | Hand refinement is dominated by general methods riding capability growth. | Sutton 2019 |
| 7 | Seed plus decoder | Product = short program run on a universal decoder. A better decoder shortens the seed. | Kolmogorov 1965; Solomonoff 1964 |
| 8 | Complementary assets | When technology is easy to imitate, profit goes to whoever owns distribution. | Teece 1986, Research Policy 15(6) |
| 9 | Preemption games | If everyone waits, someone preempts and rents equalize. | Fudenberg and Tirole 1985, Rev. Econ. Stud. 52 |
| 10 | Increasing returns | Users, data and network effects compound regardless of code vintage. | Arthur 1989, Econ. J. 99 |