Spandrell (0:00): It really got me thinking about loyalty and how loyalty works. Back to Chateau Heartiste — this was a quip he would say all the time. He was like, loyalty only goes as far as your options, right? Women are only loyal as far as their options. The more options they have, the less loyal they will be. And men should be the same way. And I was like, you know what, yeah, that's exactly how it works. People are loyal depending on the options they have. Why were proletarians loyal to [bleeped] Soviet Russia, with all its misery and depression? Because they had nothing else to do. They had no other options.

Lukas (0:33): We might have actually — dude, I think we just cracked it, Saila. We've had this ongoing thesis on the podcast, basically that an increased efficiency of markets tends to destroy humans' enjoyment of the market. And we know that this is true observably, right? If you have a dating market that is hyper-efficient, you can go on Tinder and swipe and find eight million people who are plus or minus two percent of whatever your algo ranking is — and everyone instantly wants to kill themselves. And this is the explanation for that, right? If optionality is the inverse of loyalty, we've now basically removed loyalty from society. It can't exist if everyone has enough options. If there's a system that provides everybody with both the workplace liquidity and the romantic liquidity that they need for everything, then no one is loyal to anything. Yes — which means you can never trust anything. If you can't be loyal, you can't trust anything. So everything is a rug now.

Saila (1:33): Well, optionality is like anything else. There should be more than one woman in your village, right? That way, if the only one who's there has some massive problem, you can go marry someone else. But if there are ten million women in your proverbial "village" — your actual realistic dating pool — it is now incredibly difficult to. Right. It's like the secretary problem, where there's some optimal search. Obviously this is very theoretical, but there's an optimal percentage of the sample that you go through in order to find the best secretary.

This optionality point is obviously true as it applies to the subjects they talked about, but it's also true now that we're faced with the existence of truly competent AI itself. I just mentioned it last message, I can now do any 3D game I want! That's endless optionality and it makes it that much harder to commit to any one game, knowing that I can realistically do anything. The Optionality Paralysis point by Justin becomes ever more relevant. Saila mentions the secretary problem, so I asked Claude for some actual math on my own situation:

If my goal is making a game every 6 months, let's say, then how would I approach it? Update myself on the tools and sample the environment for ideas for like 1 month and then commit? Or something else?

Your guess is nearly exactly what the model spits out, which is worth seeing derived, because the derivation also tells you which errors are cheap and which are catastrophic — and that part is more useful than the optimum itself.

The model

A slot is T = 26 weeks. You split it: S weeks sampling (prototyping candidate ideas), 26 − S weeks executing the winner. Two ingredients:

Idea quality from sampling. At your speed a tested prototype — core loop playable, fun-gate verdict — is 2–3 days; KVP went from ABAgames catalog entry to shipped game in about a week, so call it r = 2 tested prototypes/week, conservatively. If each idea is an independent draw from your personal idea distribution, and we measure quality as the quantile q within that distribution (dodges all assumptions about the distribution's shape), then the expected best of n draws is:

$$E[q_{best}] = \frac{n}{n+1}, \qquad n = rS$$

One prototype → 50th percentile on average. Nine → 90th. Nineteen → 95th. Note the shape: violently diminishing. Going from 1 to 9 prototypes buys 40 percentile points; going from 9 to 19 buys five.

Value from execution. Your own thesis (the long tail of details is most of the job) says polish time has real returns, but concave ones — the fifth month of polish adds less than the second. Standard concavity: polish multiplier = (26 − S)^β with β = 0.5.

Multiplicative value (a bad idea polished forever stays capped, a great idea unpolished stays capped — this is the fun-gate worldview in equation form):

$$V(S) = \frac{rS}{rS+1}\,(26-S)^{\beta}$$

Solving it

Set V'(S) = 0. The algebra collapses to:

$$26 = S\left[1 + \beta(rS+1)\right]$$

With r = 2, β = 0.5: S² + 1.5S − 26 = 0 → S ≈ 4.4 weeks, i.e. ~9 prototypes, committing to roughly your 90th-percentile idea, with 21.6 weeks left to execute.

So: yes — one month of sampling, five of building. Your instinct priced it correctly.

Sensitivity (same equation, different parameters):

Everything reasonable lands in 3.5–6 weeks. Also note what it isn't: the time-translated 37% rule would say 9.6 weeks, and it's beaten by every row — because 37% optimizes "probability of the single best," and you're optimizing expected value with polish time as the opportunity cost. Real payoffs cut exploration roughly in half.

The part that actually dissolves paralysis

Evaluate V(S) across the range (base parameters, V(4.4) = 4.17 = 100%):

The curve is an inverted U with a wide flat top and steep cliffs. Anywhere from 2 to 10 weeks costs you under 10%. The two disasters are the extremes: committing blind to the first idea (no calibration, −40%) and sampling half the slot away (no polish, −43%). Which means the thing your instinct wants to agonize over — exactly when to stop, is week 4 or week 6 right — is worth almost nothing. The only decisions with real money in them are "don't take the first idea" and "don't still be browsing in month three." This is the quantitative version of "the feeling of settling carries no information": near the top, all stopping points are within noise of each other.

Turning it into a policy

The fixed-S plan is already near-optimal, but a threshold rule strictly beats it and costs nothing:

1. Weeks 1–5, sample — and let tool-updating BE the sampling. Don't do a tools month then an ideas month; a prototype is how you find out what the tools can now do. Your last month is the proof: Skyland was simultaneously "can the engine do a vista" and a shipped artifact; Robots was simultaneously "can Claude do machines" and the menagerie that robot-punch-out now stands on. One activity, two outputs.
2. Early exit: if something clears your remembered bar from all previous exploration — the "can't stop thinking about it" signal — commit immediately, even in week 2. With your 15 years of calibration, your prior threshold is already meaningful; the model's n/(n+1) treats you as starting from zero, so it understates how early you can safely stop.
3. Hard stop at week 5: take the best thing seen, no extensions, no "one more idea." The table above is why the hard stop is safe: the downside of stopping slightly early is a few percent; the downside of the exception-that-extends is the right-hand cliff.

The level above, which is doing the real work

You're not playing this game once — at 6 months per game, you have on the order of 40–60 slots left. Each shipped game is itself one draw at the level that actually matters (which directions are worth your decade), and best-of-40 at that level gets you to ~97th percentile of your own possibility space regardless of per-slot optimality. The across-slot search happens automatically as long as slots keep closing; it stops entirely when one stays open. That's the cleanest statement of why per-slot "settling" is cheap and per-slot browsing is the only fatal move: a mediocre committed slot still feeds the portfolio search, and — your Chibi → Robots → punch-out chain being the live example — even the failed object produces the calibration and pipeline that the next slot spends. The one thing the portfolio can't recover from is slots that never close.

Numbers are illustrative, of course — but the conclusion isn't fragile. It's driven by two structural facts, not the constants: best-of-n saturates fast (n/(n+1)), and polish is concave. Any parameters with those two shapes land at "explore ~15–20% of the slot, and precision doesn't matter." One real month, nine real prototypes, commit, close the slot.

I thought this argument was interesting, some of the math goes above my head and I asked him to explain it to me as he might to a young child, or a golden retriever, it wasn't brains that got me here I can assure you that, and, you know, he makes some assumptions here and there that I'm not so sure about but it all seems roughly right. In general, when dealing with LLM output like this, it's useful to not take it too literally (if the details don't matter that much) but use the exploration to get some interesting insights, like the idea that doing ~10 prototypes gets you 90% of the way there in terms of exploration, which feels intuitively correct, and that polishing mattering less and payoffs being hit-driven both increase the sampling time, which does not feel intuitively correct to me but the logic seems sound. In any case, the optionality problem is more real now than before but I think it's solvable.