[FRIAM] To repeat is rational, but to wander is transcendent

[Marcus Daniels]

On 3/30/22 10:28, Marcus Daniels wrote:
> The whole motive of adopting the reference implementation -- being lazy -- locked me in to a certain performance for the solver.   I would expect the same sort of thing would happen with inheritance or horizontal gene transfer.    I could start a divide and conquer search (as you say with last common ancestor nodes), but I would never be confident in what I had if I did that.    It would probably take as long as starting over to gain that confidence.    The foundationalist view, if I understand what you mean, is that there are ideal ways to do this and that the melting and freezing of abstractions could find them once and for all.

Glen wrote:

< Hm. I'd have thought your motivation to be execute and compare against the reference implementation, not to adopt it. To me, it's a validation of your solver that, in the lower DoF condition, you got the same performance. So re: foundationalist, yes, you've understand my usage right. "Ideal" is a little ambiguous in that a foundationalist can allow for multiple "equivalent" foundations. E.g. let's say you have two foundations that span the same space. But in 1 foundation, an end point is reached with more steps/hops than the other one. If "ideal" means shortest path from origin to end point, then the other is more ideal. But if it means something else, like "stays close to some other, more ubiquitous foundation", then maybe not. Jon made the comment that he rejects all the old proposition-style foundations as wrong-headed. But culturally, because many of us are so entrenched, such may be more tractable (were they to actually *work* as foundations, of course). >

The so-called "reference model" (implementation of the model) in this context could be optimized to profit X, whereas with more DoF (that the reference model could not expose), one could reach profit X + Y where Y was positive.    No solver could get the reference model to a X +Y profit.    The term "reference model" to me would suggest to me that profit X + Y was possible with a sufficiently good solver.