Content
@
https://warpcast.com/~/channel/phil
0 reply
0 recast
0 reaction
Phil Cockfield
@pjc
"prefer duplication over wrong abstraction." truth. timestamp: https://www.youtube.com/watch?v=8kMaTybvDUw&t=889s
3 replies
0 recast
7 reactions
Metaphorical
@hyp
Abstraction is everything!
1 reply
0 recast
1 reaction
Metaphorical
@hyp
(I mean, literally, name something that is not an abstraction, you can't!)
1 reply
0 recast
0 reaction
Phil Cockfield
@pjc
Right!! You can’t spell “abstraction” without abstractions. The pursuit of “good*” abstraction, though, there’s the rub! *powerful
1 reply
0 recast
1 reaction
Metaphorical
@hyp
Yes, actually related to some work I'm doing. It's about predictive power in compression. what's the most compressed you can make an abstaction and have it still be nearly deterministic. What would the equation (a useful abstraction) of compression to determinism look like?
2 replies
0 recast
1 reaction
Phil Cockfield
@pjc
Crazy interesting @hyp !
1 reply
0 recast
1 reaction
Metaphorical
@hyp
Talks with Claude (not Shannon) There are several well-established measures in the research literature that quantify the relationship between compression/abstraction level and predictive power or determinism. Here are the key ones that would be highly relevant to your project: ## **1. Effective Information (EI) - Erik Hoel** This is perhaps the most directly relevant to your question. EI measures causal power by quantifying how much an intervention constrains future states. It tracks the trade-off between compression (moving to higher scales) and causal determinism. The measure quantifies how well a system's mechanisms constrain its past and future states, and can be higher at macro scales than micro scales when abstraction increases causal power. ...
1 reply
0 recast
1 reaction
Metaphorical
@hyp
## **2. Minimum Description Length (MDL) Principle** MDL is a model selection principle where the shortest description of the data is the best model. It balances compression (model simplicity) with predictive accuracy. The principle holds that the best explanation is the one that permits the greatest compression of the data while maintaining predictive power. This is exactly the E=mc² type relationship you're describing. ## **3. AIC/BIC Information Criteria** AIC estimates the quality of statistical models by dealing with the trade-off between goodness of fit and model simplicity. BIC attempts to resolve overfitting by introducing a penalty term for model complexity, generally penalizing parameters more strongly than AIC. Both measure the compression-prediction trade-off quantitatively. For a start.
1 reply
0 recast
1 reaction
