4 ms·

No Shortcuts To Knowledge: AI Needs To Ease Up On Scaling And Learn How To Code

dane-pgp 4y ago
> If skills are programs, and intelligence is the ability to efficiently learn skills, then intelligence is, broadly speaking, programming ability. I can see that definition becoming very popular among certain groups of people.
- inphovore 4y ago
  Intelligence is the mitigation of uncertainty. If it does not mitigate uncertainty it is not intelligence.
  - dane-pgp 4y ago
    I'm not sure if "mitigation of uncertainty" is a concept that is any more rigorously defined than "intelligence", so although it might be helpful to link the two, it still leaves a lot of questions unanswered. To give an example of how I think "mitigation of uncertainty" differs from "intelligence", consider a primitive chess computer that just looks ahead a couple of moves. You could say that its algorithm is "mitigating the uncertainty" of what move its opponent will play next, but this sort of simplistic program seems mechanical with no signs of learning or creativity. In contrast, consider a hypothetical AI that can read all the existing mathematical literature and develop new conjectures, and even prove them. I wouldn't say that such an AI was mitigating any uncertainty, but it would be doing something that for humans would require a lot of intelligence.
    - inphovore 4y ago
      Potential is the domain of uncertainty. Potential collapses as state through resolve. Mitigation is any negotiation between potential and resolve. (Naturally potential resolves to some state, whatever it may be, intelligence is altering disposition such that the state may be directed or discerned. > To give an example That is how it is the same, not different. Learning is further dispositioning, and there are presumably infinite variations of how this may be so, thus learning to learn is an intelligence, and so is learning chess, and so is learning to learn chess, and so on. Creativity is something else which couples with intelligence. Achieving novel disposition? Creative includes artful failures of intelligence. We shouldn’t put the burden of every expectation upon intelligence. > In contrast “Reading all existing” is removing uncertainty (a definition of INFORMATION[1], though I know you will not let me rewrite modern information theory in this short post.) “develop new” is certainly mitigation as what would be their purpose? And PROVING them is exactly mitigation. These are merely folds in a recursive process. The difference between modern AI and True Intelligence is that modern ML is feature programmed statical automation. Indexing brute force variations; choosing the highest statistical match, with a splash of random for character. It is the programmer who was intelligent, the AI is a statistical automation tool. [1] Information is the removal of uncertainty, if it does not remove uncertainty (distributed potential toward resolve) it is not information. In this view, state is certainty, and in abstract non-terminal domains true certainty is impossible, one may only be less uncertain and make the best weighted selection possible. > it still leaves a lot of questions unanswered No doubt!

> If skills are programs, and intelligence is the ability to efficiently learn skills, then intelligence is, broadly speaking, programming ability. I can see that definition becoming very popular among certain groups of people.

inphovore 4y ago
Intelligence is the mitigation of uncertainty. If it does not mitigate uncertainty it is not intelligence.
- dane-pgp 4y ago
  I'm not sure if "mitigation of uncertainty" is a concept that is any more rigorously defined than "intelligence", so although it might be helpful to link the two, it still leaves a lot of questions unanswered. To give an example of how I think "mitigation of uncertainty" differs from "intelligence", consider a primitive chess computer that just looks ahead a couple of moves. You could say that its algorithm is "mitigating the uncertainty" of what move its opponent will play next, but this sort of simplistic program seems mechanical with no signs of learning or creativity. In contrast, consider a hypothetical AI that can read all the existing mathematical literature and develop new conjectures, and even prove them. I wouldn't say that such an AI was mitigating any uncertainty, but it would be doing something that for humans would require a lot of intelligence.
  - inphovore 4y ago
    Potential is the domain of uncertainty. Potential collapses as state through resolve. Mitigation is any negotiation between potential and resolve. (Naturally potential resolves to some state, whatever it may be, intelligence is altering disposition such that the state may be directed or discerned. > To give an example That is how it is the same, not different. Learning is further dispositioning, and there are presumably infinite variations of how this may be so, thus learning to learn is an intelligence, and so is learning chess, and so is learning to learn chess, and so on. Creativity is something else which couples with intelligence. Achieving novel disposition? Creative includes artful failures of intelligence. We shouldn’t put the burden of every expectation upon intelligence. > In contrast “Reading all existing” is removing uncertainty (a definition of INFORMATION[1], though I know you will not let me rewrite modern information theory in this short post.) “develop new” is certainly mitigation as what would be their purpose? And PROVING them is exactly mitigation. These are merely folds in a recursive process. The difference between modern AI and True Intelligence is that modern ML is feature programmed statical automation. Indexing brute force variations; choosing the highest statistical match, with a splash of random for character. It is the programmer who was intelligent, the AI is a statistical automation tool. [1] Information is the removal of uncertainty, if it does not remove uncertainty (distributed potential toward resolve) it is not information. In this view, state is certainty, and in abstract non-terminal domains true certainty is impossible, one may only be less uncertain and make the best weighted selection possible. > it still leaves a lot of questions unanswered No doubt!