[agi] MindForth 15.JAN.2008
Mind.Forth Programming Journal (MFPJ) Tues.15.JAN.2008 Yesterday on 14 January 2008 the basic scaffolding for the Moving Wave Algorithm of artificial intelligence was installed in Mind.Forth and released on the Web. Now it is time to clean up the code a little and to deal with some stray activations that interfere with the proper generation of meandering streams of thought. First in psiDamp we are re-introducing the single call to psiDecay, so that post-thought "lopsi" concepts will gradually lose their activation over time. We need to get rid of the "newpsi" and "prequel" and "psicrest" variables, because with "lopsi" and "hipsi" we were able to get the job done in exemplary fashion. Having eliminated or commented out the obsolete variables, now we are trying to debug the stray activations. When we enter a known word like "kids" or "robots" and we press [RETURN], we get a line of output such as follows. Robot: ROBOTS WHAT DO ROBOTS DO The AI speaks the word "ROBOTS" because it is starting an SVO sentence with "ROBOTS" as the subject. At first, the activation of ROBOTS sends a "spike" of twenty (20) to the verb NEED -- which has no subconscious activation because it exists in enBoot and not as a recent thought. ROBOTS #39 w. act 48 at I = 186 sending spike 20 to seq #74 NEED at act 0 yields 20 and zone = 181 20 (lim = 63) for t=183 NEED engram; spike = 20 R The enBoot verb "NEED" gets rejected with a message-line. verbPhr: detour because verb-activation is only 12 We see from the following diagnostic output display that the Audition module has been calling psiDamp to de-activate the ROBOTS concept after the hearing of each individual letter in the ROBOTS word. R psiDamp called for urpsi = 39 by module ID #104 Audition psiDecay called to reduce all conceptual activations. O psiDamp called for urpsi = 39 by module ID #104 Audition psiDecay called to reduce all conceptual activations. B psiDamp called for urpsi = 39 by module ID #104 Audition psiDecay called to reduce all conceptual activations. O psiDamp called for urpsi = 39 by module ID #104 Audition psiDecay called to reduce all conceptual activations. T psiDamp called for urpsi = 39 by module ID #104 Audition psiDecay called to reduce all conceptual activations. S psiDamp called for urpsi = 39 by module ID #104 Audition psiDecay called to reduce all conceptual activations. psiDamp called for urpsi = 39 by module ID #104 Audition psiDecay called to reduce all conceptual activations. psiDamp called for urpsi = 39 by module ID #104 Audition psiDecay called to reduce all conceptual activations. Well, isn't that result weird? By briefly changing the "module ID #" above to "42" for external input and to "35" for internal flow, we discovered multiple psiDamp calls during the internal reentry of each word being thought. R psiDamp called for urpsi = 39 by module ID #35 psiDecay called to reduce all conceptual activations. O psiDamp called for urpsi = 39 by module ID #35 psiDecay called to reduce all conceptual activations. B psiDamp called for urpsi = 39 by module ID #35 psiDecay called to reduce all conceptual activations. O psiDamp called for urpsi = 39 by module ID #35 psiDecay called to reduce all conceptual activations. T psiDamp called for urpsi = 39 by module ID #35 psiDecay called to reduce all conceptual activations. S psiDamp called for urpsi = 39 by module ID #35 psiDecay called to reduce all conceptual activations. psiDamp called for urpsi = 39 by module ID #35 psiDecay called to reduce all conceptual activations. The concept of ROBOTS keeps being set to the same "residuum" by psiDamp, but the NEED concept keeps getting psi-decayed until its activation drops too low for validation as a good verb to go with the word ROBOTS as a subject. We vaguely suspect that the reentry of each character in ROBOTS is being treated as if the character were a whole word by itself, so that the "trough" triggger code gets activated not merely once, but many times. Aha! During reentry, the SPEECH module is setting "pov" to "35" and calling AUDITION for each character being "pronounced" by the SPEECH module. Therefore the Audition "trough" trigger is being set to one (1) for each and every character being reentered from the SPEECH mind-module. There should be some easy fix for this bug, such as perhaps creating a special flag to indicate that reentry is in progress. However, at this point we would like to remark that, after the extremely difficult lopsi/hipsi coding of yesterday, we may finally be in the close-to-True-AI phase where the major bugs have been solved and we are only clearing out minor bugs -- which nevertheless prevent the AI from functioning flawlessly as True AI. We had better check the table of variables and the Jav
Re: [agi] AGI and Deity
James, your comments are appreciated. a few comments below Stan James Ratcliff wrote: Your train of reasoning is lacking somewhat in many areas, and does not directly point to your main assertion. thanks for the feedback. As I follow other discussions and read the papers they refer to, I realize that my writings are lacking. Perhaps they are more blog like than scientific. The problem of calculating values of certain states is a difficult one, and one that a goo AGI MUST be able to do, using facts of the world, and subjective beliefs and measures as well. I'm not sure I get the MUST part. Is this for troubleshooting purpose or for trust issues? Or is it required for "steering" the contemplation or attention of the machine? Whether healthcare or education spending is most beneficial must be calculated, and compared against eachother, based on facts, beliefs, past data and statistics, and trial and error. And these subjective beliefs are ever changing and cyclical. As a better example, would be a limited AGI whose job was to balance the National budget, its job would be to choose the best projects to spend money on. Maximizing Benefit Units (BU) here as a measure of 'worth' of each project is needed and required. One intelligence (human) may be overwelmed with the sheer amount of data and statistics to come to the best decision. An AGI with subjective beliefs about the benefit of of each could use potentially more of teh data to come to a more maximized solution. It is the "future scenarios" that are often the most compelling justification or "evidence" for value of something, and in my opinion the most unreliable. Whether it is man or machine giving his case, there will be speculation involved in the common sense domain. Will the scenario be "You say this... now prove it. If you can't prove it don't use that in the justification..." Very limiting. On your other note about any explanation being too long or too complicated to understand.. Any decision must be able to be explained. It can be done at different levels, and expanded as much as the AGI is told to do so, but there should be NO decicions that you ask the machine, Why do you decide X? and the answer is nothing, or 'I dont know' If the architecture of the machine is "flow" based, that is, the prior events helped determine current events, then the burden of explaining would overwhelm the system. Even if only logic based, as you pointed out the "values" will be dynamic and to explain one would need to keep a record of the values that went into the decision process - a snapshot of the "world" as it was at the time. What if the system attempted to explain and finally concluded "if I were making the decision right now, it would be different." We wouldn't consider it especially brilliant since we hear it all the time. Any machine we create that has answers without the reasoning, is very scary. and maybe more than scary if it is optimized to offer reasoning that people will buy, especially the line "trust me." James Ratcliff */Stan Nilsen <[EMAIL PROTECTED]>/* wrote: Greetings Samantha, I'll not bother with detailed explanations since they are easily dismissed with a hand wave and categorization of irrelevant. For anyone who might be interested in the question of: Why wouldn't a super intelligence be better able to explain the aspects of reality? (assuming the point is providing explanation for choices.) I've placed an example case online at http://www.footnotestrongai.com/examples/bebillg.html It's an "exploration" based on becoming Bill Gates, (at least having control over his money) and how a supercomputer might offer "explanations" given the situation. Pretty painless, easy read. I find the values based nature of our world highly relevant to the concept of an emerging "super brain" that will make super decisions. Stan Nilsen Samantha Atkins wrote: > > On Dec 26, 2007, at 7:21 AM, Stan Nilsen wrote: > >> Samantha Atkins wrote: >>> >> >>> In what way? The limits of human probability computation to form >>> accurate opinions are rather well documented. Why wouldn't a mind >>> that could compute millions of times more quickly and with far >>> greater accuracy be able to form much more complex models that were >>> far better at predicting future events and explaining those aspects >>> of reality with are its inputs? Again we need to get beyond the >>> [likely religion instilled] notion that only "absolute knowledge" is >>> real (or "super") knowledge. >> >> Allow me to address what I think the questions are (I'll paraphrase): >> >> Q1. in what way are we going to be "short" of super intelligence? >> >> resp: The simple answer is that the most intelligent of future >> intelligences will