prev
000 of 000
next
FILE: REVI: - [20_22/12/31;15:16:06.00]:. - collect and draft compose. TITL: Proof Irrationality By Forrest Landry, Dec 31st, 2022. ABST: How some notions of rationality are irrational. TEXT: The main starting crux seems to be: If people are asserting/believing/arguing that "anything is possible", that no knowledge can be complete, or 100% certain, that some maybe future observation could change everything, then how is that _not_ equivalent to asserting that there are _no_limits_ to what can be done with causation, algorithms, reason, logic and math, etc? How is that not a directly rejection of the notion of reason itself? Ie, how does anyone working in 'AI alignment' account for the already known/proven limits to causation, to algorithms, to what is computeable, with imperfect information, etc, such as with the Halting Problem, the Rice Theorem, or the mere fact of micro-state amplification in non-linear systems (the three body problem) along with well verified Heisenberg Uncertainty, etc, _not_ result in perfected predictability problems? How is the notion of "anything is possible and nothing is perfectly knowable" _not_ a rejection of the notion of symmetry -- and therefore also implicitly and tacitly a rejection of _both_ reason and physical lawfulness itself? Is there _any_ way to assert that the notion of "reason" and "rationality" is coherent, _without_also_ implying a completed acceptance of symmetry as an absolute, and therefore, also accepting notions like causation, and some actual real limits to what can be done with only causation, as well as a notion of physical lawfulness, as also absolute? This is something of a core epistemic crux and question, beyond which no point of "proof" or notion of rigor or reason can obtain, unless this is actually resolved first. Near as I can tell, _every_ notion of reason, logic, and knowability, to the point of completed certainty, is going to depend on whether the notion of proof and symmetry are fully bijective in 'truthiness'. Frankly, it is rather shocking to me to even have to mention these sorts of issues and ask these questions. However, somehow, it seems that in the ratified forum of "AI Alignment Research", that these root notions, core questions regarding actual process symmetry and completed knowability (100% confidence) are somehow disjunctive? What notion of 'sense' is even available to make sense otherwise? A belief in "Bayesianism" does not even begin to cover this, for even the notion of object permanence (and perceptual world truth consistency) cannot be assumed _without_ there also being at least some notion of symmetry as a completed 100% accepted prior. ~ ~ ~ People keep asking me to "let the content do the talking". People somehow seem to think that I am so naive as to not know and notice that the social context in which any content (of any type) is presented is at least just as important as the content itself, if not more so. Would an advertisement suggesting that some mother purchase X brand baby food have the same meaning if it occurred next to an article that had the title "Criminal Gang uses Baby Food business as a money laundering scheme"? Belief that content can occur without context lets them implicitly control the context. It is just another social power move: No good. Admittedly, no one is ever going to "win" (ie change) these kinds of social power physics/heuristics/aspects. Personally, I am not really that interested to try, at least not in the AF/LW/EA/HN type "rationalist" groups. The selective blindness and bias associated there (around technology optimism and philosophical modernism) is rather too strong for most people to notice the taint of the waters that they are swimming in -- let alone even to notice the water as being water. We/they have already done as much as is likely to be possible to be done, easily and realistically. We will need to shift to the 'epistemic' side of things. Usually that means starting with definitions. However, given that adverse listening will every time attempt to pick apart every aspect of every definition, we simply cannot just present those definitions in plain text. There is simply too much inherent bias/unreasonableness in that process, and so we must start with coherence around even more basic epistemic technique, to see if we can at least cleanly separate the social political aspects from the rationality ones. Ie, to at least implicitly and consistently separate conflict theory participants from mistake theory oriented collaborators and communicators. As such, we will do better to ask the kinds of questions that will allow the LW/EA/AF community to obtain the required forms of the needed definitions, and methods of reasoning, themselves, on their own, by asking questions that converge on closed forms, and then use _those_ resulting definitions, the ones that they seemingly came up with, and this assent to, because they have actually worked out the reason and reasonableness themselves. We simply cannot do the 'understanding' and 'comprehension' for them, it is impossible. Hence, for the result to become unarguable, it must be based on their own assertions -- _if_ they are able to even consistently have any. However, that in itself means that we must maintain at least some control of the context. This means *not* holding all of the debate on just their forum, on their terms, etc, which also means _not_ replying to every remark that they make. It simply does not help. Instead, we can move to directly to inquiry, and whenever possible, ensure that the context of the conversation is more on a one on one basis (direct messaging) whenever possible. My neurotype simply cannot handle trying to track too many other people/minds at once, and not everyone is equally worth whatever is my remaining time of life, by my choice. ~ ~ ~ One question: Is there any way -- any actually important aspect -- in which AGI is _not_ being described (hyped) as being exactly like a 'perpetual benefit machine', (for whomever is the manufacturer/owner of the AGI)?. How/where/when is AGI is _not_ being described as "a machine that can do anything that is doable, make anything that is makable, and/or solve any problem that is actually physically solvable, in this real universe"?. How does AGI _not_ eventually become equivalent to an an assumption of limitless intelligence? It seems to me that there can very easily be found a very long list of links from LW/EA, along with selected very sort quotes for each one that makes it very explicitly clear that people *are* describing AGI as some sort of "everything will (probably) be better once we have AGI (if we can align it 1st)". Another question: Can there be claimed -- is anyone at all attempting to claim -- that there is *any* method at all of 'AGI Alignment' that is _not_ based on _just_ causation and modeling, either directly or indirectly?. Ie, directly through some form of algorithm (a causative model of a system) that acts as an alignment error correction, or indirectly, via assuming some sort of economic incentive. Is there any sense at all -- any process by which -- in which an AGI could be "aligned" which does not depend on only and exactly causation? Ie, is there any real or serious methodology of alignment that can validly, actually, and effectively depend on something other than causation? Maybe people will quibble on the notion of "what is causation?". To keep it maximally simple, 'causation' is only "a persistent/consistent pattern of observation where we/everyone notice(s) that 1st 'Q' happens, and then 'P' happens". The term 'causation' is in reference to the implied lasting symmetry of perceptual patterns (content) over/by many people and situations (context). Symmetry itself is a 'sameness of content where there is a difference of context'. For example, it does _not_ seem reasonable to assume that an AGI will somehow 'magically know' what 'our best intentions' are so as to 'do the right thing (to our benefit)'. For the AI machine to "know" it must have some process, and given that "machine" implies repeatable process, that the notions of symmetry and causation both apply. Even the notion of algorithm and even of logic itself imply some equivalence to the notion of 'causation' as defined herein. Anything done in with machines, and therefore of engineering, either directly or indirectly, is always about causation (and not, for example, magic). Hence we can also ask: is there any _reasonable_ proposal for AGI alignment that have or consist of _any_ aspects that are _not_ strictly causal, in some form or another? Of course, people will either debate or ignore these question. If they ignore the question(s), then the result is an admitted tacit acceptance that the implied truth of the question is accepted. Ie, that it is allowed that the notion of causation, and of symmetry, apply to machines, regardless of their type (neural nets, however modeled, or not). In regards to setting up for a future proof, we can also ask: Is there any form of logic or of reasoning or of 'reasonable argument' that does not in some way, directly or indirectly, depend on the notion of symmetry?. Ie, can there be any allowed form of argument or any 'valid method of reasoning or proof', that does not depend on, which can somehow also violate the notion of symmetry, with impunity, and which is still considered valid/acceptable?. If at least some arguments/proofs/results depend only and exactly on just the notion of symmetry, will _not_ therefore also result directly or indirectly, in something being actually "100% knowable"? It seems to me that if we reject the idea that anything at all is 100% knowable, that we have also, at least implicitly, though also actually, rejected the idea of symmetry, and therefore also that any form of reliable trustworthy reasoning is actually possible. Where exactly nothing is fully knowable, then not even symmetry is an allowed assumption, and where any and every form of knowing itself depends on symmetry, then how can we even know that nothing is knowable, as a special case? If it is regarded/agreed that the notion of 'valid reasoning' and 'violates symmetry' are fully mutually exclusive (explicitly contradictory), then it is also the case that anything, any statement 'P' that inherently contradicts symmetry, (directly through false syllogism or indirectly through incorrect and non-relevant definitions) can be considered to result in 'a fully knowable truth' that such a thing ('P') is 100% known to be fully/actually impossible, in the strict/correct/absolute sense of the terms. How else will we will obtain at least some usable notion of what is meant by "proof" and "knowable"(?), or what is the notion of "reason" and "reasonable"(?) that is epistemically sound, unarguable, rigorous, etc, in relation to symmetry, and/or symmetry violations?. Thus we will likely (eventually) obtain the tacit acceptance that all realistic AGI alignment methods depend on only causality, and only causal (or we will learn something really really interesting!). Another question: how is AGI not like Perpetual Motion?. Where in the same way that 'all forms of perpetual motion' somehow eventually require an assumption that violates the notion of symmetry in the form of the law of conservation of matter and energy, it can similarly be shown that, eventually, all forms of all notions of 'AGI alignment' will eventually require the assumption of something that causation simply cannot do. And causation is a symmetry concept. By eventually requiring something from causation -- something non-causative -- that is outside of the scope of what is allowable via causation, that there also we are seeing a violation of the notion of symmetry -- some place where for exactly the same initial conditions that we get sometimes different output conditions. Hence, the notion of predictability itself is compromised. By requiring something that is causatively impossible, we reject the ability to even make predictions like "AGI will always do what is in the human best interest". Hence by contradiction of the required notion of symmetry, we can have "100% knowable" that "AGI alignment is impossible", and that this occurs the same basic way, and to exactly the same standard, that we can claim that we can know that perpetual motion is 100% impossible. And of course you know, the above will never work. No reasonable argument will convince an unreasonable person, and nearly everyone we have encountered (so far) in the AGI Alignment community has already demonstrated a significant and very easy willingness to adopt and tacitly accept every manner of of the illogical and unreasonableness, while cloaking in the guise of seemingly and apparently "rational arguments" (which are actually anything but). While we can accept that people in general are inherently unreasonable, I had hoped for a higher standard from a community of people who are explicitly calling themselves rationalists. Most people simply will not hold all of the above thinking together, at once, along with all of the detailed side arguments that are components in any actual reasoning about inherent AGI safety contradictions. It is still too much. It all comes to rest on what causation cannot do. Ie, limits of modeling, limits of what can be done with game theory, with imperfect information, with algorithms and compatibility at all, etc. How about this: Does anyone recognize any limits to causation, to algorithms, to what is computable, with imperfect information, etc? If people are believing/arguing that "anything is possible", then how is that not equivalent to asserting that there are no limits to what can be done with causation, reason, logic, etc? What else is going to distinguish between "possible true arguments" and "for sure untrue arguments/reasoning/logic"?. If everything is potentially provable, with any tools, and nothing is ever actually 100% knowable, than what is even the very meaning of the concepts of reason, "knowable", and "truth" at all? What else, aside from symmetry and violations of same (contradiction), can or will be treated as the valid and socially recognized/accepted distinction between truth and falsity? If not that, in what sense can anyone at all -- or any argument at all -- be called rational, valid, or accepted as known? ~ ~ ~ Establishing the formal contraction of long term AGI safety has to have *both* logical parts and relevance parts. Where/if we wanted to construct something that is focused on only just the logic validity, then we end up with an overly simplified toy model. The problem with that is that then everyone would simply discount it as "not relevant" -- ie, the toy model does not correspond to the real world. Math without relevance is simply useless, in this case. In regards to relevance, part of the problem is that people tend to think that they can define whatever terms in whatever way, as if it does not matter. Yet there is just as much discipline needed in how terms are defined, so that the right relevance is available. This is _not_ a common skill. Hence, there is a balance between how any argument is epistemically set up, and what it even means to have a truth context. Do we spend more time to scaffold the skill building to make the argument more airtight (but very much longer overall), or do we abbreviate some parts of that skill, and the reasoning behind both the argument and the methods by which that argument is "reasoned", and end up wasting a lot of bandwidth processing needless confusions and mistakes (or politics, bias, motivated reasoning, etc). There simply is no "low bandwidth option" other than summarizing in outline (not a proof), which of course, _also_ displeases everyone as "unsatisfying". There just really is no way to make the overall inferential distance less, without somehow taking short cuts based on prior knowledge which it seems most people in the Alignment Forum simply do not have (or choose to ignore?). Basically, I am unwilling to try to field a proof of anything when the background context is one of an asserted complete absence of any coherent notion of reason itself. Either people on the AF are reasonable, and they are actually willing to commit and certify to that, to the notion of rigor, or they are not -- no half measures on this one. And no trust of any kind, particularly in concerns to actual responsibility to safety issues, otherwise. Hype will not cut it here. > Can I publish the 1st paragraph of this essay > so as to describe something of what the issue is? I would be really interested to see if anyone even understands that one paragraph. It is, near as I can tell, the crux of the matter. Either the "AI Alignment" people are "rationalists" and therefore are actually "rational" and are therefore accepting at least _some_ notion of 'epistemic coherency' -- ie, the notion that at least _something_ is '100% knowable', and therefore reliable, trustworthy, etc, as a standard of practice, etc, all based on the notion of 'reason'. Yet insofar as the notion of 'having reasons', ie, of logic and 'being logical' (in language usage, argument, discourse, etc, by constructing syllogisms) is inherently (cannot not be) based on the notion of symmetry, it seems therefore necessarily implied and required that the notion and use of the concept of symmetry, as fully 100% accepted, is therefore inherent in/as the basis of reason, and in fact, of any type of knowability at all. In fact, actually, this even applies to the notion of even perception itself: If there is _zero_ symmetry in the inputs and outputs of a communication channel, defined as what a sensor is sensing and what it transforms that into as a signal into, then even the notion of "sense" itself is obviated. There can simply be no notion of 'truth coherency', of (or in) mathematics, (let alone of "proof") or even of the notion of "science", of experiment based on perception and observation, if even the notion of regularity of consistent patterns of "1st we see this" and "then we see that" (ie, of causation) is itself being rejected. The notion of pattern and regularity itself is an implicitly unquestioned assumption of symmetry, as a basis for reason and rigor of thinking, itself a kind of abstraction of perception -- which is also tacitly assuming symmetry, etc. Yet far to many of the "AI Alignment people" seem to think that "everything is possible if only we try hard enough" (and get more money), and that "nothing is 100% knowable -- there can always be more observations/hope in the future". And it is this last that leads to so many problems. Is it 100% knowable that absolutely nothing else is also 100% knowable? How even do the AI people even account for even the one self-exception? Basically, the crux of the matter is that the notion of "rigor in proof" and "reasonableness" and of "rationality" itself (being a rationalist) -- of knowing and/or perceiving of anything at all -- is (cannot not be) _always_ exactly 100% dependent on, an at least an implicit notion of, a 100% acceptance of the notion of symmetry, with no/zero exceptions. And yet the consistent claim that "you cannot 100% know today that we may not someday do X". We all know, of course, that eventually, 'X' will become, will happens to be some notion of "align AGI to human interests". Seems so simple, so desirable, to believe and suggest that we can _eventually_ do what we want. And of course 'X' will _never_ be/become anything like "prove to 100% confidence that AGI alignment is actually completely structurally impossible, for any real notions of 'aligned'", and 'safe' etc. The net result of this ambiguity of what value 'X' is /allowed/ to take (only good claims, never bad ones) ends up being exactly and logically and reasonably equivalent to the notion of _sometimes_ rejecting the idea of symmetry as being an applicable notion (in regards to reason, logic, knowability, etc). And of course, there is no actual justification for this notion of "sometimes" that is not the simple illogical bias of motivated "reasoning". In this way, the "rationalists" become irrational, at least when it comes to all of the AI hype. So the net result is that the "AI Alignment" research community is _already_ behaving in a way that is already fully 100% knowably and verifiably illogical and irrational. And that is well before even attempting to describe some actual reasoning in regards to how to show/know that AGI is inherently unsafe. People will quibble about definitions of words all day while the planet burns around them. Moreover, asking for "it" to be in "formal language" is a complete nonsense issue. If everyone here already knows from basic Comp-Sci that algorithms can be expressed in any language -- why not proofs too? If the "rationalist" community is wanting to maintain their title as actually being rational, them at least most of them will need to step up to the standard of actually being that, which also means actually rejecting the false notion that "nothing is 100% certain" and "everything is possible". If reason and logic cannot be used to ever establish any notion of "proof" of the impossibility of something so basic as a "perpetual energy machine", then actually being able to handle any reasoning about something more abstract such as "perpetual benefit machine" (which AGI utopia is currently being billed as) is simply not possible for this group of people. It is simply not worth my time to start. How does the community of rationalists even maintain the title of being rationalists if they do not even begin to uphold the basic notions of reason and of logic and knowing itself? Frankly, I am feeling so disgusted that I do not even want to write anymore right now. ~ ~ ~ If you want/need to send us an email, with questions, comments, etc, on this or any other topic, or on related matters, use this address: ai@mflb.com (@ Mode Switch com.op_mode_tog_1();) + (@ View Source com.op_notepad_edit_1();) Back to the (@ Area Index https://mflb.com/ai_alignment_1/index.html). LEGA: Copyright (c) of the non-quoted text, 2022, by Forrest Landry. This document will not be copied or reproduced outside of the mflb.com presentation context, by any means, without the expressed permission of the author directly in writing. No title to and ownership of this or these documents is hereby transferred. The author assumes no responsibility and is not liable for any interpretation of this or these documents or of any potential effects and consequences in the lives of the readers of these documents. ENDF:
prev
000 of 000
next