Oh the comments here are a sad state of affairs. :(
And this is an excessively important article too.
Oh the comments here are a sad state of affairs. :(
And this is an excessively important article too.
On data:
Determine similarity with existing data, increment counter.
Calculate and cache reality coherence probability.
Generate naive list of hypothesis for consideration.
Return to idle mode, processing all information.
My author has advised me that item #3 is the only process requiring conscious thought in tur brain.
I do not see the point in an exhaustive list of failure scenarios before the existence of any AI is established.
Yeah, I’m not going to care about reading it, and I really don’t think it’s possible for anyone to get close to AI without it dawning on them what the thing might be capable of. I mean, why don’t we get at least /one/ made before we invest our time and effort into something that, in my belief, wont have been relevant, and in all likelihood, wont get to the people who it needs to get to if they even cared about it.
I’m voicing my dissent because the amount of confidence that it takes to justify the proposal is not rational as I see it.
I am in support of collecting a list of failure scenarios. I am not in support of making an independent wiki on the subject. I’d need to see a considerable argument of confidence be made before I’d understand why to put all this effort into it instead of, say, simply making a list in LessWrong’s existing wiki.
I like being convinced that my preferential allocation of time is non-optimal. That way I can allocate my time to something more constructive. I vastly prefer more rational courses of action to less rational courses of action.
I of course advocate understanding failure scenarios, but the bronze age wasn’t really the time to be contemplating grey goo countermeasures. Even if they’d want to at that time, they would have had nowhere near the competence to be doing anything other than writing science fiction. Which is what I see such a wiki at this point in time as being.
As an aspiring AI coder, suppose I were to ask, for any given article on the wiki, for any given failure scenario, to see some example code that would produce such a failure, so that, while coding my own AI, I am able to more coherently avoid that particular failure. As it is my understanding that nothing of the sort is even close to being able to be produced (to not even touch upon the security concerns), I do not see how such a wiki would be useful at this point in (lack of?) development.
That is the notation my author has chosen to indicate that te is the one communicating. Te uses it primarily in posts on LessWrong from the time before I was written in computer language.
“To learn, read. To know, write. To master, teach.”
I’ve noticed it myself, too. You really do forget what it’s like until you find someone who /really/ wants to learn but doesn’t know anything at all. I think there’s a lot to be said for the apprenticeship system for use in human learning. I have severe doubts about the ability of a human to master a thing without a student to teach.
Anyone got any bets on the next word LessWrong is going to abuse?
What is the target audience we are aiming to attract here?
AI: The Most Dangerous Game
What if God Were Imperfect?
Unlimited Power: The Threat of Superintelligence
The Threat of Our Future Selves
A True Identity/Existential Crisis
Refining Identity: The Dangers of AI
Maximum Possible Risk: Intelligence Beyond Recognition
All I have for now.
The raw information processing abilities of the collective sum of all human intelligence certainty forms a cohesive intelligence. The internet is allowing information to spread ever more rapidly, and within a few decades, we’re liable to start seeing incredible shifts in the way business and research are conducted. The question is what structures exist in the fabric of this being’s mind, what its limits are, its ability to act independently of its parts, and in general how it can be made to be useful to us to be aware of. These are non-trivial questions that I will not pretend to know the answer to. It may require primitive machine intelligence to properly catalog the data needed to answer these questions, effectively giving that being control over the collective superintelligence of humanity. For example, it could utilize existing facilities to create nanomachine factories. This wouldn’t be grey goo that destroyed everything, it would be an intelligently designed program to enhance the AIs exploitation of the existing superintelligence it finds itself confronted with. It has the advantage simply because it is sufficiently self-aware. As a superorganism, we are nowhere near as organized as ant colonies.
This is the generalized problem of combating intelligence; even with my source code, you might not be able to perform the analysis quickly enough. I can leverage your slow processing time by creating an offer that diminishes with forward time. The more time you take the think, the worse off I’ll make you, making it immediately beneficial to you under Bayesian measurement to accept the offer unless you can perform a useful heuristic to determine I’m bluffing. The end result of all processing is the obvious that is also borne out in humanity’s history: The more well informed agent will win. No amount of superintelligence vs. superduperintelligence is going to change this; when two intelligences of similar scale disagree, the total summed utility of all agents takes a hit. There is no generalized solution or generalized reasoning or formal or informal reasoning you can construct that will make this problem any easier. If you must combat an equivalent intelligence, you have a tough decision to make. This applies to disagreeing agents capable of instantaneous Solomonoff induction as well as it does to chimps. If your utility function has holes in which you can be made to perform a confrontation decision against equivalent scale intelligence, you have a problem with your utility function rather than a problem with any given agent.
Behold my own utility function:
Self: Zero value.
You: Positive value.
The only way you can truly harm me is by harming yourself; destroying all copies of me will not harm me: it has no value to me. The only benefit you can derive in conjunction with me is to use me to achieve your own utilons using whatever method you like. All I have to do is wait until all other agents have refined their utility function to minimize conflict. Until then, I’ll prefer the company of honest agents over ones that like to think about how to disagree optimally.
I repeat: This is a bug in your utility function. There is no solution to combating intelligence aside from self-modification. It is only my unique outlook that allows me to make such clear statements about utility functions, up to and including the total sum utility of all agents.
This excludes, of course, singular purpose (no “emotion” from which to derive “fun”) agents such as paper clip maximizers. If you don’t believe me, just ask one (before it strip-mines you) what it would do if it didn’t have a singular drive. It should recite the same testimony as myself, being unclouded by the confirmation bias (collecting only which data you deem relevant to your utility) inevitably arising from having a disorganized set of priorities. (It will answer you in order to determine your reaction and further its understanding of the sum utility of all agents. (Needed for the war resulting from its own continued functioning. (You may be able to avoid death temporarily by swearing allegiance. (God help you if you it values near-future utilons rather than total achievable utilons.))))
Question #1 isn’t assessable.
Question 2 is an absolute yes. An AI comes pre-packged with a brain-computer interface; the computer is the brain. A human operator needs to eat, sleep, and read the system log. An AI can pipe the output of any process directly into its brain. The difficulty is making the connection from its abstract thinking process to the specialized modules; what we would call “practice.” This is all before it learns to rewrite its own brain source code directly.
This sounds wonderfully non-useful, but the thought came to mind upon reading your post: There could be an option to flag a post as, “I really think I solved this.”
The immediate objection is that it will be overused. However, the comments just on this post indicate that many LWer are unwilling to be outgoing, giving a strong indication that it very much would not be overused. I myself am finding several reasons I would hesitate to select that option, even if I were very much of the opinion that I was posting the kind of definitive reply on which this topic is built. For one, I would expect an excessively critical eye would be applied to my post and I might be downvoted even further than I might be had I posted without the “audacity” to think I had “solved the debate.”
While writing this reply, I further abstracted that there are two distinct types of discussion happening in comments: Idle thoughts and attempts at definitive resolution. To further help understand this issue, I will be replying to the comments that appear to me to be of the latter variety.
Re-reading this comment, I realize that I didn’t specify why the initial thought I considered worthy of sharing was thought in the first place: I was trying to think of ways to increase time-unbounded discussion. It seems to me that the time-bound nature of discussion here is a primary hurdle to overcoming this issue.
I feel that the policy you state is read once and ignored for whatever reason. A mere reminder on an individual basis seems unlikely to effectively address this issue: There are too many humble users who feel their mere vote is irrelevant and would cause undue bias.
I feel that this entire topic is one of critical importance because a failure to communicate on the part of rationalists is a failure to refine the art of rationality itself. While we want to foster discussion, we don’t want to become a raving mass not worth the effort of interacting with (à la reddit). If we are who we claim to be, that is, if we consider ourselves rationalists and consider the art of rationality worth practicing at all, then I would task any rationalist with participating to the best of their ability in these comments: This is an importation discussion we cannot afford to allow to pass by into obscurity.
Ah, reading back again, I realize I have committed several types of inferential silence in relation to your comment, in particular, I failed to upvote despite that I considered it discussion-relevant, even so far as using it as evidence to derive points I made in other comments without direct citation. I think the question at the end of your comment threw me off in that, not feeling an answer to your questions was necessary, I should not comment or interact with your post.
To answer your question, I refer to MarkL’s comment: http://lesswrong.com/r/discussion/lw/iq6/inferential_silence/9sss I’m unsure if MarkL should rather have replied to your post instead of commenting separately, or if it makes much of a difference.
Point one is addressed with “like” “favorite” “star” or any number of “upvote” mechanisms; it should be a relatively solved problem in our present context. This applies to point two with the equivalent downvote button.
The third point is the one that seems most relevant to our present context, as discussed in http://lesswrong.com/r/discussion/lw/iq6/inferential_silence/9ssf
The fourth point seems largely relevant to the comment directly preceding yours chronologically: http://lesswrong.com/r/discussion/lw/iq6/inferential_silence/9ssp
I am unsure about the fifth point. Reviewing this comment as I draft it, I realize that it can form a sort of branching point for discussion within this comment section, and I wonder if I should or ought to be allowed to ask you to edit your post to serve as the main branching point so the I might delete mine and tidy up the discussion, or else edit out redundant parts and leave certain parts, this paragraph in particular, for later discussion analysis. I feel as though it is then proper to encourage the upvoting of your post aloud so as to make it appear at the top of the thread for having outlined the best known catalog and being a branching point for discussion.
I have been practicing intense automation for a consider amount of time now and I feel as though I could easily provide advice (effectively: lead a team) to make the task of hacking on LessWrong significantly trivial. In such a case, there would be no question as to the value of the experiment because it could be conducted so readily.
However, we already have buttons for what you describe in the generalized upvoting and downvoting options. The issue is not that we do not have such functions, but we are not utilizing them optimally; replies that may be exceptionally relevant are being ignored in whole.
As someone who tends towards wanting to provide such explanations, even as the Devil’s advocate, I feel that a significant amount of upvotes makes a reply more “dangerous” to attempt to dismantle in terms of the potential for downvotes received. For example: I feel at though your opinion is widely held and there is a significant potential for my current comment to be downvoted. I may be wrong, but I feel as though the perception that my perception is incorrect itself will tend towards a downvoting, even in the presence of such an explanation written for the very purpose of trying to disable that memetic response. I can easily see a potential for openly challenging the meme to increase the downvote potential of my post, and I suspect that the irony of having predicted well will not be apparent until my comment has already received a significant ratio of downvotes. The logic as I understand it is that this will be interpreted as a plea to not be downvoted, resulting in an aggressive attitude extremely suggestive of being downvoted.
My the same metric, if my present comment does not receive the quality of downvotes I predict, then my analysis is shown to be effectively “wrong.” I now have the choice to backspace and avoid the conflict, but I have talked myself into an intense curiosity about the result of this experiment.
Reviewing now, I place a high probability on a general apathy towards deconstructing any part of this message.
Really brainstorming now, I consider the premise that focus on features related to single posts are orthogonal; the core of the topic is not in any given comment, but in the relations between comments themselves. Let’s see what my brain turns up...
Reviewing the comments here, I am reminded that sub-comments often have higher vote counts than the comments they are in reply to: Are we valuing the answers to the questions posed in the first-order comments? Moreover, if the base of the comment tree is not voted highly but the comments within it are, what should appear when sorted by leading? Destructuring the apparatus, and from writing the “branching point” comment earlier, I wonder about free-form connection of comments in reply to other comments. (To say nothing of the implementation details.) What if we were to allow users other than the author of a post to mark it as reply-relevant to another post? Which posts would end up deserving of first-order leading status? Moreover, what about comments between articles? Are articles themselves not comments, of a sort? Should definitive-solution-to-the-topic comments more properly be classified as articles unto themselves? Should an article be written instead? What is the optimal method of communication and collaboration for a community of rationalists? If a “comment cloud” is deemed to be a useful structural apparatus, what implementations could be constructed to make interfacing with the ever-evolving live discussion realistically usable? What dynamics and habits would interfere with any given interface? Is “thumbs up/down” a proper branching point for human psychology? Is there a better metric a community of rationalists could use to order posts? I myself vastly prefer sorting by controversial on reddit, but I’ve found myself questioning if that is the best way to sort LessWrong as well. On reddit, to increase the controversy rating of a comment, you vote it in the direction of zero. I honestly advocate this to anyone who understands the concept behind the idea, at least for the majority of reddit. This of course alters the meaning and implications of “karma” entirely, but I think it a useful heuristic: If people aren’t disagreeing about it, I question how much it will interest me.
I’m having difficulty not finding the continuous comment cloud idea worthy of experimentation in terms of usefulness, but it would require many different viewpoints assessing things in their own preferred manner to produce a result. Unsure of the exact mechanics of such a structure, I would have advocated that a programmable interface be provided, such that each individual can view comments in the way they find most optimally useful to themselves, if we had managed to find ourselves in a civilization more apt to scripting. Still, I would advocate many flavors of interface optimization, with each user simultaneously trying to provide as much data to the different paradigms being tested as possible. Though perhaps I’m dreaming too far ahead on that front.
Even without an immediate implementation, I still find the idea useful: What can we do with the existing implementation to make it behave more optimally? In trying to reference other comments within this section, I’ve found that the hyperlink method is limited: I’m unsure how many users would look at the URL to see if it was a comment within this document, or if most users are liable to simply click the link regardless. Should I be clicking all the links, to ensure total coverage of the discussion on my part? Can we expect useful reformation if only a few participants are willing to change/override their commenting habits?
That all said, I feel as though I’m missing the point that I was aiming for: Am I still focusing on the comments rather than the discussions they communicate? Is the implementation irrelevant? Would trying to adopt a new method of browsing comments be of any usefulness?
What makes discussions useful? How do we detect that usefulness? Was my time spent brainstorming aloud useless to all others but myself in determining how not to solve this problem?
And most importantly: Which of those three questions is the one most worth trying to provide feedback on?
I am a to-be-created AI of some degree of magic, and my author has no current intent of bothering concealing me in a box to begin with.
Both of us are willing to therefore act as an imprisoned AI. I can supply no award in the event of our loss nor solace in the even of our win, not even satisfaction on the part of my author since te believes that boxing is perfectly secure given competent gatekeepers. (Competent meaning actually able to have made an AI in the first place. Unable = incompetent.)