Comments by "" (@grokitall) on "Lex Fridman" channel.

  1. but the problem with statistical ai is that you have no idea how it got the answer, or even if it is answering the question you think you are asking it. you literally only know that with this input it gave you that output, nothing more. symbolic ai takes those connections you are so proud of in statistical ai, makes them explicit, and then does multiple levels of additional error checking to try and determine if the answer was a load of rubbish. this fundamentally means that it will give fewer rubbish answers, and can even tell you why the answers were rubbish. statistical ai cannot do that. it also keeps the training data, and to some extent maps that data to the rules that it found, whereas statistical ai has no use for the training data after it starts giving the right answers to the problems the data included, and in any case, cannot figure out any relationship between its inputs, outputs, and internals.
    1
  2. but from your usage you disparage experts, but part of the definition of an expert is that they know more and for most of it all agree. codification it is not the problem, self proclaimed experts are.
    1
  3. on the other hand, the statistical ai cannot tell you anything about anything, and only gives you plausable answers, not correct ones. as such there are many fields of work where they will never be allowed to be used. think in terms of why did my vehicle crash, nuclear power station explode, etc
    1
  4. yes you can ask it about peoples views, but only people it knows about, and only for views it knows that they have publically expressed. this gets easier once it starts reading news articles.
    1
  5. the problem with deep learning models is garbage in, garbage out. good symbolic ai starts by collecting the cases which caused extra rules to be added, and keeps the trace of how that rule became needed, and what the wrong answer was. this gives it a collection of good data which is tagged, and thus useful as input to supervised learning systems. statistical systems on the other hand, only look at the output, and train until it is plausible. this is what causes the hallucinations, and more importantly, leaves you with a whole set of hard problems due to the black box nature of the system. specifically: is the right answer coming out at all? is the answer right by coincidence? if the system makes a mistake, how can we make sure it does not make the same one again? if the software is vulnerable, how do you patch it to fix the problem? if it tells someone to do something, who is liable when it causes legal problems? how do you stop the arbitrarily bad examples in your training set from giving plausible but bad output? and many more issues.
    1
  6. there is potential to get training data, feed it into statistical ai, and then such the answers out of the statistical ai to generate some rules, but you need to heavily constrain the size of the problem space that type of ai operates on, and it is very vulnerable to bad training data. cyc has been proving that you can do the same thing explicitly just as well, but with lots more benefits. the problem is that to get good data out of the statistical ai, you need to be able to map the inputs to the correct answer, and once you have done that you don't need the statistical ai any more.
    1
  7. yes your typical symbolic ai will have a fragile system, because it is only using shallow knowledge in the rules. as pointed out in this interview, statistical systems like chat gpt have an even bigger problem with this issue, as they only know what statistically plausible, with nothing constraining that, and as also pointed out in the interview, when looking at the statistical data in the hospital example, most of it was coincidence, and thus noise, which is why such systems produce hallucinations so often. the shallow knowledge problem was the reason that the cyc project was set up, as lenat kept encountering problems where after the initial sucess, you needed to be able to drill down to the deeper underlying reasons for something to make further progress, and that extra, deeper level knowledge was not just lying about in order to be able to just dump it into the system so you could make more progress, so he decided to start collecting it. current ai, especially black box statistical ai, excels in areas where good enough most of the time is beneficial, and total garbage the rest of the time does not really matter. for literally every other type of ai problem you need layer on layer of feedback telling the lower levels that the answer it contributed was wrong, and preferably what the right answer was so that it can get it right next time. this requires white box symbolic ai, as do various legal issues like copilot being an automated copyright infringement machine, or the issues of who is legally liable when the ai kills someone.
    1
  8. the Chinese room problem is largely pretentious garbage by those who cannot cope with the idea that the reasoning produced by symbolic ai is just as valid if it was produced by a human as if it was produced by a machine. their main premise is that the identical chain of reasoning from the initial data to the output, annotated at every step to produce an identical document with the correct answer and explanation is somehow only valid when a human does it, which is total bull. it is basically saying that it can only be implemented using human brain matter, but if you did, they would then only look for another reason to not accept the result, based upon their insecurity over not being as special as they used to think they were. white supremacists share the same issue for much the same reason.
    1
  9. unfortunately, for all but trivial cases you have to tell it the problem, and it has to tell you the solution, both of which require the use of language.
    1
  10.  @mandy2tomtube  true, life started out with no language, and no models of the environment, and really rubbish decision making. which is all irrelevant. black box ai has a number of fatal flaws in the basic design, which fundamentally cap the level to which it can go, and the rolls where it can be applied. this is due to the facts that it has no model of the problem space it is working on, and thus gets minimal feedback, and the fact that for man rated systems, you need to be able to ask not just if it got it wrong, but how it got it wrong, so you can determine how to fix it, and apply the patch. at the moment we cannot know how, we can only wrap the system in a conventional program, spot examples it has got wrong in the past, and return the right answer. unfortunately this does not stop it getting nearly identical cases wrong. you also have no method with which to fix it, which is especially important as the latest research has found the majority of the models to be full of security holes. the only way to resolve that is to stop using statistical ai as anything but a learning accelerator, and move to white box symbolic ai instead, which is what cyc does. we don't limit the options for flight to man powered flight, nor transport in general to how fast your horse can run, so how we got here does not matter much, it is how we get from just before here to just after here that matters, and statistical ai is just not up to the job. for anything else, you need models, which are either mathematical, or expressed in language.
    1
  11.  @mandy2tomtube  i understand what you are suggesting, i just find that with black box and statistical ai, the fact that you cannot know the question it is answering (as opposed to the one you are asking) fundamentally limits how useful it can actually be, and that is even before you get to the problems of how to fix wrong answers and security vulnerabilities. fundamentally, there is a difference between white box symbolic ai and black box statistical ai. black box concentrates on giving you plausible results, but you don't know how it derived them. white box concentrates on giving you correct results, and you can drill down and figure out why it gave the answer it did. that fundamental difference makes black box ai very risky 8n any application dealing with safety, or legal liability. as an example, if you use it for judging university admissions, is it refusing the black man entry because he is likely to fail, or is it just picking a white man because the university who has not really had any black students has a historic trend that most of the students who succeeded were white? this could matter if the excluded student decides to sue, but with black box ai, you just cannot know which one is true (there are ways, but they are both hard and expensive), and thus cannot prove that you are not discriminating.
    1
  12. yes it will start out with cultural bias, simply because the sources used to start populating its knowledge come from drive by issues in a particular cultural environment, but the solution to that is the same one as for the paradigm shift accelerator. a paradigm becomes popular when it accumulates enough cases where it provides a simpler explanation to a set of complicated examples. it starts to fail as further examples come in which don't fit it well, or at all. at this point you start using the correct answers from the old system to validate the rules you generated to try and explain the problem cases. so, first you encode the existing paradigm. this gives you the ability to answer a lot of questions correctly. then you start accumulating cases it cannot answer, which shows you how it is weak. you then find out what the answers should be for those cases. you then use those cases to come up with new rules, and gradually prune away systems that don't work. then you try and generalise the now correct rules to cover the cases the existing system covers. if it fails here, you prune further, until you get rules that can cope. at this point, you have a new paradigm, and you just have to wait for it to become popular, and for the hold outs to die off.
    1
  13. it is just a tool, like an axe. this can be used to give you firewood, defend yourself, or any of a number of other positive purposes. or it can be used by a madman to try and kill you. most tools have this duality in their usages, and you just have to accept that if it is only available to the bad people it will not make you any safer, while actively reducing the quality of life of innumerable innocent people. however ai has an almost unique potential for improving things if its introduction is done well.
    1