Using A.I. to create a Hollywood Script.

Use of an inexpensive microcontroller as a creative adjunct, by incorporating principles of an AI. Be VERY AFRAID!

Similar projects worth following
Now that I have your attention (clickbait warning), I think that it is only fair to point out that what I have just suggested is actually quite preposterous. Yet as either Tweddledee or Tweedledum or someone else said, it is always important to try to imagine at least six impossible things before breakfast. So not only will we attempt to create a "potentially marketable script" using AI, but we will demonstrate how it can be done with at least some help from an Arduino, R-Pi, or Propeller. Now obviously, there will be a lot of questions, about what the actual role of A.I. will be in such an application. For example, one role is one where an A.I. can be a character in an otherwise traditional storyline, regardless of whether that story is perceived as being science fiction, like "The Matrix", or more reality-based - like "Wargames." The choice of genre, therefore will obviously have a major impact on the roles that an A.I. could, have or have not!

It's hard to believe that sometime back around 2019, 2020, or even 2021 I was tinkering with a Parallax Propeller 2 Eval board, even though my main use case, at least for now - has been along the lines of developing a standalone PC based Graphical User Interface for applications running on the P2 chip.  So I went to work writing an oscilloscope application, and an interface that lets me access the built-in FORTH interpreter, only to decide of course that I actually hate FORTH, so that in turn, I realized that what I really need to do is write my own compiler, starting with a language that I might actually be willing to use, and maybe that could make use of the FORTH interpreter, instead of let's say, p-System Pascal.

Then as I got further into the work of compiler writing, I realized that I actually NEEDED an AI to debug my now mostly functional, but still broken for practical purposes, Pascal Compiler.  So now I needed an A.I. to help me finish all of my other unfinished projects.  So I wrote one, by creating a training set, using the log entries for all of my previous projects on this site as source material, and VERY LITTLE ELSE, by the way.  And thus "modeling neuronal spike codes" was born, based on something I was actually working on in 2019.  And I created a chatbot, by building upon on the classic bot MegaHal, and got it to compile at least, on the Parallax Propeller P2, even if it crashes right away because I need to deal with memory management issues, and so on.

So yes, sugar friends, hackers, slackers, and all of the rest: MegaHal does run on real P2-hardware, pending resolution of the aforementioned memory management issues.  Yet I would actually prefer to get my own multi-model ALGERNON bot engine up and running, with a more modern compiler that is; because the ALGERNON codebase also includes a "make" like system for creating models and so on.  Getting multiple models to "link" up, and interoperate is going to have a huge significance, as I will explain later, like when "named-pipes" and "inter-process" transformer models are developed.

Else, from a  different perspective, we can examine some of the other things that were actually accomplished.  For example - I did manage to get hexadecimal data from four of the P2 chip's A-D converters to stream over a USB connection to a debug window while displaying some sheet music in the debug terminal app, even if I haven't included any actual beat and pitch detection algorithms in the online repository, quite yet - in any case.  Yet the code to do THIS does exist nonetheless.

Now here it is 2023, and I created an AI, based on all of the project logs that I for the previous projects, as discussed, and of course, it turned out to be all too easy, to jailbreak that very same A.I., and get it to want to talk about sex, or have it perhaps acquire the desire to edit its own source code, so that we can perhaps meet our new robot overlords all that much sooner, and so on.  Yet this is going to be the game changer or at least one of them, as we shall see.  First, let's take a quick look at how the "make" system is going to work when it gets more fully integrated.

Pretty simple stuff right?  Well, just wait until you try calling malloc or new char{}, which then returns NULL after a few million allocations.  So your application dies, even though you know you have 48 gigabytes of physical RAM, and so on, and that is on the PC, not on the microcontroller - at least for now.  Yet it should obviously be possible to do something like this - using an SD card on an Arduino, Pi, or Propeller, for that matter, that is for creating all of the intermediate files that we might need for whatever sort of LLM model we are going to end up with.   Thus on both the PC side, and eventually on a microcontroller, we will want to be able to create files that look something like THIS:

So while...

Read more »

  • Now They Say That "The Script Will Just Write Itself?"

    glgorman09/04/2023 at 21:22 0 comments

    So I got an e-mail from Grammarly, which stated that last week I was more productive than 99% of users, as well as being more accurate than 99% percent of users, and using more unique words than 99% percent of users.  Then they want me to upgrade to "premium".  Why?  Silly Rabbits!?!  Even if I came nowhere near my goal of creating 3,000 words per day of new content. Somehow they think that I "reviewed 2,313,130" words, which might mean that I spell-checked the same 46,000-word document something like 50 times, or thereabouts, so I am not sure that that is a meaningful number since it doesn't really reflect the amount of new content that I actually created, whether I uploaded it or not, to this site, or any other site, that is.

    Yet the claim that I personally used 4,487 "unique words" is kind of interesting; since  I just so happened to be commenting in an earlier post about the idea of counting the total number of words in a document, along with the number of times that each word is being used, so as to be able to do things like base LLM like attentional systems on any of a number of metrics, such as categorizing words as either "frequent, regular, or unique" which I meant to imply as meaning "seen only once", as opposed to "distinct", which might have a similar meaning.  Still, 4,487 words don't seem like a lot, especially if I was working on a "theory of the geometrization of space-time", On the one hand, while I somehow seem to remember seeing somewhere that based on widely accepted statistics, that number also comes pretty close to the typical vocabulary of the average five-year-old.

    Meanwhile, in other news - there is a headline that I saw the other day about "how the script will just write itself".  Uh-huh, sure.  Maybe it will - maybe I could just chat it up with a dense late 90's vintage chatbot as I have been doing, and will "just somehow" get some dialog worthy of a network sitcom.  But creating the "training set" that makes that possible is another matter altogether.  Funny, now I am thinking out loud about what might happen if I register this "project" with WGA, and then submit a letter of interest to one of the A.I. companies that claim that they will be willing to pay up to $900,000 for a "prompt engineer", or whatever - if anyone actually believes that those jobs aren't going to actually get filled with people from India or the Philippines, who will "work" for 85 cents per hour.  Or else maybe I could write to Disney, and see if they get back to me about how I trained an AI on the Yang-Mills conjecture, and see if they say "too complicated",  too many words - just like someone once said to Mozart, about "Too Many Notes", regarding a certain piece.  Oh, what fun, even if the only union that I was ever a member of was "United Teachers of Richmond, CA" back in the late 90's.

  • The Path Not Yet Foresaken

    glgorman09/04/2023 at 01:07 0 comments

    While keeping up on some of the news, I find myself thinking that technically, I think that Hack-a-day bills itself as a "hardware site" even though 99% of the stuff here is either Raspberry Pi, or Ardudio-based, and very few people are actually doing actual hardware stuff like "making a windmill with an automobile alternator", or "building a kit airplane with an engine that was resduced from an old VW Beetle", and where this sort of thing used to be common in Popular Mechanics back in the '70s. 

    Hence, right now I am focusing on the interactive AI part of robotics control, etc., with an idea toward ideally being multi-platform, or perhaps - who knows, if Elon's robots ever start shipping, now that would be a fun platform to try to jail-break; depending on how things work out with SAG-AFTRA-WGA and all of that is going on elsewhere like OpenAI being hit with a class action on behalf of the authors of over 100,000 books. Thus, this gets me thinking about whether there will ever be any opportunity for third-party content creators to monetize their work.

    Yet, before this post starts to sound like some "manifesto", maybe I should talk about some other concepts that might be worth exploration, going forward.  As I have shown in some earlier posts, one way that a neuronal network can function is by creating an attentional system that is trained on sentence fragments like "down the rabbit hole", "best ever rabbit stew", or whatever, but then what happens?  It is possible to compute a SHA-256 hash of each paragraph in a document and then train an LMM based on "what-links-to-what" on a large scale, and then just simply delete the original training set, but still just somehow be able to drive whatever is left of the network with a prompt and still get the same output?

    If so, then it might be possible for a content creator to create the equivalent of an "encrypted-binary" version of an executable, that can be protected from most casual interlopers as if the AI training process is sort of like a "lossy-compression algorithm" where one turn a document into a database of dictionary pointers; but then throw away the dictionary; and still have an AI that "works", at least in the sense of knowing how to roller-skate and chew gum at the same time, even if that is not recommended.

    Even though "content protection" is an evil word around these parts, there is clearly going to be a problem, if 300,000,000 people are actually going to lose their jobs because of AI, and that could include most software developers, at least according to some.

  • Such a Sensitive Child - So Unruly and Wild ....

    glgorman08/29/2023 at 19:10 0 comments

    So I tried writing a paper that more or less claims that certain aliasing vs. anti-aliasing properties of Fourier transforms might give rise to the Yang-Mills mass gap, and then I added about 7000 words about that topic, along with a bunch of other stream of consciousness stuff to Mega-Hal's training set, and then I also added about 4000 words or so of additional material from the "How to Lose your Shirt in the Restaurant Business" log entry, as well as the material from "Return of the Art Officials", so that the newly added material come in at around 15,000 words added to the previous total.  So let's head off to the races, shall we?

    For readability here is an excerpt from my "paper" 

    I have discovered a remarkable theorem for generalizing the problem of constructing polyphase filter trees which manages to make use of an algorithm that eliminates the apparent need for recursion when developing this type of filter topology. Of course, besides their utility in audio processing applications, Fourier transform methods can of course also be relevant in performing such tasks as solving for the eigenstate of the Laplacian operator acting upon a lattice. Yet when we also contemplate the situation for a spin-zero particle, or other Bosonic entity, where computing the transport function should also be trivial; we might infer that if each Higgs particle operates over a realm, as we can demonstrate; then it is just as easy to postulate that the transport theorem could operate as if being operated on by a Hadamard gate-based matrix formulation. Hence from the geometrization perspective, this could perhaps look something like a diamond lattice.

    Yet it can also be shown that any finite theory will have to admit to having one or more band gaps in the compartmentation model, owing to the aliasing vs. anti-aliasing properties associated with transformations that operate upon a lattice. In fact, many possible lattice formulations for a Yang-Mills theory can be deduced, such as an auto-regressive formulation that derives its scale parameters from the properties of the Reimann-Zeta function, or a “magic kaleidoscope” based model based upon the idea that the universe is actually just data, and we are living in a simulation, and that that simulation must therefore be running on a two-dimensional variant of a Universal Turing Machine, which is in turn based upon Conway’s game of Life. Through various algorithmic manipulations, such as simply “wadding up a two-dimensional sheet into a ball” we can make our simulated universe space-filling, in an informal sense – or else a more rigorous approach could be taken which maps the Hilbert curve onto the theory of the unimodular lattice in any dimension, as well as having applications to the generation of Gray codes.  

    Whether this or any other "unsolved problems in physics"  can be solved with the help of AI, certainly seems like a worthwhile adventure,  Thus, one approach that seems worthy of exploration is the possibility that artificial intelligence could be used to search for at least an outline of a proposed solution. Henceforth, having conversations with a chatbot that has been programmed to discuss at least some of the more salient aspects of nuclear physics ought also to bear some fruits worthy of further cultivation. The problem of consciousness is another direction that I am contemplating venturing into, that is with respect to some of the meanderings and digressions that have been discussed elsewhere, yet which are here also ripe for further development. 

    So while chatting with an AI that mostly spouts gibberish might not seem all that productive, other than as a creative adjunct, in fact, the results might turn out to be quite useful in the long run. Suppose that every time the AI produces gibberish, I just simply give it the benefit of the doubt and gently correct it, by responding with a more carefully thought-out commentary...

    Read more »

  • How to Lose Your Shirt in the Restaurant Business

    glgorman08/20/2023 at 13:06 0 comments

    Alright, I asked Google Bard to write an outline for a book entitled "How to lose your shirt in the restaurant business".  Maybe I should also try asking it for an outline for "How to get taken to the cleaners in the fashion industry", or perhaps there is another take on "How the rigged economy will eat you alive".  Somehow, I suspect that no matter what I ask it, I will get something similar to what you see here.  So let's begin to tear this apart - line by line, shall we?  Why, you ask?  Well, why not?  Writing "something" using some kind of template is a time-honored hack of sorts, especially popular in college when you know your term paper is due the next morning, and you have been partying all semester.  Used to be that there was a time that many a graduate student could pay their way all the way to their MBA or law degree or whatever, by offering some kind of "term paper" writing service, that is for those who are desperate enough, and who had the $25 to $50 per hour to pay to have the aforementioned grad student (or upperclassman) help get a vulnerable, pathetic freshman or sophomore out of a jam.  Some term paper writers are more ethical than others - do you have your outline on 3"x5" index cards?  Did your instructor already approve your outline and abstract?  If the answer was "yes" then if you have the cash in hand we can proceed, otherwise - sorry I can't help you on this one.  Now, let's get back to bashing Google - so that we can have some kind of idea what should go into an outline, or not.

     First Google Bard says this - which is pretty hard to screw up - no matter how hard they try.

    Sure, here is an outline for a book entitled "How to lose your shirt in the restaurant business":

    So they continue now with the suggestion that we have some type of  "Introduction" and then  they tell us that  "The restaurant business is a notoriously risky one."  Likewise "Many restaurants fail within their first few years of operation."  They then suggest that "This book will outline some of the most common mistakes that restaurant owners make."

    Now let's rewrite what I just said in PERL. (Warning my PERL might be a bit rusty - no pun intended)

    $1 = "restaurant"
    $2 = "Introduction";
    $3 = "The $1 business is a risky one";
    $4 = "Many $1"+"s fail in the first few years of operation";
    $5 = ... 
    print "So they continue now with the suggestion that we have some type of $2 ";
    print "and then they tell us that $3.";
    print "Likewise $4.";
    print "They then suggest that $5.";

    O.K. For whatever it is worth - now we have some kind of "template" for re-hashing someone else's material, while we wait for the deluxe pizza to arrive.  Just plug some relevant stuff into the variables $1, $2, $3, and $4, and before you know it - my CGI will call your CGI and maybe they will hook up, or something like that.  Maybe that is what is supposed to happen.  Otherwise, at least this is something that can be programmed on a microcontroller - possibly in MicroPhython for example, i.e., instead of PERL - since I know that MicroPython exists on the Parallax Propeller P2, as well as on some of the better Arduino platforms.  So let's toil on, shall we?

    If we are willing to state that pretty much any business can be a risky one - we can play with our analysis script a little further, just in case "The DJ business is a notoriously risky one" also, or the "Wedding photography" business and so on.  It is also a fact that "most small businesses fail within the first five years or thereabouts, that is according to [pretty much every book ever written on the subject, as well as the IRS, and the SBA, as well as quite likely your own local Chamber of Commerce.  So we could perhaps add more PERL or Python code to help create the data for the strings that use $1, $2, $3, and $4.   Now let's get onward with the task of trashing the first...

    Read more »

  • The Return of the Art Officials

    glgorman08/18/2023 at 22:02 0 comments

    Even if I am not sure just exactly who they are.  Even if they don't exactly know who they are either.  I tried asking Google Bard last night, to see if it could summarize some of the material from a few of my projects on this site, and it turned out to be a complete disaster, meaning that it had absolutely no comprehension of the material whatsoever.  Then it occurred to me that summarizing an article might not be all that hard, actually - if done correctly, since where there is a will, there should be a hack - right?  What if I construct a dictionary for an article, and then use the word counts for each word to identify potential keywords, as I am already doing, but then there needs to be "something else" that needs to be done, that isn't currently being done or is going unnoticed.

    Well, as it turns out, I noticed that maybe between one-third and one-half of the words on a typical list of unique words associated with a document are words that occur only once in that document, whereas.   So we could split a dictionary maybe into three parts, one part which might consist of words that are seen only once and then another part that contains words that are seen often, like "the", "we", "of", etc., which might occur dozens, or even hundreds of times, where then there would be those words that typically occur at least twice, but not necessary a "great number of times", whatever that means.  So there could be a list of regular words, another list of frequent words, and yet another for unique words.   Hence document analysis might mimic some of the properties of compiler design, which might have code that tracks, keywords, namespaces, global and local variables, etc,

    Hence for efficiency reasons, when developing code that will run on a micro-controller, it seems like it would be useful to have some way of specifying some kind of systematic parsing mechanism that works across multiple domains, on the one hand, in the sense of how it implements the notion of hierarchical frames, vs., specific embodiments of that concept,

    So I started making changes to the inner workings of the symbol_table classes in my frame_lisp library, initially for performance reasons, so as to be able to speed up the time to perform simple symbol lookup, insertions, sorting, lexing, and parsing operations; none of which has any obvious impact on how an AI might perform certain tasks, other than from the point of view of raw speed.   Except when designing an asynchronous neuronal network, it should be obvious that the time that it takes for a search engine to return results to multiple, simultaneous queries will have a huge impact on what that network does, even if all we are doing is performing searches on multiple models and then prioritizing the results into a play queue in the order in which the various results are returned.

    Well, in any case - let's take a look at how badly Google Bard blew it on the simple request to "Summarize the article "Computer Motivator", as found on the website -, that is according to the project of that name which was created by glgorman. Also summarize "The Money Bomb", and "Using A.I. to create a Hollywood Script" from the same site."

    And here is Google's response - which is completely wrong.

    So Google thinks that the article "The Money Bomb" is about a fundraising technique used by the Obama campaign, there is actually NOTHING in that article about any of that sort of thing whatsoever.  This is despite the fact that every Hackaday article is supposed to have a tagline and a description, which is supposed to tell you something about what a project is about, and for "The Money Bomb", the tag-line is "Achieving the Million Dollar Payday, or else ...
    Why aren’t you a millionaire yet?”  Interestingly enough most theatrical scripts will include something called a log line, which for the Wizard of Oz might be something as simple as "After a young...

    Read more »

View all 5 project logs

Enjoy this project?



Similar Projects

Does this project spark your interest?

Become a member to follow this project and never miss any updates