• [split] future video about AI, all input welcome /kg
  • [split] future video about AI, all input welcome /kg

    joben > 25-11-2024, 02:08 PM

    (25-11-2024, 11:00 AM)Koen G Wrote: You are not allowed to view links. Register or Login to view.Mike Plet, are you reading anything we say? If so, please give a sign.

    New Voynich video idea: "Why AI sucks and why your Voynich Theory is still wrong"  Smile
  • RE: AI says I my translations work 2024 NOVEMBER

    Koen G > 25-11-2024, 03:27 PM

    It's on my shortlist alright. I'm working on another one first, but I won't wait too long for this one Wink


    EDIT: I'll use this space to make a proposed structure for the video, all ideas remain welcome. Note: this is not a script, just a quick structure outline.


    1) Intro

    2) Explain that in current popular discourse, usually AI = LLM so I will be focusing on LLMs

    3) Advantages and disadvantages of current generation LLMs in Voynich research

    Advantages
    • ChatGPT is great at translating natural languages, much better than google Translate. This even works for imperfect inputs, older forms of languages... 
      --> I have an interesting example about this from personal experience, and I can easily show examples, this one is no problem
    • ChatGPT can even read image files and help you read handwriting in manuscripts 
      --> this was highly imperfect when I last tried it - any experiences?
    • ChatGPT can help you write code to use in your research
      --> I have an example of this where I used code to help me analyze text files for a video, not the most exciting one; other stories are welcome.
    • ChatGPT could be used to... brainstorm? Jog creativity? Something like that? How do I put this? Any examples?

    Disadvantages:
    • Current generations of LLMs hallucinate frequently, especially when they don't have a lot of data on a subject.
    • When they make a mistake, they do so with great confidence. You, the user, have to notice this mistake and correct them. This means that you still need to do your own research and fact check important elements.
    • ChatGPT is a "people pleaser". It prefers to make up an example just to give you something, and will gladly go along in whatever you propose if you are insistent enough. This means that for Voynich theorists, using ChatGPT is a bit like playing with their own echo.
    • Current generation LLMs still struggle with certain tasks essential to our work (not sure yet how to put this)
      --> this might change in the future. we may want to add the test here?

    General issue with AI

    LLMs work by analyzing a vast amount of linguistic content, and tap from this to give you a human-sounding output. It produces the text output that is most likely, based on all the text it was trained on. RadioFM: "I would argue that all current and future LLMs' 'reasoning' capabilities will continue to be bound by their underlying Language Model nature - they might be able to tackle tasks involving natural languages or some formal ones (algebraic expressions, predicate calculus, etc.) but will probably suck at translating or manipulating low resource languages like Voynichese."

    I can do a lot with the quote found by Matthias: 

    Quote:2. Research on problem-solving

    Research on problem-solving has a long tradition in both psychology and education. Jonassen (1997) classifies problems on a continuum from well-structured to ill-structured. Well-structured problems have a well-defined initial state, a known goal state or solution, and a constrained set of known procedures for solving a class of problems. In contrast, the solutions to ill-structured problems are neither predictable nor convergent because they often possess unknown aspects. Additionally, they possess multiple solutions or solution strategies or often no solutions at all (Funke, 2012).

    --> I can show that Voynich research is very much an ill-structured business


    (WIP)
  • RE: AI says I my translations work 2024 NOVEMBER

    Mike Plet > 27-11-2024, 05:15 PM

    (25-11-2024, 03:27 PM)Koen G Wrote: You are not allowed to view links. Register or Login to view.
    (25-11-2024, 02:08 PM)joben Wrote: You are not allowed to view links. Register or Login to view.New Voynich video idea: "Why AI sucks and why your Voynich Theory is still wrong"  Smile

    It's on my shortlist alright. I'm working on another one first, but I won't wait too long for this one Wink


    Please Do,.. AI sucks in a major way,.. need input?  Tongue
  • RE: AI says I my translations work 2024 NOVEMBER

    Koen G > 27-11-2024, 06:05 PM

    Any input is welcome -this is one of those topics you guys know much more about than I do. I'll split this off onto a separate thread.
  • RE: [split] future video about AI, all input welcome /kg

    Koen G > 27-11-2024, 06:09 PM

    Also in general, several commenters to my videos ask if AI will solve it. All ideas about this are highly appreciated.

    Also if you know someone I should interview about this.
  • RE: [split] future video about AI, all input welcome /kg

    Bluetoes101 > 27-11-2024, 07:10 PM

    Elon Musk, @ him on Twitter! Wink 

    Seriously though, it would be good to see a balanced view, the good and the bad. 
    A good experience I have had is having AI write simple code for me and amendments to it for performing data analysis, also looking up information available online and where to find specific info, such as "Are there copies of X manuscript available online and where to find it". 

    Bad experiences seem to be any sort of "thinking" or giving it images to decipher, for example I stumbled on this recently (You are not allowed to view links. Register or Login to view.) and I was very interested in the style of writing and specific bits such as the underlined word endings below (amongst others), but from experience using it I could tell it was just "very confidently guessing", if anything it just reminded me I need to stop cutting corners and go learn Latin.. 

    If you have anything specific in mind please let me know, I went through a bit of an "AI stage" and have all sorts Voynich related stored away.. 

    [Image: Capture8.jpg]
  • RE: [split] future video about AI, all input welcome /kg

    bi3mw > 27-11-2024, 08:26 PM

    (27-11-2024, 06:09 PM)Koen G Wrote: You are not allowed to view links. Register or Login to view.Also in general, several commenters to my videos ask if AI will solve it. All ideas about this are highly appreciated.

    While searching for information on the subject of AI solution strategies, I came across the following statement. I think that AI solutions for VMS actually fail because of the problem description.

    Quote:Opportunities of artificial intelligence for supporting complex problem-solving: Findings from a scoping review
    Srecko Joksimovic, Dirk Ifenthaler, Rebecca Marrone, Maarten De Laat, George Siemens

    ......
    2. Research on problem-solving
    Research on problem-solving has a long tradition in both psychology and education. Jonassen (1997) classifies problems on a continuum from well-structured to ill-structured. Well-structured problems have a well-defined initial state, a known goal state or solution, and a constrained set of known procedures for solving a class of problems. In contrast, the solutions to ill-structured problems are neither predictable nor convergent because they often possess unknown aspects. Additionally, they possess multiple solutions or solution strategies or often no solutions at all (Funke, 2012).

    You are not allowed to view links. Register or Login to view.
  • RE: [split] future video about AI, all input welcome /kg

    nablator > 27-11-2024, 09:07 PM

    Anything we think we know about problem solving limitations after testing ChatGPT 3.5 or 4 is already obsolete with o1, "designed to spend more time thinking before it responds. It can reason through complex tasks and solve harder problems than previous models."
  • RE: [split] future video about AI, all input welcome /kg

    R. Sale > 27-11-2024, 09:55 PM

    AI is a combination of amazing technological progress and good old excessive expectation. The individual balance between the two to be determined by personal sci-fi preferences. AI cannot solve the VMs at this time, but it can assist in working toward a potential answer. AI is *artificial* intelligence, not yet fully *actual* intelligence. That's why some of its output makes sense even when it has no connection to reality.
  • RE: [split] future video about AI, all input welcome /kg

    RadioFM > 28-11-2024, 04:04 AM

    Piggybacking on what others have just said: keep in mind we've seen huge leaps in AI research in these past years, and that some problems current LLMs may not be able to solve correctly, may very well be solvable a couple of models down the line. For all we know, ChatGPT-5 might be able to crack simple substitution ciphers in English. The thing is, those abilities may not translate to better odds at solving the VMS.

    I would argue that all current and future LLMs' 'reasoning' capabilities will continue to be bound by their underlying Language Model nature - they might be able to tackle tasks involving natural languages or some formal ones (algebraic expressions, predicate calculus, etc.) but will probably suck at translating or manipulating low resource languages like Voynichese.

    You can choose to delve into the theory behind LLMs and even go as far as making statements like 'If ChatGPT, Llama, Gemini and Claude all suck at [task X], theres no way in hell they'll ever crack the VMS' in your video. You can count on AI enthusiast to come back and ask the same question as soon as a new model capable of solving [task X] shows up though.

    I think it'd be best to give people a base test, something reproducible from the comfort of their PCs, so that they can check for themselves whether the latest chatbot can or cannot crack Voynichese.
    I think a good base test for future chatbots would be to give them a message coded using any of the advanced Steganographia technique or a pseudo-Voynichese cipher (I think we got a couple of those here in this forum). If an AI ever manages to extracts the plaintext from one of those without any hints at all, then I'd say it's worth looking into.