#1291. Slightly Off Topic: AI and photography blogs

By pascaljappy | News

Jun 07

Before becoming a photo-loving marketer, I did research in AI. My PhD in Machine Learning is – literally – from another century, but my interest in the topic never waned. And seeing AI spread to our hobby, as well as to many other areas of our daily lives, has made me dive in again.

 

So, you can expect posts about AI and photography in DearSusan, in the near future. Plus, I have just started a new blog on Artificial Intelligence. Yipee.

It’s called “Highway” or, rather Hiw.ai, because unlike so many observers, I’m very bullish about the technology and what it can add to our lives 💪

The two blogs will reference one another on occasion, but will live separate lives, with separate subscription bases and intended audiences.

 

DearSusan will remain a collaborative, hobby-centric creative photography blog, in which we will occasionally discuss the uses of AI in photography.

Hiw.ai is intended to be more how-to and professionally-oriented in its approach, but with non-specialists in mind. There are already multiple excellent technical blogs and websites online, dealing with Artificial Intelligence, but far fewer that can help readers outside of the technical sphere understand the mechanics of AI and Machine Learning, and how to use them to their benefit, in their personal life.

In hiw.ai, I want to provide an intuitive, yet deep, dive into how AI works, so that anyone can make informed choices about technologies to use, form their own opinion about our future, and cut through the polarized drivel often found in non-specialized media.

 

The two blogs will work together, in spite of their different target audiences, as I believe hiw.ai can be interesting to most DS readers. For instance, I hope to interview the developers of AI software used in photography. The uses will be described on DS, but the technical details will be on hiw.ai.

And if DS mentions that an AI app is generative, the definition and functional explanations will be linked to on hiw.ai.

As the email linking to the present post mentions, I have created a separate mailing list for the two blogs. This post and maybe a couple more in the future, is emailed to you using the hiw.ai list. I hope that’s OK !! If you are not interested in learning about AI, please just click unsubscribe on any of those emails and you will still get the DS messages. After the first 2 or 3 messages, I will remove from the list all members that have clicked on none of the included links, so you will stop receiving those messages if you are not interested and do not unsubscribe by yourself.

And if you do like what you read on hiw.ai, please share! As with DS, I will not create social accounts or advertise. The blog will grow or die based on its success with its readership, which is expressed in comments and shares. If it is useful to you, you will share and it will continue. If not, well, it’ll just dwindle 🙂 Such is life.

 

What’s on offer so far?

  • So, what is ChatGPT and how does it work? ChatGPT is the current darling of the media. It makes content creation fast, but unreliable. And its ability to converse makes anthropomorphism natural and tempting. But it’s just a piece of software. This post explains the meaning of its name, and how it works, based on in-depth analogies with relatable, everyday, concepts.
  • Ada Lovelace and Artificial Intelligence. The mediatic fuss may be recent, but research on AI has been going on for centuries. And that history is both enlightening (!) and fascinating. One of the most important figures in the development of AI, equally so than Alan Turing from my perspective, is Ada Lovelace (1815 – 1852), the prodigy daughter of Lord Byron. Many biographies about her already exist, so this post focuses on why I think she was so successful and important, and why AI opens up huge avenues to non-specialists today.
  • Should we fear the emergent properties of Artificial Intelligence? Large AI models can do things they were never (explicitly) programmed for. For instance, ChatGPT can code, and create functional software. Yet no one taught it to! This is the most enthralling – and scariest – aspect of modern AI. Here is an explanation on why, and my take on whether we should be afraid (spoiler: I’m very bullish about AI 😉 )
  • I made cookies and an analysis of AI ethics with ChatGPT. Tongue-in-cheek title, but serious content. AI produces shape-shifting software that is difficult to assess in usual (specification-centric) ways. It’s often better to test for yourself, to see what works and what doesn’t. In this post, I try two very different types of tasks (finding recipes, and performing a theoretical analysis of a complex topic) and highlight the essential findings that can apply to other contexts.
 

As you can see for yourself, the blog has no logo, no menus, no contact form, it has received no graphical design work, it doesn’t have a subscription form, it is as bare bones as they come, and I haven’t really figured out the editorial tone and line, yet. All of this will follow if you find the contents interesting 😉 So do let me know, over there or here.

The current articles are all longform. All future ones will be written (video is fantastic for entertainment, but we all learn much better from text). But I might experiment with a mix of very short posts and longer ones. Again, please let me know what works for you.

Back to photography for now. The next post will be our collective “Best of May” post 🙂 Cheers.

PS: Unrelated but related. The Northrups have recently revealed what they feel is Sony’s game for the future: integration with Android (hat tip to Philippe for sharing this). We’ve been suggesting something like that for nearly a decade on DS, and think this could be a great move for Sony. One that will include … AI.

 

​Never miss a post

​Like what you are reading? Subscribe below and receive all posts in your inbox as they are published. Join the conversation with thousands of other creative photographers.

  • jean pierre (pete) guaron says:

    ” the blog has no logo, no menus, no contact form, it has received no graphical design work, it doesn’t have a subscription form, it is as bare bones as they come, and I haven’t really figured out the editorial tone and line, yet. All of this will follow”

    Sounds like a job, already, for AI.

  • Kristian Wannebo says:

    Pascal,
    Great Stuff!

    And I mean both your plan with that blog and your first four posts!

    [ hiw.ai does seem to have a contact form now, but to be on the safe side I comment here.]
    – – –

    ( First, a minor typographical detail, but maybe irritating for someone whose eyesight is less than perfect.
    The purple links (not the red ones) in the text have very low contrast to the background, both in light and in dark reading mode, at least on my iPhone 12 mini.)
    – – –

    Now just some random (hm) thoughts…

    On
    “So, what is ChatGPT and how does it work?”:

    Very explanatory!

    I remember reading about the beginnings of neural networks in the (sadly long diseased) BYTE magazine in the 1970s. It was about the possibility of improving OCR, and of generally processing simpler patterns.
    Later, about the time of rumours of the first hardware neural network chips, articles of neural networks stopped appearing in the few magazines I followed.
    ??.

    So these, and future, articles from you, are just what I’ve long wanted to read.
    And your explanations are better and your descriptions more vivid and to the point!
    Especially (if I understand them rightly) about the newer more advanced structures.
    – – –

    Pascal,
    *Can* one believe Google Translate in this case,
    *does* it “know” any Latin? –

    – I never learned *any* Latin, except the Swedish palindrome
    “Ni talar bra latin.”, meaning
    “You speak good Latin.”
    🙂 !

    ( On the other hand, Google Tr. seems to make a good job of the (unfinished?) text at the very bottom of the blog page?)

    Anyway, that Google translation is just lovely…
    – and, if anywhere close to correct, quite a nice read – wherever the original comes from…

    Towards the end of the 1960s (in Sweden) a computer generated poem collection was published – selected from a rather larger number – that was readable, barely.
    And one author published a novel (Bröderna Cartwright) – *most* probably human generated – where the story more or less randomly shifted every few sentences. It was considered to be very “modern”.

    Speaking of Google Tr., with the song text about ‘ticians by Georg Kreisler that I pushed through it (my comment in #1290), I was surprised at the result – compared with the lot of gibberish it would have included only a few years ago.

    [ One serious grammar fault,
    Two (understandable) “it” instead of “he”,
    One colloquial slang word understandably misunderstood,
    And a few badly chosen synonyms.]

    So I conclude that some of the more advanced neural network software you describe is already included.
    – – * – –

    On
    “Ada Lovelace and Artificial Intelligence”:

    A *very* good read, I think, Thank you!
    And you do include a lot of interesting detail that isn’t usually mentioned in texts about her!
    – – * – –

    On
    “Should we fear the emergent properties of Artificial Intelligence?”:

    ( Re. *human* translation:
    I remember my father telling me (in the 1960s) about Information theory research on human language. I was surprised how very redundant it is. And yet in experiments with translations back and forth with a new translator every time, even the context of the text faded rather quickly and was mostly gone after just a few iterations.)

    Pascal, for me a few questions emerge from your text.
    I understand that ChatGPT was exposed to much text in different languages.
    If *general* Internet material was used, that would be easy to arrange and also the same content could easily appear in different languages. Also part of the text would include mathematical “text” and computer code examples.

    With *very* much such material I guess an enough advanced network could “learn” also translation, math and coding without it really being emergent – or is this what “emergent” means?
    – – –

    In Swedish public service media AI and its dangers have recently been discussed a few times in programs of a half to one hour.

    But never as thoroughly as in this post!
    – – * – –

    On
    “I made cookies and an analysis of AI ethics with ChatGPT”

    Bon appétit, Pascal!

    Re.
    those fake articles ChatGPT claimed to have found.
    Is there perhaps an analogy to the far simpler systems of picture recognition? I’ve read of cases where a totally unrecognisable image had been “identified” as (let’s say) that of a dog…

    And I wonder if it learnt such cheating because its input perhaps contained also fake stuff?

    But the structure of both text and content in the ChatGPT’s answers you give impresses me.
    It certainly already must be hard to recognise such AI stuff when used as if human!
    – – * – –

    As I said , Pascal, great stuff!
    I’m very much looking forward to what’s coming!

    God speed!

    – and there are still your good links to read…

    • pascaljappy says:

      Thank you Kristian!

      Where did you find purple links? Everything is black or yellow on hiw.ai. Or at least is supposed to be 😉 It’s the template’s default colour and I don’t like it at all and will soon change that.

      Here are some tentative answers to your questions.

      Google translate doesn’t know latin. Neither does GPT. They may have memorised patterns of text that, when assembled; can create latin text. But they have no clue what is written.

      AI went through several “winters”, when the real abilities didn’t meet the hype and unrealistic expectations. That’s why neural networks disappeared from the scene for a while. I think we may be on a verge of a new one and this is the topic of my next post.

      I’m not sure anyone on the planet knows for sure how GPT can traslate this well. It’s called an emergent property because it wasn’t part of the programming. GPT knows how to disassemble text (a question, for instance), associate it with a set of probabilities that allow it to reassemble new text, using the most likely patterns. I’m guessing something like “how do you say XXX in French” is easy to map to the French translation. But a request for translation hidden in a more complex sentence might not actually lead to a translation. It would be interesting to try.

      The citations are indeed like finding a dog in a picture. It assembled a dog using an ear, a nose, eyes, paws … and the result looks like a dog. It’s just not a real dog that lived and present in an actual photo. Likewise, GPT assembled an article using authors, a publication, a date, a title and a publication number, which when assembled very much looks like a citation, but doesn’t correspond to one that ever existed. The same thing happened to that lawyer looking for precendents. GPT made them up, and they were very believable, but none were true.

      Caveat emptor 😉

      Cheers

      • Kristian Wannebo says:

        Right, Pascal,

        those links *are* yellow – i only checked the color in my dark mode, where they look purple.
        And you seem to agree that they need to have a higher contrast to the background.
        🙂
        – – –

        You write
        > “Google translate doesn’t know latin. … have no clue what is written.”
        ( As you also explain in your post.)

        With “know” Latin I meant just having the capacity to output text that to us appears to be Latin.
        ( Apologies, I ought to have been clearer.)

        Re.
        that Latin text.

        I meant my text more as a joke, because that Google translate output for the Latin-looking text was so funnily queer.
        I just hoped to hear how that text came into being.

        So let me ask:
        ( – remember, I know nil Latin.)
        Did Google Translate not “know” (as above) – 😉 , 🙂 – enough “Latin” and so totally “failed” me?
        Or was ChatGPT “guided” to output text looking like Latin – but failed?
        Or was such an output expected, as input in Latin was known to have been too meagre?
        Or was it “guided” to output random text that would look like Latin nonsense?
        [ Sorry if I’m now too verbose… 🙂 ]
        – – –

        Re.
        trying more complex questions on ChatGPT.

        A perhaps too simple one:
        The grammar fault Google translate made for me (it became, of course, different when I added punctuation to the lyrics).

        “Ein Philharmoniker ist ein Staatsmusiker, der Pension kriegt, wenn er nicht mehr gut gefällt.”
        became
        “A philharmonic is a state musician who gets a pension when he no longer pleases him.”,

        instead of
        “A philharmonic is a state musician who gets a pension when he no longer pleases.”.

        Is, perhaps,
        “… when he is no longer pleasing.”
        better English?
        Or?

        ( How would ChatGPT make it?)
        – – –

        Thanks
        for your answer on lies!
        And on emergent!

        I thought – as Alice in wonderland might say – that AI theories & knowledge had developed much further…
        I imagined knowledge, or at least ideas, about different levels of complexity in GPT’s capacity to produce emergent properties – so I wondered about how high math it could “produce” in relation to how high a level of logic was *explicit* and how much *implicit* in the input texts.
        – – * – –

        And when some future AI machine perhaps “produces” a *new* proof of Gödel’s theorem and then begins to give further as yet unknown examples of mathematical postulates that can neither be proved nor disproved – *then* I’d be impressed.
        😉 , 🙂
        But – unless simple enough for a human – how could they be verified, possibly that could be far too complex!

        Or will the then over next AI be able to produce proofs as simplified as logically possible?

        We’ll anyway have to learn to build also a truthful AI machine…

        Mathematicians would greet the new explorative possibilities!

        Cheers!

  • jean pierre (pete) guaron says:

    Ah, latin once again.

    Having spent much of my life learning various languages, latin was actually high on my bucket list. Because so many other languages, or words in other languages, are derived from latin. And anyway, being the strong-willed, fiercely independent bratty little child that I was, I’d secretly decided I was going to read law when I got as far as the university, and in those days, where I was living at the time, you couldn’t enrol at the Law Faculty unless you matriculated in latin (amongst other equally irrelevant things).

    But I’ve never come across that palindrome before, Kristian – so I presume it only works in Swedish.

    • Kristian Wannebo says:

      🙂 !

      In my 9th school year I chose the natural science branch – modern or classical languages were the other choices.

      But the correspondence Latin course I began soon petered out…

      I guess I wasn’t strong-willed enough.

      Cheers, Pete!

  • >