• BakedCatboy@lemmy.ml
    link
    fedilink
    English
    arrow-up
    209
    ·
    2 months ago

    I’m curious if it could solve the traffic light and crosswalk ones, I would try but I’m out of free image uploads from asking it to explain memes to test its cultural knowledge.

      • jballs@sh.itjust.works
        link
        fedilink
        English
        arrow-up
        39
        ·
        2 months ago

        I’m sure eventually someone will make a bot called something like ai-explains-the-joke that does this automatically.

        • WldFyre@lemm.ee
          link
          fedilink
          arrow-up
          33
          arrow-down
          1
          ·
          2 months ago

          I mean it still parsed the specific text in the meme and formulated a coherent explanation of this specific meme, not just the meme format

                • GreatDong3000@lemm.ee
                  link
                  fedilink
                  arrow-up
                  7
                  arrow-down
                  3
                  ·
                  2 months ago

                  Man the models can’t store verbatim its training data, the amount of data is turned into a model that is hundreds or thousands of times smaller than the original source data. If it was capable of simply recovering everything that it was trained on this would be some magical compression algorithm and that by itself would be extremely impressive.

                • Hexarei@programming.dev
                  link
                  fedilink
                  arrow-up
                  4
                  arrow-down
                  2
                  ·
                  2 months ago

                  They do not store anything verbatim; They instead store the directions in which various words and related concepts relate to one another in some gigantic multidimensional space.

                  I highly suggest you go learn what they actually do before you continue talking out of your ass about them

      • kromem@lemmy.world
        link
        fedilink
        English
        arrow-up
        17
        arrow-down
        6
        ·
        2 months ago

        The majority of people right now are fairly out of touch with the actual capabilities of modern models.

        There’s a combination of the tech learning curve on the human side as well as an amplification of stories about the 0.5% most extreme failure conditions by a press core desperate to feature how shitty the technology they are terrified of taking their jobs is.

        There’s some wild stuff most people just haven’t seen.

        • Miaou@jlai.lu
          link
          fedilink
          arrow-up
          13
          ·
          2 months ago

          I can just as well say that the screenshot above is the top 0.5% pushed by people trying to sell the tech. I don’t really have an opinion either way tbh, I’m just being cynical. But my own experience with those tools hasn’t been impressive.

          • kromem@lemmy.world
            link
            fedilink
            English
            arrow-up
            4
            arrow-down
            2
            ·
            2 months ago

            At a pretrained layer, the model is literally a combination of a normal distribution curve of capabilities.

            It can autocomplete a flat earther as much as a Nobel physicist given sufficient context.

            So it makes sense that even after the fine tuning efforts there’d be a distribution in people’s experiences with the tools.

            But just as the average person’s output from Photoshop isn’t going to be very impressive, if all you ever really see is bad Photoshops and average use, you might think it’s a crappy tool.

            There’s a learning curve to the model usage, and even in just a year of research the difference between capabilities of the exact same model from then to now is drastically different, based only on learnings around better usage.

            The problem is the base models are improving so quickly the best practices for the old generation of models goes out the window with the new. So even if there were classes available I wouldn’t bother pointing you to them as you’d just be picking up info obsolete by the time the classes finished or shortly thereafter.

            I’d just strongly caution against betting against the tech’s continued capabilities and improvements if you don’t want to be surprised and haven’t taken the time to look into them operating at their best.

            The OP post is pretty crap compared to the top 0.5% usage.

            • ThunderclapSasquatch@startrek.website
              link
              fedilink
              arrow-up
              1
              ·
              2 months ago

              It really does depend on what you ask and how, I can get some really nice music recommendations from Chatgpt but it also cannot comprehend GURPS skill rules, it’s actually funny how it manages to get it wrong a completely different way each time

        • The Picard Maneuver@lemmy.worldOP
          link
          fedilink
          arrow-up
          7
          arrow-down
          2
          ·
          2 months ago

          At the risk of sounding like a tech bro who’s desperately trying to secure funding: this truly does feel like a major leap in technology that is going to change the world.

          Anytime I hear it dismissed as “basically auto-complete”, I feel like it’s being underestimated.

          • kromem@lemmy.world
            link
            fedilink
            English
            arrow-up
            6
            arrow-down
            3
            ·
            edit-2
            2 months ago

            It’s not just underestimation, it’s outright misinformation.

            There’s so much research by this point over the past 18 months that there’s an incredible amount going on beyond “it’s just a Markov chain, bro.”

            It was never a Markov chain as that ignored the self-attention mechanism which violated the Markov property. It was just some people trying to explain it used a simplified description which went viral.

            Sometimes talking to people who think it’s crap feels like talking to antivaxxers. The feelings matter more than the research and evidence.

          • AdrianTheFrog@lemmy.world
            link
            fedilink
            English
            arrow-up
            3
            arrow-down
            2
            ·
            2 months ago

            Its kind of funny because autocomplete on phones is definitely moving in the direction of using LLMs. Its like it wasn’t true when people started saying it, but it will be literally true in a couple of years at most.

    • DarkCloud@lemmy.world
      link
      fedilink
      arrow-up
      30
      arrow-down
      4
      ·
      2 months ago

      Yes it probably can… CAPTCHAs don’t work based on your answers (many types you can answer wrong and still sometimes pass) - they work by tracking your mouses movements and timing and deciding whether they human-like.

      • Match!!@pawb.social
        link
        fedilink
        English
        arrow-up
        26
        arrow-down
        1
        ·
        2 months ago

        Why do i fail the “choose all images with motorcycles” challenges all the time then :c

          • ElderWendigo@sh.itjust.works
            link
            fedilink
            arrow-up
            2
            ·
            2 months ago

            I’m stubborn. I refuse to give the machine the answer I know it wants. And no, that overpass is not a bridge. Usually there is an option to skip or verify another way, This is when the captcha drops the ruse and it’s clear that the machine was just analyzing my mouse movements and response timings anyway to verify that I was behaving randomly in a human way. Still a better game than any of those in YouTube ads.

        • TexasDrunk@lemmy.world
          link
          fedilink
          arrow-up
          6
          ·
          2 months ago

          Are you the same guy who didn’t see me riding my motorcycle and tried to run over me? Because I think maybe you just can’t see motorcycles.

          No, that didn’t actually happen. I just wanted to give this person a hard time.

      • BakedCatboy@lemmy.ml
        link
        fedilink
        English
        arrow-up
        2
        ·
        2 months ago

        I don’t remember actually but I checked the file metadata and I have the template in my downloads folder next to this which has an exif tag of 2 minutes later with gimp metadata so I’m pretty sure I must have made it, which makes it a bit more impressive since I probably just sent it to friends privately and didn’t post it anywhere it could have been scraped for training.

  • jol
    link
    fedilink
    arrow-up
    132
    arrow-down
    1
    ·
    2 months ago

    Captchas te not meant to deter all bots. It’s meant to make it ever so slightly expensive that a mass DDOS attack would be extremely expensive to perform. Think like thousand sof requests per second, all being Captcha’d and how much it costs to run AI. It’s current not a feasible solution.

    There is cheaper AI that can solve Captchas though, and it’s only gonna get cheaper.

    • frezik@midwest.social
      link
      fedilink
      arrow-up
      31
      ·
      2 months ago

      It’s long been cheap enough that you can pay a call center full of people in a developing country to solve them for you. Going to be a while before AI is cheaper than that.

      Having used them to protect a few web sites from spammers filling up forms, they do cut down on the bullshit. This makes things more convenient for the people reading the information coming in from those forms, but I sometimes wonder if it’s worth the cost of everyone else having to pick out the bicycles in the picture.

    • Cethin@lemmy.zip
      link
      fedilink
      English
      arrow-up
      28
      ·
      edit-2
      2 months ago

      Also, captchas are meant to gather data to train on. That’s why we used to have pictures of writing, but that’s basically solved now. It’s why we now have a lot of self driving vehicle focused ones now, like identifying busses, bikes, traffic lights/signs, and that sort of thing.

      Captchas get humans to label data so the ML algorithms can train on it, eventually being able to identify the tests themselves.

      • AwkwardLookMonkeyPuppet@lemmy.world
        link
        fedilink
        English
        arrow-up
        5
        ·
        2 months ago

        Now it’s making me identify developed pictures from a photo negative. I’m not quite sure what they’re going to do with that training since computers can already perform that task.

        • TheOakTree@lemm.ee
          link
          fedilink
          arrow-up
          3
          ·
          2 months ago

          Also the “select the image below containing the example image above.”

          Like… we already have computers that can recognize image repetitions.

          • Cethin@lemmy.zip
            link
            fedilink
            English
            arrow-up
            5
            ·
            2 months ago

            So that’s almost certainly trying to gather data to defeat data poisoning. The other image is probably slightly altered in a way you can’t detect.

        • bitwolf@lemmy.one
          link
          fedilink
          arrow-up
          2
          ·
          2 months ago

          A common OCR tactic is to turn the image negative and bump the contrast to make text easier to recognize.

          It could be a precursor for that step.

    • Ballistic_86@lemmy.world
      link
      fedilink
      arrow-up
      15
      ·
      edit-2
      2 months ago

      I believe this is why Google, and a few other companies, have started using behavioral analysis to figure out if you are human. Did your mouse wonder around the page before clicking to verify? Did you come from another website as if browsing the web? What device are you using and have you used it on this site before? Are you logged into an account? I’m sure they use many more factors, but it’s something that would be hard to replicate with bot behavior on a consistent basis (for now).

      • jol
        link
        fedilink
        arrow-up
        1
        ·
        2 months ago

        Basically yeah, but that has been the case since almost the start of recaptcha. Fingerprinting just got so good to the point that the visual test was not even needed anymore.

  • Cistello@reddthat.com
    link
    fedilink
    arrow-up
    78
    arrow-down
    1
    ·
    2 months ago

    I have an extension which solves most Captchas for me It does it better than me which is why I use it

    • User@lemmy.ml
      link
      fedilink
      arrow-up
      4
      arrow-down
      1
      ·
      edit-2
      2 months ago

      Based on the serif of the font I’d go with P as the ai says

      • c0ber@lemmy.ml
        link
        fedilink
        English
        arrow-up
        3
        ·
        2 months ago

        based off the font i’d say “p”; but also mainly because the tail goes below the bottom of the other letters even after accounting for the distortion

    • problematicPanther@lemmy.world
      link
      fedilink
      arrow-up
      14
      ·
      2 months ago

      it’s the recaptchas that they should have trouble with. since it’s not just about finding the right picture, it’s also about the time between clicks, the way the mouse moves, etc.

        • problematicPanther@lemmy.world
          link
          fedilink
          arrow-up
          6
          ·
          2 months ago

          but us humans aren’t truly random, we probably behave in similar ways to each other, but also have individual ‘fingerprints’. like the time it takes between keystrokes, or the length of time we spend holding the button on the mouse down while clicking. we could probably come up with a way of identifying someone based only on that kind of data. what was i talking about?

          • snooggums@midwest.social
            link
            fedilink
            English
            arrow-up
            3
            ·
            2 months ago

            Not without enough context to know what time of day, if the person is ill, or a ton of other things that would make someone respond differently at different points in time.

            The anti bot stuff is going to be looking for too much consistency, which is hard look for on its own before trying to look for some kind of ‘fingerprint’

      • MonkderDritte@feddit.de
        link
        fedilink
        arrow-up
        7
        ·
        edit-2
        2 months ago

        reCaptcha never works for me. Probably something with thirdpartyisolation.enabled. Can’t snoop all the history and stuff.

      • lad@programming.dev
        link
        fedilink
        English
        arrow-up
        2
        ·
        2 months ago

        Yeah, and a couple of people I know who were consistently reported to be robots because they’ve been shown captcha too much and as a result solved it too well. Which in turn led to more captcha and improved solving speed. Well, you see the problem, I guess

  • nucleative@lemmy.world
    link
    fedilink
    English
    arrow-up
    38
    ·
    2 months ago

    There’s a program called Xevil that can solve even HCaptcha reliably, and it can solve these first gen captions by the thousands per second. It’s been solving Google’s v3 recaptchas for a long time already too.

    People who write automation tools (unfortunately, usually seo spammers and web scrapers) have been using these apps for a long time.

    Captchas haven’t been effective at protecting important websites for years, they just keep the script kiddies away who can’t afford the tools.

    • edgesmash@lemmy.world
      link
      fedilink
      arrow-up
      14
      ·
      edit-2
      2 months ago

      Captchas haven’t been effective at protecting important websites for years, they just keep the script kiddies away who can’t afford the tools.

      To be fair, keeping the script kiddies away has some good value. Whether that value outweighs all the wasted time and impact to sight/hearing impaired people is another discussion.

    • lud@lemm.ee
      link
      fedilink
      English
      arrow-up
      7
      arrow-down
      2
      ·
      2 months ago

      ?

      Lots of websites force capchas when on a VPN they don’t even have to be provided by Google. Rarbg for example forced a terrible captcha which I usually solved by using OCR with the OCR tool in powertoys. They letters were barely edited or fucked up at all.

      • SomeGuy69@lemmy.world
        link
        fedilink
        arrow-up
        21
        ·
        2 months ago

        Google now keeps you often in endless loop until you disable your VPN. This was never this bad.

        • no_name_dev_from_hell@programming.dev
          link
          fedilink
          arrow-up
          4
          ·
          2 months ago

          It’s extremely bad if you come from a country like mine, Iran, where we have to use VPNs religiously in order to circumvent censorship and it has become painful to Google anything especially when you’re not logged into your Google account.

        • johannesvanderwhales@lemmy.world
          link
          fedilink
          arrow-up
          5
          arrow-down
          1
          ·
          2 months ago

          They appear to have degrees of blacklist. Usually when this happens if I get a new ip it resolved the issue.

          Note that VPN users share IPs with other users and many of the people using the same IP may very well actually be doing malicious things. Not everyone uses VPNs for just “privacy”.

        • Quack@lemm.ee
          link
          fedilink
          arrow-up
          2
          ·
          edit-2
          2 months ago

          If you use the audio captcha it’s done in just one go. That’s been my experience at least after having been stuck in one too many endless loops with pictures.

        • lud@lemm.ee
          link
          fedilink
          English
          arrow-up
          1
          arrow-down
          1
          ·
          2 months ago

          Ah alright. I never use Google search.

  • ZwoofBlaf@sh.itjust.works
    link
    fedilink
    arrow-up
    36
    arrow-down
    1
    ·
    2 months ago

    Yeah captchas are done. Soon they will be easier to figure out for AI than for humans.

    This is why Sam Altman is doing his worldcoin thingy with the iris scanners. His idea: One iris (well, two…) is one real human. I’m sure this will be abused though and I absolutely vehemently don’t trust him with my biometrics so no way I will join that.

    I think what we should do is just get used to the fact that the internet now consists of humans and AIs. Learn to take things with a grain of salt.

  • MindTraveller@lemmy.ca
    link
    fedilink
    arrow-up
    32
    ·
    2 months ago

    Some disabled people have trouble with captchas, so these days you can download an extension where a robot solves the captcha for you.

  • Zeppo@sh.itjust.works
    link
    fedilink
    English
    arrow-up
    21
    ·
    2 months ago

    This is why a lot of sites have moved to something more complex than text, like the weird “rotate this to match” stuff that LinkedIn uses.

  • BluesF@lemmy.world
    link
    fedilink
    arrow-up
    16
    arrow-down
    2
    ·
    2 months ago

    The “puzzle” isn’t the test, the test uses your browser history, mouse activity, etc to identify you as human (or not). The puzzle is used to generate training data for ML models.

  • Voroxpete@sh.itjust.works
    link
    fedilink
    arrow-up
    14
    arrow-down
    1
    ·
    2 months ago

    There’s a lot of misunderstanding in this thread about how captchas work.

    What modern captchas examine isn’t actually your ability to solve the puzzle… It’s how you solve it. Things like mouse movements and how you type are big factors. So a bot would process for a moment, and then basically copy and paste in the answer, whereas as a human is going to type at a normal pace, often with pauses as they double check the details. Same goes for the click the tiles challenges. A bot will work through systematically, a human will bounce around, and their timings will be very different.

    • Lets_Eat_Grandma@lemm.ee
      link
      fedilink
      arrow-up
      12
      ·
      2 months ago

      Captchas have largely been solvable by machines at a rate higher than humans for a long, long time.

      It is very easy to train a model to behave like humans do by simply having a sample of human inputs.

      Here is an article from august 2023 covering how much better machines are than humans at accomplishing captchas of many flavors. Sauce