AI Is Dumb

—

by

ljndawson
in

I don’t have a lot of patience for the claims being made by companies who are heavily investing in AI. It definitely has its uses – image recognition, mining through large repositories of text, etc. But one thing that AI has not yet perfected is decisionmaking.

Computers – even sophisticated ones – are stupid. They can only operate on the information they’re given. They completely lack context, and rely on…you guessed it…metadata to help interpret the tasks they’re being asked to perform. They are really bad at making sense.

We can laugh at a lot of AI goofs – like generating images of people with the wrong number of fingers or two left feet – but a lack of context can be extremely dangerous. Timnit Gebru, in her paper “View of the TESCREAL Bundle” cites a number of mistakes that bots have made, such as encouraging people to eat crushed glass, exploring the benefits of suicide, and other horrific examples. Computers do not have a sense of right, wrong, or morally ambiguous – they have a sense of yes and no. On or off. Zero or one.

Which is why throwing bucketloads of content at an array of bots is not going to generate much that’s useful. Bots are not, ultimately, that good at extracting meaning. Or intent. Or mood. These are human functions, born of emotions and philosophies.

Metadata provides direction and context – say, for example, the bot encouraging people to eat crushed glass had run across a category of “satire”. That would help! (There’s no excuse for the suicide suggestion.)

AI spits out what it’s been given. Gebru’s paper is a great summary of why we should be questioning what it’s been given. So many LLMs are predicated on faulty (or stupid, or evil) assumptions. And without a metadata framework to serve as a guardrail for AI inputs and outputs, we just have a lot of big dumb machines sucking up natural resources and not even citing their sources.

The fundamentals of computing will always hold: Garbage in, garbage out.

Discover more from New Virago

Subscribe to get the latest posts sent to your email.

Comments

7 responses to “AI Is Dumb”

  1. delicatelykitten0650091f8e Avatar
    delicatelykitten0650091f8e

    Brilliant observation and summation of A.I. inadequacy

    Like

    Reply
  2. Nick R Avatar
    Nick R

    I agree with the context that it’s all about context, metadata, and how you interact with the model, but I believe there is something worthwhile to tease out and has helped me contextualize this shift and what makes today’s AI models different from promised in the past…

    The code – or at least the process – behind how LLMs (and now LCMs) work is very similar to the way a human brain works – pattern recognition at different levels, prediction, reward for being correct. The same way a child wants to touch a hot stove, eat dirt, and yes even chew on glass – AI models will occasionally do those things as well.

    Chat GPT first made it to the public in the end of 2022. While AI (especially this concept of learning for AI) is significantly older (my work was in the late 90s). So, for simplicity sake, lets call it 3 years old.

    It’s reached a level of complexity that – like a human child – we don’t really know how to best teach it, guide it, etc. Now, we’re spending a whole bunch more money that we throw at child psychology, so it’s learning and growing much quicker than the field that encompasses Baby Einstein.

    Where I’ve seen the latest models fail is when we fail to provide them enough context. But, I’ve found that if you provide enough context, and give enough information, it’s absolutely stunning what responses it can have…

    If anything, I’m more scared of the raw power that is now available to us. It’s revolutionized the scammer market. It’s a technology that needs to be learned to be respected for it’s potential.

    And, to conclude with the same line but a completely different direction – Garbage in, Garbage out. If we disrespect this tool and feed it all of our hatred, feed it all of our anguish, feed it our lies, and feed it all the worst of humanity, it will become not only a powerful mirror, but a manipulative master unlike we have ever seen.

    Like

    Reply
    1. ljndawson Avatar
      ljndawson

      Context is EXACTLY what’s been missing, historically. And why the metadata that can provide that context (and scoping) is so critical – what’s scary to me is that it’s being ignored.

      Additionally, the basis of many common models is, as Gebru points out, eugenicist. She cites a lot of excellent sources proving this. And I take her very, very seriously. Imagine a racist, misognynistic, anti-disability three year old. A psychopathic three year old.

      Garbage in, garbage out, indeed. I agree with your final statement wholeheartedly.

      Like

      Reply
  3. jcsimonds Avatar
    jcsimonds

    What’s even dumber is the corporate response to this. I’m sure C-suite dolts are thinking they can fire all their workers. But the credulous rah-rah you hear on business TV is so appalling. They have no idea what it is good for, just that it’s “disruptive” (which is now my least favorite word in the language, and I immediately distrust anyone who utters it). These morons get paid million$ to think this tripe? Arg.

    Like

    Reply
  4. What I Saw at the Evolution – New Virago Avatar
    What I Saw at the Evolution – New Virago

    […] Where will the next disruption come from? […]

    Like

    Reply
  5. Prey – New Virago Avatar
    Prey – New Virago

    […] talked about this before. But it gets […]

    Like

    Reply
  6. Ozymandias – New Virago Avatar
    Ozymandias – New Virago

    […] my friend Dalia calls “the human in the system.” Because, as I’ve said before, AI is dumb. It can only work with what it’s […]

    Like

    Reply

Leave a reply to Ozymandias – New Virago Cancel reply