Photo by PhotoObject.neton Freeimages.com
GIANT DISCLAIMER: I’m not an AI expert by any stretch. I
know enough to be dangerous, and to protect myself.
GenAI IS NOT A SEARCH ENGINE!!!
This may seem like it’s targeted at businesses, but with the
consumerization / commoditization / democratization (i.e., everyone with an
internet connection and a cool idea for a picture of Mother Theresa having a
swordfight with Tiny Tim while riding a unicycle & playing a sousaphone) of
GenAI, there’s likely some nuggets in here that could apply to you personally.
This post is about generative artificial intelligence, which
is the latest shiny new thing in AI, with ChatGPT being the (arguably) most
headline grabbing example. There are other types of AI, some of which you
probably come across everyday. Examples include things like Netflix
recommendations (how they consistently fuck up their recommendations is pure
magic) and automated call attendants (press 1 for English, press 2 to be put on
hold until you give up - you should really be allowed to choose from a
selection of music – the stuff they force on you blows chunks).
Chris, have you actually used GenAI?
Why, yes, dear reader, I have.
For work I’ve used GenAI to:
- Create synthetic (test) data about people. I had to write the prompts so that no real SINs (Social Insurance Numbers – Canadian equivalent to social security numbers) or phone numbers were created.
- Create draft user stories for data migration. If you understand AGILE you don’t need me to explain. If you don’t understand AGILE you don’t want me to explain.
- Draft performance review comments for myself and the people that reported to me.
- Created deepfakes for training. Don’t worry, the people being faked knew it was happening, and we did not use any publicly available models.
For fun I’ve:
- Created an image to be printed on a friend’s birthday cupcakes. They were really good and went well with the whisky we were drinking that night. Nine expressions from Laphroaig, if you’re curious.
- Summarized and extracted themes from a list of my career accomplishments. It did a decent job, but I did have to do some editing. I also made sure to scrub anything that could identify me, employers, or clients.
- Created generally silly images for shits & giggles.
As I’m currently in job search mode, if I had any faith that
GenAI could do a brilliant job, I’d use the shit out of it to write cover
letters and target my resume. What a pain!
GenAI has the potential to be a real benefit. In my opinion,
one of its greatest potential benefits is to let people focus on tasks /
activities that are valuable. What’s valuable is really context dependent, so
I’m not gonna get into here.
Unfortunately, there are also some negative aspects to GenAI. However, if we work together and some people & organizations stop being dicks about it, we can mitigate the negative aspects. Those negative aspects include:
- Unintended bias – if the training data includes biases, those biases will be included in the output. Efforts need to be made to identify and remove bias. Not limited to GenAI, by the way.
- Hallucinations (no not the fun kind) – GenAI can on occasion spout garbage (white glue in pizza sauce anyone?). Fact-check the output.
- Resource consumption – GenAI consumes mind boggling amounts of power and water (for cooling). Yeah, I’m stumped on this one, but we need the tech vendors to step up and do something.
- Plagiarism, copyright infringement – if you’re using GenAI to produce something that could be mistaken for someone else’s work, stop it. If you’re doing it intentionally, you’re a dick.
- Deepfakes & NCII (non-consensual intimate images) – Yeah, this is just nasty and abhorrent and the people doing this shit need to be locked up.
- Misinformation / disinformation – If you’ve been paying attention lately, I don’t need to explain this.
I think the promise that GenAI holds is worth it, but we
really need the vendors, regulators, and legislators to move their asses and
impose responsible AI. We need consequences with real teeth for violators. I
believe we’ll get there because I believe that we’re generally good, kind, and
fair.
The GPT part of ChatGPT stands for Generative
Pre-trained Transformer, by the way. For a comprehensive description of
what GPT means, read this IBMarticle.
The generative pre-training bit refers to training a large-language model (LLM). If you want to know what an LLM is, go ask your favourite search engine. The basics of pre-training are:
- The model is pointed at a bunch of unlabelled data. This data could be everything on the internet, or it could be very specific corporate stuff.
- The model learns to detect patterns in the data.
- Based on what it has learned, the model applies the patterns to new data.
Easy, so far, no?
The transformer bit is more complicated. Again, some of the basics:
- A type of neural network
- Understands, interprets, generates human language
- Guesses (really good guesses) or predicts what the output should be, based on its pre-training.
Sometimes the GPT model you’re using messes stuff up, which
is how you find yourself adding white glue to your pizza sauce. ALWAYS
VALIDATE THE OUTPUT!!! This cannot be stressed enough. If bad shit happens
to you because you didn’t validate the output, that’s on you, not the AI.
Okay, that’s as far as I’m willing to go in explaining GenAI
without resorting to using it. The articles I’ve linked are all from IBM for no
reason other than my convenience.
At work, one of my primary jobs is to ask what business outcome is supposed to be achieved with some tool or technology. It’s no different with GenAI; what is your objective? Understanding the objective will answer two key questions:
- Is GenAI suitable for my purposes?
- If GenAI is suitable, which one should I use? They’re not all the same and I’m not going to document the differences.
NB: if you’re using a public model (e.g., ChatGPT)
assume anything that you put in (prompts, sample data) will be used to further
train the model. Make sure your inputs don’t include anything personal,
private, or proprietary.
Muy importante: six principles of responsible AI:
- Fairness
- Reliability & safety
- Privacy & security
- Inclusiveness
- Transparency
- Accountability
If you’re curious about what the principles mean, a quick
internet search will satisfy your curiosity.
This isn’t an attempt to dissuade anyone from using GenAI;
just think about how & why you’re using it.
GenAI cannot produce art. Only humans can produce art because
they’re, uh, human.
What are your thoughts about GenAI?
What’s your experience with GenAI?
Be great today, be better tomorrow.
Cheers!