Write: The Rise of the AI Assistant

Welcome to the age of the AI Assistant!  You may not have known you were there, but you’re there.  You may not be using one, but most of you are.  If you’re not using one, you probably will and you are almost certainly underestimating the impact of this coming change in how we interact with the world.  I hesitate to call it revolutionary, but it kind of is and it’s kind of amazing!

Now, this comes from someone who is eternally excited about change.  I grew up reading Science Fiction and Fantasy, so imagining a world that’s coming or might eventually be is kind of second nature and I am happy to embrace the new.  I once told someone that I’ll know I’m nearing the end of my time when I stop caring about the new changes in our world.  

I also used to say that I will be concerned when a new technology or application comes along and I just don’t get it or why it exists.  I had to deal with this one because that has certainly come along (*cough* Snapchat), but just because something came along and was temporarily popular (and I didn’t get it), doesn’t mean I’ve stopped understanding technology, but it does mean that things are being created that are, decidedly, not for me.  And that’s okay.

The current/prior age of AI Assistants includes things like Googling what the last movie Vin Diesel did or asking the current time, temperature or weather.  They could parse a bit of natural language to determine what you were likely asking, Google (to use the kleenex equivalent of an Internet verb) the question and give you the results.  There may have been some basic ability to follow up a question with another question (Follow-up Mode on Alexa, for example), but essentially it required you, the human, to string together responses to figure out if you got your question answered.  Additionally, it was pretty clearly limited to doing a better job of looking up information on the Internet than you were likely to do and rank-ordering the results based on what it had determined was the quality of the answers.  It was fun, it was kind of neat.  I got just a tickle of excitement, but it was by no means the computer-based Assistant I imagined from a life of reading Science Fiction.

2022 was kind of a blow-out year for AI assistance and AI Assistants and I’m going to just talk about a couple and use a recent interaction as a demonstration.

2022 saw the launch of several AI-aided art generation programs.  Such programs as Midjourney, Stable Diffusion and DALL-E 2.  You can find out more about these and others here.

Most of these work by feeding them text prompts about what you’re looking for and the AI will generate some responses and you can tune them or have them focus on a particular image or direction.  You can also tell it to generate art in a particular style, resolution or genre.

Now, these AIs do come with some caveats.  Since they are effectively trained on a large block of imagery that was scraped from the Internet, the AIs have effectively been trained by absorbing other artists’ art and styles, so in some cases, the results can and will look like something generated by a particular artist which raises some troubling questions about the generated art.  In fact, AI-generated art has been banned in several venues.  

Is the art “stolen”?  In most cases, not really.  I would assert that most any artist absorbs the art and styles from everyone they have looked at, studied and perhaps even emulated.  But, in most cases, they eventually land on their own unique style, even if it’s derived from someone else.  AI “artists”, though, don’t yet create their own style so much as mix together all the inputs from all the art they’ve been trained on.  Is it derivative?  Absolutely.  Specifically, because it’s derived from the work of other artists.  Still the results are interesting and very worth exploring.

The other exciting new release on the AI Assistant front is ChatGPT.  ChatGPT is a (beta) product from OpenAI.  OpenAI states as their mission, “OpenAI’s mission is to ensure that artificial general intelligence (AGI)—by which we mean highly autonomous systems that outperform humans at most economically valuable work—benefits all of humanity.

We will attempt to directly build safe and beneficial AGI, but will also consider our mission fulfilled if our work aids others to achieve this outcome.”  It’s also taking a look at their Charter.

Now, OpenAI is also a company, a corporation whose effective charter will always be to make a profit for their investors and stakeholders, but I still appreciate the stated sentiments.

So, the big advancement in ChatGPT is roughly two-fold.  First, it has an even more advanced ability to parse and respond to natural language questions and to reply using a similar natural language.  It’s like chatting with a (nearly) human expert who has access to a large portion of the information on the Internet.  Second, it maintains state.  This means that if you ask it a follow-on question in a session, it recalls what you were talking about in the previous question.  This means the AI recalls the prior interactions and will take those into account in the next answer.  It’s kind of amazing.  It really is like having the ability to have an ongoing conversation with an expert assistant.

The version of ChatGPT that we are able to play with today is based on GPT3.  GPT, by the way, stands for Generative Pre-Trained Transformer (GPT).  That means it’s been (pre-)trained on a body of knowledge.  GPT3 was trained on 175 billion parameters and it was 100x bigger than GPT2.  GPT4 is expected to have 100 Trillion (with a T, so an order of magnitude or roughly a 1000x larger) parameters.  From this article, we find that “GPT-1 was trained on 4.6 gigabytes of data, GPT-2 was trained on 46 gigabytes, GPT-3 was trained on 750. GPT-4 will be trained on considerably more, a significant fraction of the internet as a whole”.

ChatGPT, like the discussion about AI-generated Art, does not come without some red flags.  Obvious things include a kid who can now generate 500 words about the themes in Catcher in the Rye in 10 seconds, including references and quotes.  So, there will be an arms race in both generating materials using AI and then catching that something was created by an AI.  Also, it’s not all that hard to get the AI to generate answers which sound correct, but are, in fact, wrong.  And it will include references to fictional sources to “prove” its point.

So, there is clearly work to be done, but every significant technological advance presents both huge upside opportunities and significant risks.  Given the invention of the gun, it can be used for both good or evil, the intention is provided by whoever wields it.

I thought it might be fun to end this with a recent foray into using these tools and being both amazed and a little startled at what these assistants can do and thought I’d share them with you.

For context, I am currently running a role-playing game set in a Weird World War 2 based on the Savage Worlds rule system.  So, there may be zombies, zombie Nazis, magic, occult and other nonsense.  As a warm-up, I asked the AI to generate a character for me that could operate in my game.  

Here’s how it went:

This is pretty good. But, my players are Seasoned (more experienced), plus I decided I wanted a French female medic who has uses Powers (magic) to add to her abilities.

Wow.  For some context, the layout of the information matches what I would expect for Savage Worlds.  Additionally, to make a character Seasoned means they have advanced from a starter player and will have stats that reflect that experience.  To give her Powers means that in our game system, Savage Worlds, she has some magical powers and the AI was able to choose powers that would make sense for a medic.  One of the steps in character creation is to create Hindrances that would make sense for a character, the AI made her Cautious and Loyal, which both exist in Savage worlds and make sense.  Finally, note the character description, which references the Hindrances as well as her Powers and note that as a French native, she speaks fluent French.  This is pretty amazing!

But, let’s try another.  Let’s create a monster for our Party to fight!

A few observations.  First, it gave our monster a great name and took the prompt very much into account.  It gave our monster Natural Armor and a bonus of +4 and remembered that in calculating its Toughness: 16(4), the four being from Natural Armor.  It’s “massive and imposing” and is Sized appropriately (+4).  The addition of the Improved Frenzy and Swallow Whole is kind of genius.  I would probably not have come up with something like that had I been creating a monster on my own.  Finally, note the description which again takes in to account the features it added earlier about the monster.  So good.

Next, I went to Midjourney and asked it to imagine the creature above using a slightly edited version of the description.  This is what might terrorize my Party in our next Session:

In this case, I’d give Midjourney a solid B because it has the sense of a sewer but has an open sky behind and it’s fairly immense.  But, still evocative and a nice image to use to set the stage for an encounter.

For brevity, I will say I asked ChatGPT to create an encounter using this monster and it did a really good job with that as well.  I may be out of a job.

The final example involved asking it something which I was sure it would either struggle with or fail entirely.  The next session involves infiltrating a POW camp in this Weird WW2.  So, on a lark, I asked it to create an NPC (Non-Player Character) based on Schultz from Hogan’s Heroes (a television show from the 60s if you don’t happen to be old).  Here is how that went:

This one made my jaw drop.  “He is not very observant, and can be easily fooled by the prisoners.” I mean, come on!? I’m willing to bet that whatever the training data was, it did not include having the AI watch episodes of Hogan’s Heroes.  Even without that, it clearly knows what the show is, who the character is and how he relates to the other characters including the main character of the show, Hogan.  Then it used that information to create this version of Schultz who has all the appropriate attributes and characteristics that fit very nicely to the character on the show!  It was able to take its understanding of what a Savage Worlds character was with its understanding of Hogan’s Heroes and create a character for me.  All in seconds.

Now, this isn’t the Turing Test, but I’ll tell you what: It’s a peek into a future where I have a very savvy assistant who I can have conversations with and who can help me create more things more quickly.  It’s a bit like having another person in the room to bounce ideas off of.  That Sewer Leviathan is a great monster!  And one I can use happily and save myself time to focus on other things.

This is really just the smallest peak into the coming next generation of AI Assistants.  To clarify, this is not the robot uprising.  These AIs need to be led by a human towards a goal, but they do have the ability to use a volume of information that we would find unimaginable to make connections and create things that would be very difficult for us alone.  It’s a pretty cool time to be alive to see these kinds of advances.

Note: The blog post image at the top was generated by AI on MidJourney.