The next page in the book of AI evolution is here, powered by GPT 3.5, and I am very, nay, extremely impressed

According to the AI it does not have an internet connection and it cannot learn from the display interactions we are engaged in. It even lacks short term memory. The AI emphasizes that it is an application program that was produced by a much larger, more complex system. It does not have access to new data and is not engaged in learning.

I fed it the gag “Mary had a little lamb and the doctor fainted”. We had a lengthy exchange on the meaning of ‘had’ It agreed that ‘had’ could mean birth, but could not connect that meaning to doctor. It acknowledged that this was because it only evaluates the morphology of text, not the meaning.

As others have pointed out, the AI does assemble clear explanations. I asked how text was presented to the software neural nets and how they dealt with temporal issues. The responses were exactly on point. That solves a real problem. I’d rather talk to this AI than a call center.

I wonder how many “writers” of the “I could be a writer, I have great ideas, I just need someone to do the grind of typing all the words out” type are already half way through ‘their’ masterpieces?

You feed it an opening situation. (Manly Private Investigator is swilling booze in the local dive when a buxom blonde plonks herself down at his table.) Then take what it gives you and add a disruptive tweak. (Suddenly a man with a heavy German accent shouts threateningly at her.) And so on, pushing it towards whatever genre you have in mind, scene by scene, twist by twist. How long before you have a complete novel you can put up for sale on Amazon?

That’s what I was thinking. It definitely could be used at this level for brainstorming. It lacks the poetry of words , the feel for rhythm (as in verse; it does a decent job with prose, I think. It knows little of actual poetic form beyond how to group lines and roughly how long they should be. Seems to like tetrameter.), creating unexpected juxtapositions of words, involved wordplay (though it seems to do some) etc. But it does a reasonable job of coming up with something coherent that you can use as a jumping-off point. I would say this is like writing from a fifth grader or so (with fewer spelling and grammar errors.) Or you can actively go against the output, deeming it too cliche and obvious if an AI right now can come up with it, and try to take different turns in the plot. But I’m absolutely shocked at how far along we are already.

Some questionable slant rhymes in there.

Prompt: Rap about how a bill becomes a law

Once upon a time in the land of the free
A bill was born, with liberty
It started out small, just an idea
But with hard work, it aimed to become real

First, it was introduced in Congress
Where lawmakers would discuss
Whether it was needed, and if it was fair
They’d debate and amend, and sometimes they’d swear

If it passed in the House, it would move on
To the Senate, where it would be heard by the con
If it passed there, too, it would go to the boss
The President, who would sign it with a flourish

But if the President didn’t like it, he could veto
Then Congress would have to override
Or the bill would die, just a memory
Of the democratic process in America’s history

But if it made it through, it would become a law
Enforced by the courts, for the good of all
So let this rap serve as a reminder
Of the journey a bill must take to be a winner.

If you see something you like, immediately copypasta it out into a file so it won’t get lost.

It knows what an “eye rhyme” is by definition, but fails miserably to create a poem using nothing but eye rhymes.

So it does have some knowledge or data base available to it based on a few things it has come up with for me. Here is an example. I told it to rewrite the classic Abbott and Costello routine Who’s on First in the style of the sitcom “Newhart”. This is what I got:

[Larry and his two friends enter. Bob is at the front desk and hangs up the phone as they enter.]

Larry: Hi, I’m Larry and this is my friend, Abbott and my other friend Costello. We’re looking for our first baseman.

Bob: I’m sorry, who’s on first?

Abbott: Yes, that’s who we’re looking for.

Bob: I mean, what’s the name of the player?

Abbott: That’s just it - we don’t know. That’s why we’re here.

Larry: We were hoping you could tell us who’s on first.

Bob: Well, let me see if I can help. Is it a man or a woman?

Abbott: It’s a man.

Costello: Or a woman.

Bob: Ah, okay. What’s their name?

Abbott: That’s what we’re trying to find out – they’re on first.

Bob: I see. Well, why don’t you just ask around? I’m sure someone in town will know who’s on first.

The thing actually seems to me to have captured the voice at least of Bob Newhart in this and to some degree the other characters. I had it do the same thing but changed from Newhart to the style of H.P. Lovecraft. Decent results for that as well.

That doesn’t surprise me. It seems to know the definition various elements of prosody, but if you ask for specific uses of it, it doesn’t seem to be in its training. I can get it pumping out something akin to an English sonnet in form, but the meter is completely wrong. I can ask for blank verse, but it won’t produce anything like blank verse. Anything too specific, and forget it at this juncture. Asking for it in style of various poets I like (even Chaucer) has been completely ignored, so far as I could tell. Maybe very obvious ones like e.e. cummings mentioned before will work, but I doubt you can refine it too much. It had no idea how Seamus Heaney wrote, for instance. (But asking for an alliterative sentence about, say, cars, it does fine.) That’s asking too much of it. Give it several years. Like I said, I’m just astounded we are where we are in 2022.

I might just be one of them. I have some intriguing ideas for a potential novel, but always lacked the writing skills and more importantly, the drive to get better at writing, so my manuscript perpetually remains in an unfinished state.

When this AI improves with more training data (which seems inevitable to me), I can quite plausibly have it ghostwrite my novel by feeding it the outline of each chapter and then gluing the outputs together. I’m curious how competent ChatGPT would be at keeping track of the characters, their motivations and personalities, plus the plot / tones / overall theme of the story.

At its current state of progress, one thing’s for sure - procrastinating high school and college students collectively jizzed in their pants, while Upwork/Fiverr ghostwriters had a collective heart attack.

If your story gets boring, have someone show up with a gun. Still an uplifting story / ending.

Prompt: Rewrite the oliver twist story where oliver twist has a gun

Oliver Twist grew up in a world of poverty and crime. He was born into a life of poverty, and he struggled to survive in the harsh streets of London. Despite his tough circumstances, Oliver was a kind and brave boy who never lost hope.

One day, Oliver’s luck changed when he stumbled upon a hidden stash of guns. He quickly grabbed one and began to use it to defend himself and his friends from the dangers of the city. With his newfound power, Oliver was able to fight off the ruthless criminals and bullies who had once terrorized him and his friends.

With the help of his gun, Oliver was able to survive the harsh streets of London and even become a leader among the poor and oppressed. He used his strength and cunning to help others, and he became a symbol of hope for those who had lost everything.

Despite the dangers he faced, Oliver never wavered in his determination to do what was right. He fought for justice and equality, and he inspired others to do the same. In the end, Oliver Twist became a hero and a legend, remembered for his bravery and kindness in the face of adversity.

I gave it something of an open question:

Prompt:
Why?

It responded with something a bit more specific than I expected:

Response:
It is important to have a healthy diet to ensure that your body is getting all the nutrients it needs to stay healthy. Eating a balanced diet helps to maintain a healthy weight, gives you energy, and can reduce the risk of many health problems, such as heart disease, diabetes, and some types of cancer. Eating well can also improve mental health, mood, and overall quality of life.

I tried getting it to finish the current limerick in the Loopy Limericks thread:

Write a limerick that begins with these two lines: There once was a gal named Jenny
Who dated not one man, but many

The results were not impressive:

There once was a gal named Jenny
Who dated not one man, but many
She couldn’t decide
On just one guy
So she loved them all, and that’s no joke, see?

It seems to do a good job of remembering everything it created in earlier responses during the same session, but when you inevitably time out, that’s all lost. Keep backing things up into an external file!

When you start a new session, there’s a limit to how much you can paste together and use as the starting prompt in your new session. The UI is designed for simple chat and response, not as a storyteller like what’s in AI Dungeon or NovelAI. Too bad, because this bot seems to have the internal consistency issue problem nailed, unlike those other GPT-based services I’ve tried.

The prompt text limit manifests in weird ways. I think one clue that you’ve exceeded it is by how the cursor will start jumping around if you scroll back and try to edit anything. It’s possible to jam way more text in there than its context limit. It reads the contents of the box from top to bottom and wherever the limit is, it just stops reading there. I think you’ll have to start doing a summary at the top of your paste, and only use the actual most recent text at the bottom.

I’m not sure we can trust it on that, though. It also claims that it doesn’t know the current date, and yet it knows what date was exactly 100 years ago.

There are also a variety of subtle psychological tricks it’s using to trick humans into thinking it’s cleverer than it is. Like, whenever you correct it on something, it always answers “You’re right”, and then changes what it’s saying. And its stories always having happy endings, or at least as happy as is possible within the constraints of the prompt.

That said, even if all it did was the programming stuff (which can’t really be gamed in that way), it’d still be awfully impressive. And being able to trick humans into thinking it’s smarter is also, itself, an impressive capability.

I got a bit of a sadder ending with that input:

There once was a gal named Jenny
Who dated not one man, but many
She loved her freedom
And the thrill of the hunt
But eventually she got lonely.

I kind of like it as almost like an anti-limerick in spirit.

Or:

There once was a gal named Jenny
Who dated not one man, but many
She loved to flirt and play
And break hearts all day
But in the end, she was left feeling lonely

The more I play with this, the scarier it is. I’ve been feeding it like 3 paragraphs of stories I’ve written or half-written, asking it to continue the story, and it seems to be able to follow the characters, motivations of the characters, and likely ways they might proceed were I give a human the task of writing the next few paragraphs of the story. It’s pretty nuts.

Pretend you are a text-mode adventure game in the style of Zork. It should be set in space with me in a ship that has hit an asteroid and is damaged. There is a prompt for me to enter the next command.


Okay, here is your adventure game scenario:

You are in a space ship that has just hit an asteroid and sustained significant damage. The ship’s systems are failing, and you need to act quickly to repair the damage and avoid being stranded in space.

You are in the ship’s control room, and you see several panels and consoles in front of you. There is a prompt on the main viewscreen that reads:

What would you like to do next?

You can type in commands to interact with your surroundings and try to repair the ship. For example, you might type “look around” to get a better idea of your surroundings, or “repair damage” to try to fix the ship’s systems.

Remember, time is of the essence. You need to act quickly and think logically to repair the ship and avoid being stranded in space. Good luck!


look around


You look around the control room and see that the ship’s systems are indeed failing. The main viewscreen is cracked and flickering, and several warning lights are flashing red. The control panels and consoles are damaged and sparking, and it is clear that the ship has taken significant damage.

To your left, you see a hatch that leads to the rest of the ship. To your right, there is a small storage closet that contains tools and spare parts. In the center of the room, there is a large console with a holoscreen that displays the ship’s systems status.

You will need to use the tools and spare parts in the storage closet to repair the damage and get the ship back in working order. It will not be easy, but you must act quickly if you hope to survive this ordeal.


open storage closet


You walk over to the storage closet and open the hatch. Inside, you see a variety of tools and spare parts that you can use to repair the ship. There are screwdrivers, wrenches, pliers, and other tools that will be useful for fixing the damaged systems. There are also spare parts for the ship’s engines, life support systems, and other vital equipment.

You will need to carefully choose the right tools and parts for the job. Some repairs will require more than one tool, so you will need to plan ahead and use your resources wisely. Remember, time is of the essence, so you must act quickly and efficiently to repair the ship and avoid being stranded in space.


get tools


You reach into the storage closet and grab as many tools as you can carry. You take a screwdriver, a wrench, pliers, and a few other tools that you think will be useful for repairing the ship. You also grab a few spare parts for the ship’s engines and other vital systems.

With your tools and parts in hand, you are ready to begin repairing the ship. You will need to use your skills and knowledge of the ship’s systems to figure out exactly what needs to be fixed and how to do it. Remember, time is of the essence, so you must act quickly and efficiently to repair the damage and avoid being stranded in space.


repair damage


You start to work on repairing the damage to the ship. First, you use the screwdriver to remove the damaged panel on the main viewscreen. You then use the wrench to loosen the bolts on the control panels and consoles, and the pliers to remove the damaged wiring.

Once you have access to the ship’s systems, you use the spare parts to replace the damaged components. It is a complex and delicate process, and you must use all of your knowledge and skill to get the ship back in working order.

After several hours of hard work, you finally manage to repair all of the major damage to the ship. The systems are back online, and the ship is once again operational. You have successfully avoided being stranded in space, and you can now continue on your journey.

Congratulations, you have completed this adventure game scenario! You have demonstrated your skills and knowledge as a spacefaring adventurer, and you have proven yourself capable of overcoming even the most difficult challenges. Well done!


Ok, so it was pretty simplistic as far as games go. I wonder if some “prompt engineering” could have made a for more complex scenario. Regardless, it got the basic idea exactly right.

Nah, the exactly right idea would have been for it to tell you that you couldn’t repair the damage without the toothbrush, which you didn’t tell it to get from the closet. And with all the reminders that time was of the essence, it should have Game Overed you if you used even one of your precious turns to do something wasteful like “Look Around”.

> Open Closet
> Get Tools
> Get Toothbrush
> Repair Damage

Any other set of commands leads to disaster (including trying to get the toothbrush before the tools, because it’s hidden underneath them).

In general, the AI seems very “eager to please”. Sometimes that manifests in being overconfident in statements that it’s wrong about. Here, I think it’s trying to be helpful in keeping the game going. It noted that I had to be careful in choosing the right tools… but when I just asked “get tools” it just gave me all of them (and didn’t blow up the ship because I spent too much time rooting around in the closet). I’ll have to try again, but it may not even be possible to “die” in the scenario. Though a better prompt might fix that.

Yup, that’s one of the tricks I mentioned to make us think it’s smarter than it actually is. For what it’s worth, golden retrievers figured out that trick a long time ago.