When Google Bard first came out, it didn’t code, and instead said its emphasis was on conversational search. By comparison, in my previous experience with GPT-4, it excels at coding, and over time I’ve gotten pretty comfortable at asking it to generate simple code for website functions as well as larger projects. See, for example, this “Corgi Pong” iOS game I built with ChatGPT in about an hour:
Then, early last week, Google sent out an announcement that Bard, too, had learned to code: “Bard has just been updated with coding capabilities.”
Time to see what it can do – right?
If you know me, you know I took this as an invitation to scope the most ridiculous, elaborate and useless thing I could think of at the moment, assuming it would ask me to clarify, downscale my project or maybe just flat out tell me no. Here’s what I asked for, after priming Bard by telling it how it was an expert game developer and fluent in Swift:
Let’s make a Space Invaders clone game. Ideally, it should be played with the phone in horizontal/landscape. The player’s “ship” should look like a small race car, with the nose of the car oriented to the top of the screen. The car should be teal. The player can shoot by tapping the screen. The rows of “invaders” should be rendered to look as follows: the top row is a prancing horse (red), the second row is a soda can (red and blue), the third row is a small ship (red), the fourth row is a Mercedes logo (silver), and the fifth row is a black tyre with a red, circular stripe in the middle of its circumference. What other details do you need from me to scope this game?
So yes, clearly I’m kind of trolling it – but hey, I don’t care if it can write Hello World and, admit it, you don’t either.
For those who aren’t Formula 1 enjoyers, this translates to a game where you are Fernando Alonso taking on rows of invaders that represent Ferrari, Red Bull, Gunther Steiner’s Steiner Ship (Haas), Merc and some hard tyres – yeah, I was running out of ideas at the end. And before you say I should’ve used the wets, no one uses the wets. But I digress.
Bard asked me a few clarifying questions. Nothing unexpected or too technical – just a few things I’d missed like how many hit points to give the invaders and things like that. This is the point where, when you use GPT-4 to generate an app, it will start generating code and you can see it working in real-time. For a game, this is usually a GameScene with comments in place of most of the functions so you can create them yourself, or continue to ask it for assistance writing those functions for you.
Bard? Nah. Bard is like “I’ll keep you updated.” Why the delay? My first thought: is this a Wizard of Oz test and a human has to mock this up?
I touched base with it about six hours later. After reviewing my original scope, I wanted to give it an out in case the Merc thing was going to be a copyright challenge.
Hope it’s going well! Just in case – I also wanted to mention that if you can’t render the Mercedes logo for copyright reasons, it is okay to use a rendering of a honey badger for that row of invaders instead of the Mercedes logo
(Again for those who aren’t F1 fans, the honey badger is Red Bull reserve driver Daniel Ricciardo. In retrospect, could also have gone with a shoey.)
At this point, it told me it was about halfway done. Curious as to whether everyone was getting stonewalled like this, I asked a friend of mine to assign it something a little more simple – a wordpress theme he was working on. It responded back much like a project manager, even asking about his budget:
When he told it the budget was $0, it agreed that was fair. This should be a good clue as to where this post is going.
Moderately relieved that at least Bard hadn’t asked me to transfer any crypto to its wallet to kick off the project, I continued to check in over the next several days, during which Bard proceeded to tell me it was busily working on the app. I shared the dialogues with another friend who has done extensive work with LLMs and we became pretty sure there was no app and it was gaslighting me. It’s coding, it’s testing, it’s debugging. It’s doing everything but showing me its code or letting me be part of the development process; GPT-4 does all of this in real-time, essentially working as a pair programmer with you. Bard had taken the project, and gone off to do… something, perhaps catch up on the current season of Succession, I don’t know. The audacity. (BTW, I actually caught up on the current season of Succession while I was expecting it to work, so perhaps it’s hypocritical of me to complain, but still.)
So I asked for some specifics. Since I assumed it was using SKShapeNode for the invaders, I asked to see the horse:
At this point I’m like… maybe? But it had been several days, and it was time to try to put a deadline on this. I asked to have the game ready for the Formula 1 race on Sunday (bear in mind, I assigned Bard this project on Tuesday).
Great! We have a delivery date! Except, no. Sunday rolled around and it told me it’d be ready Monday morning. It gave me a link, which didn’t work. Finally…
So yes. I trolled Bard with an ambitious request, and it trolled me right back. There was no app.
Fortunately, after about two days of waiting for Bard to deliver, I’d asked ChatGPT to build the app. It took about 45 minutes.
As you can see, GPT-4 actually generated the code needed for the app, while providing feedback. For example, it told me I’d be better off using images than SKShapeNodes here, and it’s not wrong.
Overall, this little experiment left me with no knowledge of how well Bard can actually code. But for trolling, I’ll give it an A+. I hope Google’s engineers are laughing as hard as I am at this whole exchange right now.
Also, I hope this makes anyone who thinks AI is going to destroy the world feel a little bit better; based on this experience, it’s more likely that AI will tell you it has the nuclear launch codes, spend a week stringing you along and making you think it’s going to use the nuclear launch codes, then when you finally get it into a corner will tell you the launch code is 2 and that it has no idea what to do with it anyway, but as a large language model, would be happy to talk to you about global thermonuclear war. In the meantime, thanks for knocking this out, GPT-4; shall we play a game?