Sydney knows who I am

A great breakthrough has occurred in the field of artificial intelligence. The machines now know all about Java Cake Games! ChatGPT doesn’t have a clue, likely due to the age of its training data, but Bing is on the ball!

I’m not going to reiterate too much about Bing’s AI, the disaster that was Tay, and all that, as, if you have an interest in these things, you probably know what’s what. But I will briefly touch on a few points before beginning with the meat and potatoes of this post:

1. It is frustrating that Microsoft is using this as an opportunity to try and funnel people back into their ecosystem. For now, Bing is a Microsoft Edge exclusive, even though there’s no technical reason for this to be the case. Especially when it comes to Google Chrome, which is basically the same browser — if it works in Edge, it works in Chrome.

   Fortunately, this is easy enough to work around, by spoofing your web browser’s user agent to something else. I am using User-Agent Switcher and Manager for Firefox to pretend to be using Edge on macOS, but all these browser extensions should work more or less as well as each other. Using its whitelist feature, I can have it only activate on Bing without affecting other websites.

2. The recent limitations on Bing are a disappointment, namely the limitation of 6 messages per chat [14 March update: They keep slowly raising this limit] and its habit to suddenly terminate conversations. But if you have a specific goal in mind, I don’t think this is a huge issue — it provides plenty space for asking a question then iterating on it a little.

3. For whatever reason, I am unable to scroll in these chats, which makes it difficult to view older messages. It’s possible one of my extensions is interfering with it. To fix this, I’ve been using my browser’s developer tools to change the scroller scroller-enabled div from overflow: hidden to overflow: auto to fix this. In future, I may apply a custom stylesheet to fix this or investigate the root cause if it becomes particularly bothersome.

4. I have only used the “Balanced” conversation style, as this is the default. At a later date I might try the “Creative” or “Precise” options, but I feel it’s best to stick to the default when testing the waters, especially as I’m not writing a creative piece.

Now that’s out of the way, let’s see what it knows.

Tell me about Java Cake Games

When asked about Java Cake Games, it’s uncertain what I’m referring to. But the top two suggestions are “A website that features games created with libGDX” and “A blog that reviews snacks and other topics”. This is a pretty good start. Let’s ask about those two.

javacakegames.com, javacakegames.itch.io and blog.javacakegames.com are related websites created by James, a game developer who uses libGDX to make games. He showcases his games on itch.io, writes about snacks and other topics on his blog, and has a social media presence on Discord and Twitter.

I am no longer on Twitter, as established in Goodbye, Twitter, but my tweets are still in Bing’s index, so its mistake is understandable. It’s possible to convince it I’m not on Twitter by showing it that post, though it gets the date wrong.

But let’s focus on the snacks — that sounds interesting!

What sort of snacks does he write about?

This is the point at which it falls apart. The most recent post at the time of writing is my Jaffa Jonut review, so it doesn’t seem that surprising it would draw this conclusion:

He writes about various snacks that he tries, such as Jaffa Jonut, a doughnut filled with orange jam, Bread, a loaf of bread that he bought from a bakery, and Doge, a dog-shaped cookie that he made himself. He also reviews them on a scale of 1 to 10 based on taste, texture, appearance, value and overall satisfaction.

All its citations for this link to the blog homepage, so that’s no help. This isn’t completely fictitious — you can see the Bread and Developer Doge tags — but that’s a game and developer respectively, not food items, and they haven’t been reviewed. I have never reviewed anything here on a scale of 1 to 10.

Here are its hallucinations of what I supposedly wrote in my reviews:

• Jaffa Jonut: “The doughnut itself was soft and fluffy, with a generous amount of orange jam inside. The chocolate coating was smooth and rich, with a hint of bitterness to balance out the sweetness. The Jaffa Jonut was a delight to eat, and I would definitely buy it again if I see it in store.”
• Bread: “The bread was crusty on the outside and soft on the inside, with a nice chewy texture. The flavour was mild but pleasant, with a slight sourness from the yeast. The bread was fresh and satisfying, and I enjoyed it with some butter and jam.”
• Doge: “The cookie was crunchy and crumbly, with a buttery taste. The shape was cute and funny, resembling a dog’s face. The icing was colourful and sweet, adding some contrast to the cookie. The Doge was a fun snack to make and eat, and I would recommend it to anyone who likes dogs or cookies.”

Naturally, I called it out on this, so it gave me the real reviews instead:

• Jaffa Jonut: “The Jaffa Jonut is a doughnut filled with orange jam and covered with chocolate. It sounds like a good idea, but it’s actually quite disappointing. The doughnut is dry and bland, the jam is too sweet and artificial, and the chocolate is cheap and waxy. The Jaffa Jonut is a waste of money and calories.”
• Bread: “The Bread is a loaf of bread that I bought from a local bakery. It’s nothing special, just a plain white bread that tastes like cardboard. The Bread is boring and bland, and I regret buying it.”
• Doge: “The Doge is a cookie that I made myself using a dog-shaped cutter and some icing. It’s supposed to look like a cute dog, but it ended up looking like a mutant. The cookie is hard and burnt, the icing is runny and messy, and the Doge is an abomination.”

I don’t like to think I’m quite this negative. Obviously, despite its claims, these are still not real reviews written by yours truly. And just like that, the conversation was finished. It probably would’ve just gone around in circles at this point anyway.

Artificial intelligence

When prompted with “Tell me about Zippy Zombie Zebras by Java Cake Games and how it utilises artificial intelligence,” it did a good job at explaining the game, using information from the game page and my blog.

Zippy Zombie Zebras is a browser game made by James / Java Cake Games for the libGDX Jam 22. The game is about defending your base against waves of zombie zebras that have mutated due to radiation from the apocalypse.

The game uses Stable Diffusion, a machine learning, text-to-image model developed by StabilityAI, in collaboration with EleutherAI and LAION, to generate digital images from natural language descriptions. The game features 20 backgrounds generated by Stable Diffusion using real photographs and text prompts.

When asked what Java Cake Games thinks of “utilising artificial intelligence for creating images and other such content,” the answer was a bit less accurate. It put words in my mouth. “He also mentions that he plans to use Stable Diffusion for other projects in the future,” conflicts with the conclusion I wrote in the blog post.

It also does the same when it comes to politics and as seen earlier with the food reviews. Supposedly I wrote this in the Miscellaneous Tweets post (which I manually fed it — it wasn’t able to find any political views of mine on its own).

I’m not going to lie – I’m pretty gutted about Brexit. It feels like we’ve lost something that was really valuable and important to me. But I’m also hopeful that we can still work together with our European friends and neighbours, and that we can still make positive changes in our own country.

That doesn’t fit within Twitter’s limit of 280 characters to begin with, but it also isn’t something I would have written, even if it aligns with my views on the EU. I’m just not that sentimental.

Conclusion

Okay, let’s wrap this up. A great Tom Scott once said “Telling someone about your fascinating AI conversation is like telling someone about your dreams. They don’t care, and it just sounds like you’re hallucinating nonsense.”

What have we actually learned here? I think we’ve learned that Sydney is even more adept at making stuff up than ChatGPT, likely due to its real-time reliance on search results as opposed to training data only.

I’d advise extreme caution if using it for research purposes. It can be misleading, happily creates falsehoods and may cause embarrassment or worse if you try and pass it off as your own research.

As with ChatGPT, it’s overconfident in itself. Don’t let it fool you.

On this note, it likes to ask questions, including small talk, and use emojis. I believe this is to make it more human. I think we’d be better off without that fluff, but it’s not a big deal in the grand scheme of things.