Not A Saggy, Old Cloth Cat Then?

I wouldn't normally post today but since it's Blaugust and since Wilhelm at TAGN gave me the idea for a post where I barely have to do anything at all...

I guess it's probably still Who The Heck Are You? week in Blaugust. I haven't checked but we haven't been going seven days yet, have we? It's all a bit of a blur. Anyway, Wilhelm reminded me of that thing a few people around this corner of the blogosphere were doing a year or two back, when AI was still quaint and amusing, which was to play Ask The AIs. 

You'd ask them some really obvious questions and then have a good laugh at the answers. It was a fun game for a while but sadly it couldn't last.

There's a blog in my blog roll called AI Weirdness that used to do nothing but demonstrate the bizarre quirkiness of LLMs and Image Generators. It's been very, very quiet for a long time but by chance Janelle Shane, who's blog it is, posted something yesterday.

As evidenced by Janelle's long silence, broken only by sporadic posts months apart, the AIs stopped behaving like funny little toddlers and started giving boringly straight replies. They still hallucinate, of course, but rarely in an amusing way. 

These days, I only ask the LLMs something if I think they're going to give me a useful reply, by which I mean if I want them to do something I already know they're good at. I certainly never ask them for anything to which I don't already know most of the answer. It's more like getting them to go fetch the files than asking them to do primary research.

They're very good at generating prompts for other AI apps, for example. If you want to get a good result out of an image or music generator it often pays to ask an LLM to frame the prompt for you. They speak the same language, I guess.

One thing that used to be amusing was to ask an LLM to tell you who you were. Or who another blogger was. That's what Wilhelm did and although he wasn't very impressed with the results, they looked to me not dissimilar from something a harried sub-editor might knock up to introduce an article about someone they'd just had to spend five minutes researching in the files.

So I thought I'd ask the three of them - ChatGPT, Gemini and Co-Pilot - who they thought I was. Somewhere in the back pages of this blog is a post where I did the same thing, once before. It might be interesting to compare the results to see if the AIs have gotten any better but that would involve more effort than I plan on putting in on this Saturday evening. Feel free to go look for it yourself if you feel so inclined. There's a search function. Or you could ask an AI to do it for you.

Here are the three responses I got, along with my comments. If anyone would like to have a go, asking about themselves, it'd be fun to get a post-chain going about it. Tried suggesting that last time, too, as I remember. Don't think it got much take-up.

Gemini is the AI I use the most often so let's start there. I've inverted the colors to make it more readable but you may need to expand it to see it properly. I probably should have cut&pasted instead of taking a screenshot.

As you can see, I pre-empted the inevitable by telling all the AIs I was not, in fact, a pink, cloth cat. I had enough of that last time.

I found Gemini's response quite odd. It seems to want to tell me who other people think I am rather than making any definitive judgments of its own. I'm curious as to who these bloggers might have been, who said such complimentary things about me - in quotes, no less. It's certainly true I have been called a "contrarian" in my time but not, as I recall, since I've been blogging. I wonder who said that. Or did Gemini just make it up?

Gemini got the name of the blog right and also mentioned Mrs Bhagpuss, which is an AI first. The final bit about the Fantasy Critic League seems extremely random but it is at least true. Why pick on that out of the thousands of things it could have chosen though? Maybe it wanted one very recent fact?

All in all, not at all bad. Doesn't mention anything I've ever written about except games, but then none of the others do either. I have almost five hundred posts tagged "Music" but no-one cares about those, so why should AIs be any different?

Granddaddy ChatGPT next. I have always found ChatGPT to be less reliable than Gemini but others have reported exactly the opposite. This result really supports my view that this is the most over-rated of the three. Gemini didn't say much but what it did say was right. ChatGPT says more without really adding anything substantive and gets some of the very basic facts completely wrong.

I am very much not well-known in the Elder Scrolls Online community. I think "utterly and completely unknown" would be nearer the mark. I have barely played the game, have fewer than twenty posts tagged either with the full name or ESO, and I haven't mentioned it for many years. My blog is also not called "Bhagpuss's Gaming Blog", which would be a terrible name. 

The rest of it is vague puffery which I'm happy to accept and preen over but which is functionally meaningless. It could fit dozens of gaming bloggers as well or better. 

All in all, quite poor. Worst of the three, as we are about to see.

Co-pilot! You're up! And as you can see, it's a hell of a lot easier to read than the others so points for that. Also for the intro, in which Co-Pilot agrees with me that I am indeed not a fictional cat. I was worried there for a moment.

I can't argue with the dismissive tone as Co-Pilot makes it clear I'm not "a widely recognized public figure" either, although I wasn't claiming to be. And once again I'm quite re-assured to be told I don't appear to be a fictional character. Nice opening.

Even more points for the color, the sub-headings and the bullet-point list. Co-Pilot is crushing it on the style front. 

Better still, every one of those bullet points is correct. It knows I'm using a pseudonym, it gets the name of  the blog right, it names the two games that I've written about most frequently and LotRO isn't a bad third choice. It also specifies MMORPGs rather than just games or gaming, although I don't really write about MMOs all that often these days. I do wonder how long I'd have to go on not really writing about MMOs for that to change. Decades, probably.

Again, I'm hardly the right person to comment on the accuracy of the description of my prose style or my place in whatever "community" Co-Pilot thinks it's spotted, but I wouldn't strenuously argue with any of this. Not even over "nostalgic". 

All in all pretty darn impressive. All the facts correct, most of them relevant and everything laid out beautifully.

I declare Co-Pilot the winner. Shame I don't really use any Microsoft products other than Windows itself, especially not for search. Maybe I should make the effort to use Co-Pilot in future instead of just lazily going back to Gemini every time I have a job fit for an AI.

And since this is supposed to be a quick filler post, I'll stop there. I look forward to seeing anyone else's AI-generated bios but I suspect I might be waiting a while. 

 

Notes on AI Used In This Post

Apart from the LLMs, obviously. Just the top image, then. Produced by Google Imagen 4.0 Fast at NightCafe from the prompt "An AI working hard, researching to write a precis biographical introduction to an article about someone it's never heard of - line art, color". Default settings, short runtime.

If you ask any image generator for a picture of an AI it will always, at least in my experience, give you a robot. I didn't want a robot but I also didn't want to waste any more credits trying to get anything else. Imagen, even the non-pro "fast" version, is relatively expensive (Two credits per image. Some of the models are .5 per image and the results seem just as good to me.)

It's a perfectly serviceable illustration for a post about AI and its limitations but I wouldn't be using it for anything else.