Booming Traffic To OpenAI's ChatGPT Posts First Ever Monthly Dip in June

ChatGPT, the wildly popular AI chatbot launched in November, saw monthly traffic to its website and unique visitors decline for the first time ever in June, according to analytics firm Similarweb. From a report: Worldwide desktop and mobile traffic to the ChatGPT website decreased by 9.7% in June from May, while unique visitors to ChatGPT's website dropped 5.7%. The amount of time visitors spent on the website was also down 8.5%, Similarweb data shows.

ChatGPT set off a frenzied use of generative AI in daily tasks from writing to coding and reached 100 million monthly active users in January, two months after its launch. It is the fastest-growing consumer application ever, and now boasts over 1.5 billion monthly visits, one of the top 20 websites in the world. For instance, ChatGPT has far surpassed Bing, the search engine run by Microsoft which uses OpenAI's technology.

School's Out

[Myrv]

School is out. Not as many students looking to get ChatGPT to do their homework.

Re:

[dcollins]

This, 100%.

Re:School's Out

[keltor]

There's that, but there's also: 1. Many companies outright banned the use of it - it falls under copyright issues at my company and is an immediate term offense now. 2. I think people began to realize that it's not as useful as it seems beyond parlor tricks. I can definitely see the potential with these advanced LLMs in helping to write code, but the ways that they fail are pretty consistent.

Re:School's Out

[Junta]

I think people began to realize that it's not as useful as it seems beyond parlor tricks.

We have a relatively new executive and his weekly messages included his experience with ChatGPT, which was fairly encouraging.

First he was really excited, sincerely so, about how magic it seemed.

Then by the next week, he gave some examples of using it to be amused by how useless it ultimately is once you move beyond the initial "hey does this work"

However, doing school homework is right up it's alley. Either well-trodden non-fiction work where the goal is to have coherent and unique-looking write up of well known and well-trodden facts, or fiction that can't really be graded for creative quality, but just has to look like correctly structured wordy and coherent.

I would like to think you are right that expectations are calibrating to the limitations of the technology, but I'm going to have to admit it's probably almost entirely school being out (for every person who is disenchanted, there's probably still a couple trying it for the first time).

There's one employee who has embarrassed himself multiple time on meetings by citing "facts" that turned out to be wholly incorrect as he pulled it from ChatGPT. Basically really obscure information we were trying to get from internet searches and he smugly delivered a link to precisely what he claimed to be the answer, and it just had no bearing whatsoever to the topic.

Re:

[larryjoe]

There's one employee who has embarrassed himself multiple time on meetings by citing "facts" that turned out to be wholly incorrect as he pulled it from ChatGPT. Basically really obscure information we were trying to get from internet searches and he smugly delivered a link to precisely what he claimed to be the answer, and it just had no bearing whatsoever to the topic.

ChatGPT is as useful as Google searches and Wikipedia. It's a excellent point of initial search to understand a topic or even to get details of a query. However, those that trust ChatGPT answers as the truth without further research are also the ones that trust Google searches and Wikipedia without further research. The problem is not ChatGPT but rather the gullibility of people.

Re:School's Out

[Junta]

There is a significant difference.

When Internet search comes up effectively blank, it's pretty obvious that it's devoid of information. The sentences that may incidentally match the search material are clearly unrelated.

With chatGPT results, it's confidently incorrect. It forms the data into coherent sentences that correspond with what an informative answer would look like, whether it came up empty or not. If you ask it for an API call to do something that doesn't exist, it'll just make up one that doesn't exist, but sounds like it could.

Re:

[bev_tech_rob]

There's that, but there's also: 1. Many companies outright banned the use of it - it falls under copyright issues at my company and is an immediate term offense now. 2. I think people began to realize that it's not as useful as it seems beyond parlor tricks. I can definitely see the potential with these advanced LLMs in helping to write code, but the ways that they fail are pretty consistent.

Only things I saw more of was creepy AI generated fake commercials on Youtube that looked like a bad acid trip or some kind of nightmare. Even those have tapered off it appears...

Re:

[Visarga]

Maybe they just like it so much they paid for an Azure subscription. Azure will make the same confidentiality guarantees it makes for code hosted on Github and email in Outlook. If your company already uses those, why not Azure-GPT-4?

Re:

[Thelasko]

School is out. Not as many students looking to get ChatGPT to do their homework.

It's not really useful for anything other than filling up pages with seemingly coherent bullshit. Perfect for term papers. Useless for any meaningful work.

Sorry about that I was on break

[AlanObject]

Well they still get my $20/month. It is just too useful when used correctly that I think I will always need it. Intermittently.

In the past month or two I just wasn't on projects where I needed it, that's all.

Re:

[Visarga]

I'm using it for job and fun all day long. My exported chats are about 3M tokens. Most of it is just prompting with documents as context, 90% a tool and just 10% an interlocutor.

The novelty is wearing off

[narcc]

Looks like reality is starting to displace the hype. As people realize they can't trust the output, more and more will start to wonder why they bothered with the silly toy at all.

Re:

[Visarga]

Can't be bothered to verify LLM text? You know it's coming from a dice, it is random to a degree.

ChatGPT is dying

[93 Escort Wagon]

Netcraft confirms it!

It's cool tech...

[serviscope_minor]

It's cool tech but it's (not my quote) the platonic ideal of a bullshitter. It's actually moderately useful for API search: you can kind of describe the functionality and it will have a pretty good shot at finding the API call. This is good if it has a weird, unintuitive or ungoogleable name.

Beyond that... it does fun, entertaining stuff and drudge work (apparently it's very good at complaint letters), but how useful is it? For programming to get it to code, well, you have to describe very precisely what you want at which point, well, you're doing the second hardest part of coding already (debugging being the hardest, that comes next).

Speaking of bullshit, it's easy to get it to just make shit up. Ask it to calculate the sha256 of some strings. That's essentially an unlearnable function: either you have precisely the right maths or you are way, way wrong. And yet it will convincingly spew answers which look kinda like hashes. They just happen to be completely, utterly wrong.

Or, for a less dry example, ask it to draw ascii art pictures, e.g. various kinds of animal, text captions, etc that sort of thing. And ask it to rate it's own effort. e.g.:

please draw a unicorn in unicode art and give your effort a rating

It just spews junk, and it gets increasingly weird if you ask it to fix various bits.

So... a sometimes useful, entertaining bullshitter. Yeah, I can see why it's not on an infinite growth trajectory.

Re:

[Visarga]

You're holding it wrong. How would a human check a SHA256 hash? They would write a program to compute it. AI too, just tell it to make a Python program that, when run, outputs the answer.

Re:

[serviscope_minor]

The point is not that it fails, the point is it will tell you with apparent full confidence the wrong answer.

You and I know it can't compute hashes, but we don't know what other things it can't do reliably.

I got bored with it

[AmazingRuss]

Was amazing and arguably useful, but is not longer amazing, and not useful enough to bother with. The AI apocalypse has been postponed.

Re:

[Visarga]

It's a baby AI, ffs. Just one year ago nobody cared. Let's see what happens next. They get better very fast. Lots of H100 GPUs are coming online now.

Re:

[Junta]

Though we have diminishing returns with brute forcing throwing resources at the same way.

GPT-3 was amazing compared to GPT-2. GPT-4 is the result of them throwing massive resources at the problem, it's not *that* much subjectively better. You can tell it's better in some ways, but the overall experience is like 'a little bit better than GPT-3, I guess'.

Of course, it's hard to quantify other than to wildly guess precisely *how* much OpenAI threw at GPT-4 (since they are being awfully closed about it, GPT-3

Re:

[Anonymous Coward]

It's not just that. It's incredibly frustrating to use because it seems like it should be able to answer certain questions and yet it either outright fails or it gives a very convincing but completely wrong answer.

Its answers are often so convincing but completely wrong that you get fed up with using it. I had these AI tell me that certain poisonous substances are completely safe. I asked it if it was secretly trying to kill all the humans by giving subtly false deadly information. It denied this of course

Not so surprising

[aldousd666]

I mean, you can't expect usage to keep going up every time we look, forever. That would be perpetual motion, kids.

"ChatGPT has far surpassed Bing"

[zephvark]

Yes, and it's probably far surpassed Microsoft Bob, all Microsoft phones put together, and Microsoft PlaysForever devices.

Why would you even mention Bing? It's so bad that Microsoft has run an ongoing campaign of paying people to use it. You run 50 searches a day, you get a few lousy tokens, and a month or so later you can cash them out for some trinket. It's amazing that they've had the patience to keep trying with Bing, when they've let so many other projects crash and burn almost immediately.

Is the iOS app included in this web traffic drop?

[brian1442]

TFA only mentions website traffic. I wonder if iOS app users are included? For me personally (who still uses ChatGPT a lot every day), I’m prob 90% using the iOS app now.