Rick Korzekwa, March 3, 2023
A significant theme in reporting on ChatGPT is the fast progress of its person base. A generally said declare is that it broke data, with over 1 million customers in lower than every week and 100 million customers in lower than two months. It appears to not have damaged the report, although I do suppose ChatGPT’s progress is an outlier.
Checking the claims
From what I can inform, the one supply for the declare that ChatGPT had 1 million customers in lower than every week comes from this tweet by Sam Altman, the CEO of OpenAI:
I don’t see any motive to strongly doubt that is correct, however consider it’s an imprecise assertion from a single individual with an incentive to advertise a product, so it may very well be incorrect or deceptive.
The declare that it reached 100 million customers inside two months has been reported by many information shops, which all appear to backside out in information from Similarweb. I used to be not capable of finding an in depth report, nevertheless it appears like they’ve extra information behind a paywall. I feel it’s cheap to just accept this declare for now, however, once more, it could be totally different in a roundabout way from what the media is reporting.
Setting data and progress of different apps
Claims of report setting
I noticed folks sharing graphs that confirmed the variety of customers over time for varied apps and providers. Here’s a rather hyperbolic example:
That’s a formidable curve and it displays a notable occasion. But it surely’s lacking some vital information and context.
The declare that this set a report appears to originate from a remark by an analyst at funding financial institution UBS, who mentioned “We can’t keep in mind an app scaling at this tempo”, which strikes me as an inexpensive, hedged factor to say. The stronger declare that it set an outright report appears to be misreporting.
Knowledge on different apps
I discovered information on month-to-month customers for all of those apps besides Spotify. I additionally searched lists of very fashionable apps for good leads on one thing with quicker person progress. You’ll be able to see the complete set of information, with sources, right here. I give extra particulars on the information and my strategies within the appendix.
From what I can inform, that graph within reason correct, nevertheless it’s lacking Pokémon GO, which was considerably quicker. It’s additionally lacking the Android launch of Instagram, which is arguably a brand new app launch, and surpassed 1M throughout the first day. Right here’s a desk summarizing the numbers I used to be capable of finding, listed in chronological order:
|Days to 1M
|Days to 10M
|Days to 100M
|Netflix subscribers (all)
|Netflix subscribers (streaming)
|Pokemon Go (downloads)
It’s somewhat exhausting to match early numbers for ChatGPT and Pokémon GO, since I couldn’t discover the times to 1M for Pokémon GO or the times to 10M for ChatGPT, nevertheless it appears unlikely that ChatGPT was quicker for both.
Scaling by inhabitants of Web customers
The whole variety of folks with entry to the Web has been rising quickly over the previous couple of a long time. Moreover, the expansion of social networking websites makes it simpler for folks to share apps with one another. Each of those ought to make it simpler for an app to unfold. With that in thoughts, right here’s a graph displaying the fraction of all Web customers who’re utilizing every app over time (word the logarithmic vertical axis):
Basically, it appears like these curves have preliminary slopes which are growing with time, suggesting that how shortly an app can unfold is influenced by extra than simply a rise within the variety of folks with entry to the Web. However Pokémon GO and ChatGPT simply appear like vertical strains of various heights, so right here’s one other graph, displaying the (logarithmic) time since launch for every app:
This exhibits fairly clearly that, whereas ChatGPT is an outlier, it was nonetheless considerably slower than Pokémon GO.
Another comparability we are able to make is to different services which have a really quick uptake with customers and the way their attain will increase over time:
- YouTube views inside 24 hours for newly posted movies offers us a reference level for a way shortly a hyperlink to one thing on the Web can unfold and get engagement. The decrease barrier to watching a video, in comparison with making an account for ChatGPT, may give movies a bonus. Moreover, there may be presumably multiple view per individual. I have no idea how large this impact is, however it might be giant.
- Pay-per-view gross sales for dwell occasions, on this case for fight sports activities, are a reference level for one thing that individuals are prepared to pay for to make use of at residence in a brief timeframe. The cost is the next barrier than making an account, however advertising and marketing and gross sales can occur forward of time.
- Online game gross sales inside 24 hours, in some circumstances digital downloads, are just like pay-per-view, however appear extra instantly similar to a service on a web site. I might guess that video video games profit from an extended interval of selling and pre-sales than PPV, however I’m unsure.
Here’s a graph of data for this stuff over time, with information taken from Wikipedia, which is included within the information spreadsheet. Every dot is a separate video, PPV occasion, or sport, and I’m solely together with those who set 24 hour data:
It might seem that very fashionable apps aren’t as well-liked as very fashionable video video games or movies. I don’t see a powerful conclusion to be drawn from this, however I do suppose it’s useful context.
I think the advertising and marketing benefit for Pokémon GO and different videogames is substantial. I don’t keep in mind seeing adverts for Pokémon GO earlier than its launch, however I did a quick seek for information articles about it earlier than it was launched and located a number of hype going again months. I didn’t discover any information articles mentioning ChatGPT earlier than launch. This doesn’t change the general conclusion, that the declare about ChatGPT setting an outright report is fake, nevertheless it ought to change how we give it some thought.
That ChatGPT was in a position to beat out most different providers with none advertising and marketing looks like a giant deal. I feel it’s exhausting to promote folks on what’s cool about it with out a number of person engagement, however the subsequent technology of AI merchandise won’t want that, now that individuals are conscious of how far the know-how has come. Given this (and the hype round Bing Chat and Bard), I might weakly predict that advertising and marketing will play a bigger function in future releases.
Appendix – strategies and caveats
Many of the numbers I discovered had been for month-to-month customers or, in some circumstances, month-to-month lively customers. I wasn’t at all times certain what the distinction was between these two issues. In some circumstances, all I used to be capable of finding was month-to-month app downloads or annual downloads, each of which I might naively count on to be strictly bigger than month-to-month customers. However the annual person numbers mirrored longer-term progress anyway, so that they shouldn’t have an effect on the conclusions.
Among the numbers for days to explicit person milestones had been interpolated, assuming exponential progress. By and huge, I don’t suppose this impacts the general story an excessive amount of, but when you have to know exact numbers, you need to test my interpolations or discover extra direct measurements. Not one of the numbers is extrapolated.
When trying to find information, I attempted to make use of both official sources like SEC filings and firm bulletins, or measurements from third-party providers that appear respected and have paying clients. However typically these had been exhausting to search out and I had to make use of much less dependable sources like information studies with doubtful citations or research with incomplete information.
I didn’t strategy this with the intent to provide very dependable information in a really cautious approach. Total, this took about 1-2 researcher-days of effort. Given this, it appears probably I made some errors, however hopefully not any that undermine the conclusions.
Due to Jeffrey Heninger and Harlan Stewart for his or her assist with analysis on this. Due to the 2 of them and Daniel Filan for useful feedback.