What an amazing set of data!
The "Generative AI services popularity" [1] chart is surprising. ChatGPT is being #1 makes sense, but Character.AI being #2 is surprising, being ahead of Anthropic, Perplexity, and xAI. I suspect this data is strongly affected by the services DNS caching strategies.
The other interesting chart is "Workers AI model popularity" [2]. `llama-3-8b-instruct` has been leading at 30% to 40% since April. That makes it hands the most popular weights available small "large language model". I would have expected Meta's `m2m100-1.2b` to be more used, as well as Alphabet's `Gemma 3 270M` starting to appear. People are likely using the most powerful model that fits on a CF worker.
As shameless plug, for more popularity analysis, check out my "LLM Assistant Census" [3].
[1] https://radar.cloudflare.com/ai-insights#generative-ai-servi...
[2] https://radar.cloudflare.com/ai-insights?dateRange=24w#worke...
Character.AI is extremely popular among youngers so it's not really surprising.
Why would DNS caching skew results?
I don’t think Cloudflare is using DNS queries to compile the stats considering they have visibility into the full http requests for sites they proxy.
Edit: Another comment mentions DNS queries. Did I miss something about how they’re compiling the stats?