It's true the rate of useful model growth is probably lower than the raw model count suggests, given what we've seen; the prior probability of any new model being a significant leap is, say, 10%. But we’ve also seen that even with the noise, the cumulative effect over two years still means a posterior probability of 75% that a user finds a better fit for a specific task today than they could last year, simply because the base rate of options has gone up so much.