I've heard that Roger Heath-Brown has presented the following "conjecture" at several conferences, most likely to illustrate our poor understanding of the topic more than because he actually believes it to be true. Let $\pi_2(x)$ denote the number of twin primes less than $x$. Then >> $\pi_2(x) \sim c \frac{x}{\log^2 x} + O(1)$ where $c$ is the twin prime constant. In other words, the twin prime asymptotic holds with error term is $O(1)$.