Loosely, ‘negligible’ means small enough to ignore. A difference in probability of $2^{-128}$ is likely small enough to ignore; a difference in probability $2^{-10}$ may not be.
Formally, in the language of complexity theory, a function $f\colon \mathbb N \to \mathbb R$ which assigns to each natural number $n$ a real number $f(n)$ is negligible if it goes to zero faster than the inverse of any polynomial—that is, if for any polynomial $p$, there exists a number $n_0$ such that for any $n > n_0$, $|f(n)| < |1/p(n)|$.
For example, $n \mapsto e^{-n}$ is a negligible function.
Sketch of proof: Let $p(x) = x^d$; at what point $n_0$ do we have $e^{-n} < 1/n^d$, or $e^d < n^n = e^{n \log n}$, for all $n > n_0$? We don't need to solve this exactly—it suffices to set $n_0 = d$; as long as $n > n_0 = d$, $n \log n > n > d$, so $n^n = e^{n \log n} > e^d$ and thus $n^{-d} > e^{-n}$, or $e^{-n} < 1/n^d$ as desired. Finding the corresponding bound $n_0$ for other polynomials left as an exercise for the reader.
In the language of complexity theory, we usually consider ensembles of cryptosystems that scale with a security parameter, and we consider the asymptotics of the growth curves of the user's cost and the adversary's cost as the security parameter scales.
The same is true for algorithms more generally, which is why software engineers will casually talk of quicksort's $O(n^2)$ worst-case time but $O(n \log n)$ average time, because it's a useful proxy for estimating concrete costs—and, unlike a measurement in seconds or joules, it doesn't change if you move from one machine to another, which is helpful for understanding how it will scale if you feed it more and more data even if you use a slightly faster machine.
Of course, we must be careful not to lose sight of the concrete numbers without asymptotics: while it is proven[1] that multiplication can be computed in $O(n \log n)$ bit operations, you'll still get an answer faster in wall clock time using a naive $O(n^2)$ algorithm for any input that is possible to work with in practice.
Similarly, while the Blum–Blum–Shub pseudorandom generator has a polynomial-time reduction theorem relating the difficulty of distinguishing BBS output to the difficulty of factoring the BBS modulus, the concrete security of BBS[2] (paywall-free) is such that it costs $2^{43}$ bit operations to generate a mere 128 KB of data at a ~100-bit security level, using somewhat old estimates for factoring costs. At 2.5 GHz, that takes a minute of computation. In contrast, e.g., AES-256 can generate the same amount of data in microseconds at a much higher security level.
