Detectability of interstellar messages
Some numbers come from a review paper by Cullers (2000), who discusses the SETI Phoenix project. There, it is claimed that the Arecibo dish is capable of detecting a narrow band, coherent signal of $f=10^{-27}$ W/m$^2$ given a 1000 second observation. Assuming that this is an isotropic signal, then the implied power at distance $d$ is $p=4\pi d^2 f$, which means that $p \simeq d^2$ MW.
So, it is clear that unless a 1MW signal is highly beamed it could not be detected by our current technology even from a nearby star. (Actually, this number is out of date, the receiver at Arecibo is somewhat more sensitive now, but I can't find any numbers). Of course we do emit more beamed signals. The Arecibo radar transmits at 1MW, but its equivalent isotropically radiated power is 20 TW. In other words, the Arecibo dish could detect the directed signals it emits (and of course does, when performing solar system metrology) at distances of about 5000 light years although the radar does not normally send a signal for 1000s.
The SETI Phoenix project, was the most advanced search for radio signals from other intelligent life. Quoting from Cullers et al. (2000): "Typical signals, as opposed to out strongest signals fall below the detection threshold of most surveys, even if the signal were to originate from the nearest star". Quoting from Tarter (2001): "At current levels of sensitivity, targeted microwave searches could detect the equivalent power of strong TV transmitters at a distance of 1 light year.". A recent survey using the Green Bank telescope was able to rule out continuous signals (between 1.1 and 1.9 GHz) at the level of 8 (beamed) Arecibo radars from a large sample of 104 Kepler planet hosts at distances of ~1000 light years.
The next generation of radio telescopes use "phased arrays" to monitor signals from many directions at once and can perform wide-angle surveys much more rapidly. The SETI project is now using the Allen Telescope Array. The claim is that over 10 years it can survey a million stars with sufficient sensitivity to detect the Arecibo radar out to distances of 1000 light years.
It has been suggested that new radio telescope projects and technology like LOFAR and the Square Kilometre Array may be capable (using a month or so of observing time) of serendipitously detecting radio "chatter" at a few hundred MHz out to distances of 10-1000 light years and over a fair fraction of the sky - see Loeb & Zaldarriaga (2007). The SKA array, due to begin full operation some time after 2025 could also monitor a multitude of directions at once for beamed signals. A good overview of what might be possible in the near future is given by Tarter et al. (2009).
EDIT: I realised I didn't fully address the question. The Arecibo dish can detect the beamed signal you talk about in 1000s at a distance of about 5000 light years. The dish has a diameter of 304m. So to detect a signal that comes from 5 times further away which will be 25 times weaker would naively require a 1.5km dish (assuming the noise levels remain the same).
This paper contains an important analysis of the different trade-off between bandwidth and energy efficiency. The interesting conclusion from that paper is that the most energy-efficient way to send and receive interstellar messages (over flat spacetime) that maximise the bit-rate requires making the bandwidth of transmission very large. In particular, this means that the traditional flavor of SETI that looks for suspicious stand-alone modulated frequency gaussians, might be too naive and restricting, and maybe be also a sub-optimal way of communication
There are other known tricks to enhance the bit-rate of communications at very low power. Using gravitational focal points of stars, it has been calculated and predicted by several authors that even a low-power cellphone could be detected at 10 light-years! So if Bracewell or Von Neumann probes are keeping up a galactic internet, gravitational focal points is where we want to place observers to sniff signals.