I used blender 2.49a and this java project(http://sites.google.com/site/gfcaprojects/openstereogram).
With these tools you too can trick people into
staring at a bunch of dots.

I'm just going to
give you the gist for now.

Stereograms need to be somewhat simple to be effective so when
you're modeling your scene think simple big shapes and use depth
with strong foreground and background elements.

We are aiming to obtain something called a depth map.
Essentially, that is data that can be stored in an image where black
represents something at the very furthest point (height = 0) and white is the most near. (height = 1)

The trick to obtaining good depth maps for stereograms is to
adjust the Camera Clipping settings.(Clipping Start/End )
They are values you find on your camera object.
Make sure that the End: setting just barely captures your furthest
element without cutting it off. This will insure that your model
spans the entire range of values for the depth map.

Thus, objects will really 'pop'.

Next, you need to Use Nodes and set up something like this.

Basically, we run our Z depth buffer through a Normalizing vector node
and then invert that to obtain our depth map.
I used an output File node to pull it directly to a file, but you
need not do so. You can simply route it to the Compositor node
and hit the Render button.

It might take a bit of work to get the image just right.
You're aiming for a lot of shading variation like so.

Once you have that...
Run the openstereogram java project

And set Hidden Element to Depth map.
Set your file and click Generate.

And then try to see if it worked.
I must say I had to quickly get good at seeing these
in order to test my output.

Best of luck, and remember this is just the gist.
I'll be motivated to do a more detailed guide when
I'm healthy and thinking straight again.
... and who knows how many more months 'that' will be.
Sheesh.