The tech you're referring to was to increase or decrease polygons based on the power of the users machine. The big benefit to this tech would be that when you got a new graphics card the graphics would scale up to match the number of polygons generated for the models. Not sure about the photo realistic part but I do remember them talking about how this tech was going to be the best looking because of the increased polygon count.
Considering the jump graphics hardware has taken over the last decade, I don't find Tim Sweeney's prediction that photo-realism within our lifetimes seems like an impossibility. Barring some kind of physical limitation with the hardware I think his assumption isn't that crazy.
In the end, it didn't sell particularly well, it cost a lot to make, and it had a lot of performance problems.
Have to wonder who thought a game about a cherub named "bob" would sell like crazy.
When Quake1 models were a couple hundred polygons, it was sometimes predicted games would never need more than 5,000 to 10,000 poly models. There was a PC game (name escapes me, had a cherub in it as the main character) from the late 1990s that talked up using "real time polygon tessellation" to deliver models with thousands of polygons on the screen, derived from models built in Maya that had tens of thousands of polygons in the original spec model. According to the developers, it was almost photo realistic! Such is the braggadocio in the gaming industry.
Now we're up to 15,000 to 50,000 in-game depending, and that's not nearly enough.