I believe he was working on model in 08. He was using the Obama stuff to calibrate the model. That's just what I recall.
Interesting - I thought the idea was that Silver had more data to plug into the model, which made the polling predictions more accurate. This is a take I haven't seen before.
In terms of setting parameters, access to internal polls wouldn't help with factors like historical accuracy or house effect, but it could aid with things like state demographics. That is possible, although Silver's never said so - if so I'd say that's kind of an unfair advantage relative to other models, but not really a critique of the model's validity, since the data were clearly beneficial to its accuracy.
Bottom line is over three elections he hasn't out performed the RCP average.
On a yes/no states and races prediction basis Silver has been only slightly better - which makes perfect sense, it would be a very rare case where the polls pointed one way and an aggregation model the other.
But on the vote margin predictions he's head and shoulders better. His overall error is lower, and 96% of this cycle's state margins were within Silver's 95% confidence interval. Another fairly major example - RCP has the popular vote at O +.7, 538 has O +2.5. Actual margin is O +2.5. The bulk of what 538 does is poll aggregation for sure, but there is substantial added value there.