Better explain Theano shared variables

Author: zaxtax

docs/source/notebooks/api_quickstart.ipynb

@@ -1496,7 +1496,9 @@
    "source": [
     "## 4.1 Predicting on hold-out data\n",
     "\n",
-    "In many cases you want to predict on unseen / hold-out data. This is especially relevant in Probabilistic Machine Learning and Bayesian Deep Learning. While we plan to improve the API in this regard, this can currently be achieved with a `theano.shared` variable. These are theano tensors whose values can be changed later. Otherwise they can be passed into PyMC3 just like any other numpy array or tensor."
+    "In many cases you want to predict on unseen / hold-out data. This is especially relevant in Probabilistic Machine Learning and Bayesian Deep Learning. While we plan to improve the API in this regard, this can currently be achieved with a `theano.shared` variable. These are theano tensors whose values can be changed later. Otherwise they can be passed into PyMC3 just like any other numpy array or tensor.\n",
+    "\n",
+    "This distinction is significant since internally all models in PyMC3 are giant symbolic expressions. When you pass data directly into a model, you are giving Theano permission to treat this data as a constant and optimize it away as it sees fit. If you need to change this data later you might not have a way to point at it in the symbolic expression. Using `theano.shared` offers a way to point to a place in that symbolic expression, and change what is there."
    ]
   },
   {
@@ -1582,8 +1584,8 @@
  "metadata": {
   "anaconda-cloud": {},
   "kernelspec": {
-   "display_name": "Python 3",
-   "language": "python",
+   "display_name": "IPython (Python 3)",
+   "language": "",
    "name": "python3"
   },
   "language_info": {
@@ -1596,7 +1598,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.6.1"
+   "version": "3.5.2"
   }
  },
  "nbformat": 4,