Hi Joelgrus,
First of all, thanks a ton for your astonishing work. I was playing around with the network and found that the non-linearity in the output layer (sigmod or softmax) would stop the learning process. How much ever I train the network, its not learning anything. Cost is not going down and accuracy is not going up. Any thoughts?

In Tensorflow 1.0.0, throws

File "fizzbuzz.py", line 52, in <module>
    cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(py_x, Y))
  File "/Users/mstraeubig/miniconda3/envs/tensorflow3/lib/python3.6/site-packages/tensorflow/python/ops/nn_ops.py", line 1578, in softmax_cros
s_entropy_with_logits
    labels, logits)
  File "/Users/mstraeubig/miniconda3/envs/tensorflow3/lib/python3.6/site-packages/tensorflow/python/ops/nn_ops.py", line 1533, in _ensure_xent
_args
    "named arguments (labels=..., logits=..., ...)" % name)
ValueError: Only call `softmax_cross_entropy_with_logits` with named arguments (labels=..., logits=..., ...)

due to API changes. Running tf_upgrade.py fixes this.

I think it may work better if you solve problem different way.
Instead creating 1 input it may be better to split number in separate digits.
I think it is more probable for network to detect divisibility rules with that input.

Last contribution was about a year ago and vital issues are still left open.

Is this project going to receive further support?