My training data was really small, though - I could only find about 360 existing messages total. So, when I decided to generate some more messages, I decided to add my favorite neural network-generated messages to the original dataset, increasing the total to almost 500. This is still really small for a neural network, but a noticeable upgrade.
The first thing I noticed is the definite upswing in the number of messages involving bears. In fact, I’m seeing worrying signs of a bear-based feedback loop that might lead to 100% bear content after a few more iterations of this.
BE MY BEAR LOVE BEAR STACK BEAR LOOK BEAR WINK BEAR HOME BEAR TWEET BEAR TAME BEAR DEARY BEAR BEAR BEAR TIME BEAR
In addition to the bears, the neural network still managed to produce some complimentary messages.
ME LOVE HAVE SWEET TANK OOO NICE CUTEY LIDS YOU RANK BEST MANE HEAT TEAM HOT GIVE
And some that are less easy to interpret.
BEE O O MAGE LOVE IN A FAN BOOK WEAR ME LET’S RIND CLOUD ME YAK O WAY FRARR ME YOU ARE BARE SWEET BOG BEELT POSWRORD U?GHCLENCY U
Use these with caution.
MEAT LOVE TOOL BIT DIRTY LOVER LOVE BAN SICK WINK OY LOVE BUM MY HAG YOU REAR FAN ME F LOVE LOVE BAG OR MY BUN COIN 2 LOVE SWEET MEAT