Ganet, Version 2.0

I have lost count of how many times I've typed (I mean, "clicked my up arrow several times and hit Enter," if I'm being honest) git push heroku main in the last two weeks.

Has it only been two weeks?!

It started with a crazy idea, inspired by my 6 year-old nephew's affection for Janet, the endearing anthropomorphized vessel of knowledge from The Good Place. There is something about Janet that my nephew finds charming and appealing, in contrast to his guardedness around Alexa's disembodied voice and wary resistance to Siri (who my sister refuses to enable on his iPad, citing the Pandora's Box she is likely to open).

A few weeks ago when we were reading a book together, my nephew seemed uncomfortable and almost embarrassed that he couldn't pronounce some of the words he came across. Even though I tried to be encouraging, he quickly gave up and moved on to another activity. It got me thinking about how handy (and patient, and non-judgmental, and tirelessly supportive) ChatGPT had been with me when I was stumbling through the basics of web development. I wondered if there was a way I could create an AI "Reading Coach" to help him with tricky words that he could spell but not pronounce, for those times when he was too shy to admit he needed help.

With a ton of input from ChatGPT and several other AI assistants, I built a Flask application and an accompanying website, integrated with ChatGPT and Eleven Labs' text-to-speech API, and, when I finished the first working version, shared it with my nephew. He was immediately intrigued and delighted.

Now, ironically, this was a week before OpenAI was criticized for using a Scarlett Johansson-like voice with ChatGPT, and to be clear, I did not choose a D'Arcy Carden-like voice for Ganet (although D'Arcy's voice is definitely one of the many things that makes Janet so SO awesome). I did, however, introduce "Ganet" to my nephew by comparing her to something I knew he already liked, and, following the "Sky" voice fiasco, I realized something fundamental about the road we have all been going down, as humans encountering shiny new AI-driven things.

Capitalizing on familiarity is what a GPT does. It uses Large Language Models and machine learning to intentionally appear more human, because more human means less scary, and less scary means more likely to be adopted. Even Edward Scissorhands figured this out. Chatbots are not helpful if they are not used, and they are not used if they are creepy or robotic. Make them novel but familiar and even if they are wrong a lot, they can get away with it, simply by still being fun to play with!

Building Ganet taught me a lot about how many times you have to inadvertently break something before you can finally deliver a finished version of it. It reminded me that every success story is really a story about a ton of failure. We don't always get to see the number of times someone got something wrong before they got it right, and I don't think this is healthy - not personally, not professionally. Without witnessing the skinned knees, bruised egos and facepalms, we forget that no one gets it right the first time.

When I first took my nephew rollerskating, I warned him that he was going to fall, and that this was okay and totally expected, because he was learning, and even I fell a lot, when I was learning. I reminded him that you get better at anything you practice at, and that with practice, he would get better and it would get easier. He did fall - a lot - and at one point, he looked at me and said, "When we're falling, it means we're learning." I could not have been more proud.

The first Ganet I showed him was just words on a screen, and instantly I realized that while it worked for me, it would absolutely not work for him, because a kindergartener can't read the sentence, "Audio processed successfully!" All that work and it was dead on delivery... even a low-literacy adult might be challenged by so many syllables! But, he was still delighted. Still excited at the promise - the intention, the need that it was meeting. So I told him that Ganet was going to get better, because we wanted to make sure that she was going to work for him, and this meant that just like him, she would learn and evolve (we might have had to then ask Ganet what "evolve" meant).

Ganet Version 3.0? Coming soon, I hope!

Note: because Ganet is not built for more than 3 concurrent requests, I'm not pasting her URL here. If you'd like a demo, feel free to reach out!