Thinking about AI for mobile devices: how much precision can you lose before it’s a problem? How do you balance quantization against the need for accuracy?

The precision is dependent on deployment environment and it varies from usecase to usecase. The mobile devices are constrained in size and computing resources to execute any full fledged AI model. But there are formats available to convert any AI model to be able to run on Mobile devices and this comes at the cost of accuracy and precision. Use cases like objects detection, people count, movement detection can work with lesser accuracy than something like facial detection.. so it comes down to usecase of what's the model is serving.. 

In my opinion, AI is more about approximation rather than precision. We live in a world where information changes by the milliseconds. Misinformation has become normal while ordinary people are struggling to find out what to believe. AI is based on the training provided by the people who programmed it. I am pretty sure it will come biased. 

It depends on the use case, as long as the objective of the program is met, the AI model can be quantized