What do you like best about the product?

The hosted aspect is great and if you avoid storing large artifacts, too many histograms, or too many images, it's very cheap or free even for heavy usage. I often run training on remote systems and checking the eval outputs remotely can be a pain with other software like TensorFlow. You need to be sure to run the server on your instance to view the dashboard. With W&B you get comparably powerful features, and you don't need to do anything but run your train script. In addition to hosting, the Sweeps functionality is excellent for hyper-parameter sweeping and pre-determined groups of runs in multi-task settings.

What do you dislike about the product?

Some UI elements, like the Runs data table, can be laggy, and sometimes the auto-refresh seems finicky.

What problems is the product solving and how is that benefiting you?

W&B helps track and debug my experiments. I use it for developing new models, observing loss scales and adjusting them so they're in compatible ranges, detecting and addressing gradient instabilities such as vanishing or exploding. I also use it for hyper-parameter searching to find the best values for my training runs. Additionally, I use the Sweeps functionality for coordinating runs that involve many related models that are deployed together.