Yesterday Sobert was migrated to a new version of the Telegram client libraries. This required many changes in the code, and introduced many bugs. One of those bugs caused responses to photos to be triggered when they shouldn't have been triggered. This bug should be fixed now. If you find any other new issues, let me know.
Multiple small incremental brain updates were installed over the last few months. The current training dataset size is 1105MB. There are some new optimizations which improve training performance of the existing model on existing hardware. New experimental training hardware is online and architecture improvements are being worked on.
Sobert's brain has been updated again, this time with 869MB of training data. Also a small problem was found with the previous brain update and things weren't working correctly. This has been fixed now.
A new brain has been installed. 830MB of training data, 228M parameters. No other changes at this time, except some minor fixes in the data preprocessing code.
Another new brain has been installed! 781MB of training data, 228M parameters. Also global_badwords option was renamed to default_badwords and its default value was changed to 10.
