SXSW Day 2: A case for bots and discovery

Whilst my first day was very top-level, on my second day at #SXSW I chose to attend some more practical sessions. I've been particularly keen to learn more about bots and the application of machine learning.

One of the reasons bots are touted as the next big thing is easily explained with this fact:

2/3 of smartphone users didn't download any apps last month

With bots, you also don't need to:

• ...download anything
• ...create a new account
• ...upload your connections
• ...learn a new interface

This makes bots a very accessible plattform. The biggest hurdle for adoption is that of discovery, but Facebook hinted that they are working that.

How to build the perfect bot

1. Build for conversation, and consider the mindset of the users
2. Start simple. The more features you add-on, the bigger the chance to screw up or make it unclear what the Bot is actually for. Considering the complexity of natural language, this is a key aspect of a successful bot. Try to be creative within the constraints.
3. Consider the users and the plattform (which provides context for the experience).
4. Be social in terms of considering the Bots possible role as extending the IRL activity of the users. It doesn't have to live in an online silo.

What roles does bots have?

The second bot session talked more about it being a great plattform for user feedback and insight. Since there will be many questions that won't be interpreted correctly, or possible to answer, there should be an active feedback loop to continuously improve the bots understanding of its users.

And finally, privacy and sensitive information should be a key consideration. Due to the personable conversations with the bot, people tend to share very personal information when asked (or even unprompted). Consequently it was advised that there'd be a safety team to manage these aspects of running a bot.

Discovery vs exploitation

The music service Pandora ran a session where they shared how they are working with balancing exploitation (giving users more of what they like) vs exploration (suggesting new tunes).

They have an impressive set-up consisting of roughly 35 analysts and 35 musicologists (it's a word...). Whilst the musicologists add meta-data to describe the song, the analyst work on algorithms to connect it all - and also incorporating user feedback. They have about 70 or so different algos, categorised by:

• Collective intelligence (wisdome of the crowd, i.e "people similar to you, also enjoyed...)
• Content-based (analysing the song based on the meta data)
• Personalised filtering (time listened, thumbs up etc..)

A solid day with practical tips, which gave me lots of ideas for incorporating into various projects.

>> If you enjoyed this, you might also want to read about my first day at SXSW: It's all about the greater good or or my third day at SXSW: bringing human-computer interaction into the physical world <<