A Robot in Your Pocket
Jeff Bonforte, CEO Xobni, @bonforte
Amit Kapur, CEO Gravity, @amitk
· Marvin Minsky
o In the 50s, predicted robots would be everywhere in 5 years
o In the 60s, it was 10 years
o In the 70s, it was 20 years
o In the 80s, it was 40 years
· It’s a fine line between tools and robots
o Robata is Czech for “hard work”
o It’s a fine line between a tool and a point where it becomes something that works for you.
· We think of robots as a hardware thing
o We want R2D2, Rosie, and Six.
o What we have are vacuum cleaners and industrial robots.
· They’re here, and they’re software.
· What’s changed in the last decade?
o Smaller and cheaper sensors.
o The more things we measure, the more accurately we can respond.
o Smartphones are a collection of sensors we carry with us all the time.
· Software, too.
o Natural Language Processing: Understanding semantically what something is about.
o Machine Learning: Software can look at data, learn from it, do intelligence tasks.
o Distinct Ontologies: Instead of a rigid taxonomy, … Humans don’t think in hierarchical structures. We think flexibly. An iRobot vacuum makes us think about things like chores, and how we don’t have time, and the cost of hiring a may.
§ Machines need to be able to understand and combine things.
· More data than we know what to do with.
o We start by measuring things we don’t know what to do with.
o Will it rain today?
§ It’s a deterministic problem. Use barometer, wind conditions, etc.
§ Stochastic: Look at 10 million shoe selections of New Yorks, and you can figure out if it’s going to rain.
· The point of stochastic is that one data point doesn’t matter. Whereas in a deterministic model, you could crash your model with a weird data point.
· After 24 hours, shoe selection is not correlated to weather.
§ The point is, we can correlate surprising things.
o Xobni does this with inboxes. The average inbox is a couple of megabytes. The Xobni inbox has 40 MB of data.
· Explicit versus implicit data
o “I’m here at this restaurant”, or “this is my favorite person”
o We look at your data, and observe what you do. If you text a person 1,000 times to the same number, why does the phone still ask you which number to use?
o Examples of implicit data:
§ Payment patterns from credit cards
§ Locations you pass when you drive, locations you stay a long time.
§ You express your preferences and patterns through what you do every day.
o For example: let’s say I get a txt message from someone with a link. How often do I click on links from that person? If it’s high, then go fetch the page in the background, so that when I do click on it, the page is preloaded.
o Implicit systems are much more accurate, because they are related to current behavior and actual actions, rather that what people think they are interested in, or what they explicitly said 2 years ago.
o Features like circles in google are explicit and they cause high cognitive load.
· Where giants tread
o IBM’s Watson.
§ Smart AI can win Jeopardy.
§ Now diagnose cancer.
o Google’s self-driving car.
§ Passes 300,000 miles driven.
o Huge R&D budgets, years of efforts.
· Startups coming into the equation
o The cost of getting technology and processing data is going down
o More tools are open source
· Big R&D innovations feel like they’re five years away, but it’s usually 10 years.
o Example of iDrive: cost and effort to do ($5.7M for 16 terabyte drive, $1.5M monthly bandwidth bill, write every component of systems) versus Dropbox ten years later (off the shelf components, cheap costs).
o Analogy: Brakes
o Digital: Antilock
o Robot: Crash avoidance
o Analog: thermostat
o Digital: timer thermostat
o Robotic: Nest
o Analog: Newspapers.
o Digital: Online representation.
o Robot (gravity): Personalized experience based on their preferences, derived from their past behavior
o A: Yellow pages
o D: Yelp
o R: Ness
o A: Encyclopedia
o D: Google Search
o R: Google Now
o A: Address book
o D: Contacts / Email
o R: Xobni
· A spam filter that’s 95% accurate is totally unreliable. 0% adoption. At 98%, still bad. 99%, still bad. You need to get to 99.8% before you get adoption.
o But for restaurant selection, 95% is great.
o Different level of expected quality for different systems.
o Personalizing the Internet
o Marissa Meyer saying that Yahoo is going to be very focused on personalization.
o Surrounding ourselves with the experts in machine learning, natural language processing.
o Mission: leverage the interest graph to personalize the internet
o The more information that flows into a system, the harder it becomes to find great content. It’s the signal to noise ratio.
o The history of the internet is of companies creating better filters to find great content.
§ Their web: directories, google.
§ Our web: use social graph, get content shared with us from friends
§ Your web: using technology to process data to understand the individual, and have adaptive, personalized experience.
o Interest Graphing
§ Semantic analysis of webpage. Match against ontologies we’ve built.
§ Watch what people do, match against interests.
§ Then personalize what they see.
§ Show examples of how sites filled with links (New York Times, Huffington Post), Gravity will find the top articles you’d be interested in.
o Why who matters?
§ It starts with the idea of attaching files to email. You know the sender, the receiver, and the email. Instead of presenting all files on the computer for possible attachment, you can prefilter the list, and it’s a 3x reduction in possible files.
o Super cool demo of voicemail.
§ Voicemail transcribes and hotlinks to contacts, doing things like resolving references to email (“see the address I emailed you”), and people (the venndiagram of people they know in common means they must be talking about this Chris), and vocabulary (this two people use words like “dude”, and “hey man”)
· Future Applications
o Trackers are digital. What’s the robot version? The equivalent of a check engine light for your health.
o Education: the creation of personalized education and teaching.
o Finance: help for your particular financial situation.
· Often people are worried about privacy. Anytime you give people data, you have to worry to what are they going to do.