Technology Review - Published By MIT
Advertisement

March/April 2009

TR10: Intelligent Software Assistant

Adam Cheyer is leading the design of powerful software that acts as a personal aide.

By Erica Naone

smaller text tool iconmedium text tool iconlarger text tool icon

Search is the gateway to the Internet for most people; for many of us, it has become second nature to distill a task into a set of keywords that will lead to the required tools and information. But Adam Cheyer, cofounder of Silicon Valley startup Siri, envisions a new way for people to interact with the services available on the Internet: a "do engine" rather than a search engine. Siri is working on virtual personal-assistant software, which would help users complete tasks rather than just collect information.

Weekend plans: Adam Cheyer participates in a conversation with the software. (Go to the next page to read the dialogue and an explanation of the artificial-intelligence behind it.)
Credit: Howard Cao
Multimedia
video  Siri's founders describe the intelligent software assistant.

Cheyer, Siri's vice president of engineering, says that the software takes the user's context into account, making it highly useful and flexible. "In order to get a system that can act and reason, you need to get a system that can interact and understand," he says.

Siri traces its origins to a military-funded artificial-intelligence project called CALO, for "cognitive assistant that learns and organizes," that is based at the research institute SRI International. The project's leaders--including Cheyer--combined traditionally isolated approaches to artificial intelligence to try to create a personal-assistant program that improves by interacting with its user. Cheyer, while still at SRI, took a team of engineers aside and built a sample consumer version; colleagues finally persuaded him to start a company based on the prototype. Siri licenses its core technology from SRI.

Mindful of the sometimes spectacular failure of previous attempts to create a virtual personal assistant, Siri's founders have set their sights conservatively. The initial version, to be released this year, will be aimed at mobile users and will perform only specific types of functions, such as helping make reservations at restaurants, check flight status, or plan weekend activities. Users can type or speak commands in casual sentences, and the software deciphers their intent from the context. Siri is connected to multiple online services, so a quick interaction with it can accomplish several small tasks that would normally require visits to a number of websites. For example, a user can ask Siri to find a midpriced Chinese restaurant in a specific part of town and make a reservation there.

Recent improvements in computer processor power have been essential in bringing this level of sophistication to a consumer product, Cheyer says. Many of CALO's abilities still can't be crammed into such products. But the growing power of mobile phones and the increasing speed of networks make it poss­ible to handle some of the processing at Siri's headquarters and pipe the results back to users, allowing the software to take on tasks that just couldn't be done before.

"Search does what search does very well, and that's not going anywhere anytime soon," says Dag Kittlaus, Siri's cofounder and CEO. "[But] we believe that in five years, everyone's going to have a virtual assistant to which they delegate a lot of the menial tasks."

Story continues below


While the software will be intelligent and useful, the company has no aspiration to make it seem human. "We think that we can create an incredible experience that will help you be more efficient in your life, in solving problems and the tasks that you do," Cheyer says. But Siri is always going to be just a tool, not a rival to human intelligence: "We're very practical minded."

See the 10 Emerging Technologies of 2009.

Comments

  • Great review - something like this exists...
    - Top Down Semantic Web -
    I've actually experienced something like this recently - Check out Headup.com - they do pretty much the same thing but they're a Firefox addon.

    Anyway...
    Siri sounds really great. I'll definitely give it a spin, if and when it becomes available...

    Cheers
    mike
    "I tweet @pop_art"
    Rate this comment: 12345

    popart
    02/25/2009
    Posts:2
    Avg Rating:
    4/5
  • Siri
    The concept of Siri sounds great. Looking forward to it.
    Rate this comment: 12345

    jayvie
    02/27/2009
    Posts:1
    Avg Rating:
    5/5
  • great article
    Great article Erica, thank you for the heads-up! We built a more "to-the-point" virtual intelligent agent software to help our clients reduce their customer service costs. The agent simply responds to user questions and redirects them to the relevant page on the website as it does so. Readers of this article might be interested, so here is our website: Botego
    Rate this comment: 12345

    ekimkaya
    03/01/2009
    Posts:1
    Avg Rating:
    3/5
  • Artificial Intelligence
    I have been designing systems that are TABLE DRIVEN and EVOLVE for over 30 years. That is how you start to create AI. The big problem was getting the authority to create new tables from a DBA. I got around that by creating a table of tables that used a program to view, add, update, and delete tables and their elements using various kinds of Database Management Systems.

    Now with Linux, Apache, Mysql, PHP (LAMP) or (WAMP) that is no longer a problem. I installed it on my PC and I wrote a script that can create just about anything including: databases, and their tables. It backs them up, restores them, modifies their structures, updates their rows, etc.

    Then I created a Data Dictionary that defined things like STATE, ZIP CODE, ETC. It defines objects giving their definition; edit rules, regular expressions, and creating Valid Value tables such as VV_STATE. My table creation script looks at the Data Dictionary to determine parameters for my standard Java Script editing program in the Header of the HTML that it creates and modifies the form creating pull downs of acceptable values populated by corresponding Valid Value tables.

    Any Bozo can use this script to do anything that a DBA can do and once a table is created my scripts will also maintain these objects by creating table maintenance scripts for specific tables using a selection menu.

    Now I could have continued and created a script for entering System Requirements with access to standard definitions of just about any type of table imaginable to allow it to create entire applications, but since that kind of AI is impossible in “EVERYONES” estimation I did not bother.

    After I was through teaching Green Cards from India and Pakistan how to write COBOL UDB programs for Y2K I was considered too old to understand today’s technology. I have been replaced by Green Cards.

    Many years ago I worked for Xerox and worked with MIT’s very expensive COBOL code generation application. I told them that my techniques could allow me and my team to develop systems faster than it could. I proved that was true and Xerox stopped using it. It is a good thing for you people that I do not approve of employment through Social Networking.          
    Rate this comment: 12345

    jadamone@emb...
    03/09/2009
    Posts:3
    Avg Rating:
    4/5
    • Re: Artificial Intelligence
      Sounds like you managed to pull a rabbit out of the hat jadamone. Although I knew something like this was eventually going to come down the pipe, I have, in addition to many other things,  been busy trying to construct an argument around the theory that the evolution of computer coding langauages has run roughly parellel with the evolutionary coding of biological processes. A bit of a long shot shot, perhaps, but no less entertaining. To get right to the point, however, if you would like to pursue the commercialization of this or any other ideas which your playing with, go to my website, www.williameady.com and leave a message. But even if you don't, your feedback on this particular subject has been interesting to say the very least.
      Rate this comment: 12345

      phoenix
      03/10/2009
      Posts:172
      Avg Rating:
      3/5
  • impossible to infer context?
    I admire their goal, but how will they infer context? We need help at the project/task level, but - for example - tools like email and browsers are so multi-purposed that they can be used for anything. It's a kind of Turing test: How do you know when someone's actually working on their computer. All activities appear identical: Browsing, reading, typing, etc.
    Rate this comment: 12345

    matthewcorne...
    03/09/2009
    Posts:1
  • I'm ready!
    If it came out today, the web designer would hit the upload button and instantly see my shadow and hear my shopping cart wheels rushing behind him.

    I have been ready for AI to help with task management, prioritization, appt scheduling, managing too much info, time management, goal management, etc.

    I have some questions about the delay in AI personal assistants. I am not a programmer, so let me preface my question with that bit of info. The question is, why can't a mobile ph, serving as an AI device, learn everything from the person using the phone, ie, build its knowledge base through inquiry? It seems that the interractive response of the device owner, who has a vested interest in teaching, would quickly develop the amount of information necessary for that person. It also seems that based on some previous comments, the number of databases that exist with definitions must be enormous. Would it be possible to have an AI device go through interractive learning throughout the day, and then when not interacting, go through connecting online to relevant databases to create the precise definitions relevant to that days interractive experience.

    It seems that each person would develop a unique personal assistant through this process. Because each one would be unique, they would not have exactly the same definitions or logic about similar info points. Yet, its probable that the combined logic of thousands of these unique assistants could be scored by their owners, and those with the higher scores could become preferred models for that particular info point.

    And again, this would continue an evolution of information gathering and testing of the use of those information models.

    It seems that would be faster and more efficient than any human entering millions of definitions to create one personal assistant. Why not create millions of unique personal assistants?
    Rate this comment: 12345

    jon.s.in.flo...
    03/18/2009
    Posts:1
    Avg Rating:
    5/5
  • AI is only part of a bigger picture
    Ubiquitous Virtual Assistants will certainly come to pass, but AI-based Assistants, though they play a key role, are only part of a bigger picture.

    Providing AI-based virtual support by assembling responses on the fly requires a significant effort in identifying and programming for all possible contexts. As a result, most AI-based Virtual Assistants handle only a small set of tasks.  Furthermore, expanding the number is limited by the availability of specialist programmers and the knowledge that they can develop and codify.

    As a Knowledge Management company, we approached the issue in quite a different way.  Our objective was to enable people with no programming skills, who already have the knowledge in their heads, to create task wizards.  Each wizard can then be called by AI-based Virtual Assistants like Siri, Botégo and Headup, by Google, an embedded link or any other method.

    The resulting product, gStepOne, taps into the existing body of “doing” knowledge – the smartest ways to produce the best outcomes, from billions of web pages (via Google) or directly from people in business, government, clubs, community organisations or individuals.

    People can create wizards to help deliver a service, sell a house, hire staff, conduct an event, adopt a child, take out insurance, apply for a grant – any task where people need an Assistant to help them do things like an expert, without the expert being there.

    They do it by drawing a map of the steps in a task by dragging and dropping shapes onto a page, then linking and naming the steps.  gStepOne Googles the web for supporting “How to” information, training, pictures, videos, maps, blogs, etc. for each step and generates a wizard.  The wizard is then available to guide and support others when they need to do the task.

    gStepOne is free and available from http://www.gstepone.com.  Please check it out!
    Rate this comment: 12345

    GregC
    03/19/2009
    Posts:3
    Avg Rating:
    5/5
  • OS supports for SIRI
    Hi, is this SIRI applicable to MS Windows, Linux or any ohter OS? Can this run on a MAC PC?

    TY
    Rate this comment: 12345

    gaspem
    04/23/2009
    Posts:1

Technology Review Magazine

The 10 Emerging Technologies of 2009
Technology Review presents its annual list of 10 technologies that could change the way we live.

FEATURES

A Zero-Emissions City in the Desert
Oil-rich Abu Dhabi is building a green metropolis. Should the rest of the world care?
By Kevin Bullis
But Who's Counting?
No one really knows how many people visit websites. A San Francisco startup and Google are both working to change that.
By Jason Pontin

Read more articles from this Issue

NOTEBOOKS FORWARD ESSAY TO MARKET Q&A PHOTO ESSAY REVIEWS HACK DEMO
Archives MIT News Subscribe Contact

Log In

Forgot your password?     Register »
Advertisement

Videos

The Marcellus Shale Gas Rush
Technology Review November/December 2009

Current Issue

Natural Gas Changes the Energy Map
The United States has vast supplies of this cleaner fossil fuel. But how should we use it?
Advertisement
Advertisement
Subscribe to Technology Review's daily e-mail update. Enter your e-mail address

TECHNOLOGY RESOURCES

More Technology News from Forbes

Advertisement
MIT Massachusetts Institute of Technology © 2009 Technology Review. All Rights Reserved.