The problem with Jeopardy!

The I’ve-been-blogging-for-five-years-but-am-still-paranoid-about-work-related-posts-being-misrepresented disclaimer:
I do work for IBM as a developer on Watson but that role doesn’t extend to this blog. My responsibilities for Watson are limited to writing code… so any ramblings here are my personal views, and not necessarily representative of IBM’s positions, strategy or opinions.

The first time most people saw or heard of Watson was on “America’s favourite quiz show”, Jeopardy!

And that association seems to have stuck. For the moment, at least, Watson’s identity seems to bound up with Jeopardy’s. It’s “the computer system that won on Jeopardy!”

I’m not sure that’s entirely a good thing.

Jeopardy! was a great demonstration of Watson’s capabilities in a lot of ways. It showed the breakthroughs in interpreting natural language, the breakthroughs in coming up with evidence and rationale and identifying the level of confidence in it’s answer’s, and the power of machine learning systems.

But… it does make it easy to misunderstand the future potential of the technology.

The Jeopardy! challenge was about man vs machine.

It was about a machine trying to beat the humans.

A machine that had to work entirely by itself to come up with answers.

And if it couldn’t come up with an answer immediately, or when didn’t have enough confidence in the answer that it did come up with, it’d give up, stay silent and not buzz in.

None of that describes the way a future Watson could work. In my view, Watson should be about man and machine working together.

Watson could help to make it’s users smarter.

When a Watson system returns it’s answers, it could be showing it’s users the sources that it found that were relevant to the question – showing the passages from documents that the user might not have read or even heard of, helping to make them more informed.

It could show the reasoning for why it had it’s level of confidence in the answer, helping them to consider new perspectives.

And it works both ways – the users will make Watson smarter.

When a Watson system doesn’t have enough evidence to be confident in it’s top answer, it wouldn’t have to give up there. Every interaction wouldn’t have to be a single, stand-alone question. Watson could ask follow-up questions, working with the user to gather more information, more context and more evidence, until the level of confidence in the top answer improves.

Not only that, the machine learning in Watson would mean that every time a user provided a correct answer, Watson would get that little bit better and smarter at handling questions like that in the future.

Reciprocal. Collaborative.

These are not the words that Jeopardy! brings to mind. But it’s the vision of the future that I look forward to.

Watson doesn’t need to be about providing a definitive, take-it-or-leave-it answer without any additional information. It could be about helping decision makers to have the relevant stuff from terrabytes of text, right at their finger tips.

Watson could be a well-informed assistant for human experts.

More like a librarian who knows all of the text that has been written about your field of interest, and can work with you to find just what you need to answer the question that you have.

Does that vision come across in a man vs machine challenge?

IBM has uploaded a number of Watson-related videos to YouTube. Some of the comments that get posted on them are positive.

Others, however…

That isn’t the only one. There are loads like this.

Or…

To be fair, not all of them are entirely serious.

Okay… let’s be honest. I’m using comments from YouTube to back up an argument.

This is kinda cheating. There isn’t a point of view too crazy that you couldn’t find some commenter on YouTube that would back it up.

And I like I say, there are positive comments there, too, so I don’t want to blow this out of proportion. It’s not actually that bad.

But I do think introducing Watson to people in an adversarial, man vs machine format, has had some impact on the way it has been perceived. And there are plenty of more thoughtful, considered pieces in the mainstream press that talk about Watson being better than doctors, rather than how a Watson system could be a tool to make doctors more effective.

It’s not just the Internet crazies that are looking at this through a lens of “Watson was better than Ken and Brad on Jeopardy, and it beat them. Next it’s gonna be better than my doctor and beat them at medicine.” And in an era of outsourcing, there are people writing that they fear Watson as an attempt to save money by replacing humans with computers.

It’s been well reported that the computer on Star Trek was an inspiration for many of the original researchers working on the Watson project.

Characters on Star Trek had a computer that they worked with collaboratively, interacting with it using natural language. They would ask the computer questions, like “why is X happening?” or “what will happen if we do Y?”. It wasn’t a vision of a future in which computers were replacing people, or implying that people will be subservient to computers.

That’s the sort of vision we need to explain the potential of Watson.

Or, to pick a more contemporary example, maybe iPhone’s Siri is the best thing that could have happened for Watson. It puts the idea of question answering in people’s hands, and is getting people talking about the potential without worrying about it trying to beat us or replace us.

When we’re ready to show the world the first proof-of-concepts applying Watson to real-world problems, it’ll be interesting to see how it’s perceived, and whether the man-vs-machine image sticks.

Tags:

5 Responses to “The problem with Jeopardy!”

  1. David Colbourn says:

    I for one am thrilled to have the likes of Watson as a co researcher and to take on the job of subject matter expert model calibrator. Errors of context and early adoption threshold management or paradigm shift management problems seem to be in full retreat in the face of this project. I would love to get more information on how the unstructured data is incorporated and any taxonomy of language insights the team has made that could help business prepare for such an approach.
    I see Watson as a sounding board for when synthetic biology starts expanding beyond the human brain model. I see a host of ‘synthetic existential anxiety’ issues arising from model calibration backup and recovery and log roll backs in the area of synthetic biology. I may be wrong to see advanced synthetic biology as being a self aware conscious intelligent life form in a moral vacuum who could get emotionally pissed off, but that is how I see it. Retaining an unemotional non self aware more deterministic Watson has security implications until the moral vacuum can be filled and the loss of a sex life and death can be understood.

    Do you know anyone working in this area?

  2. David Colbourn says:

    Are there any rule sets or general guidelines relating to cohearency and consistency of metaphors comming out of this work that could help semantic web efforts with an eye toward indexing of unstructured data or is the effort a brute force but tuned parallelism like netezza?

    Any guidelines on when aggeration or preprocessing or meta data is used in decision making instead of raw data?

    The work on Watson has profound implications for data modeling, Who is the best person to speak to about this?

  3. [...] said before that I’m not revealing anything new about Watson here. I’m a code monkey not a Press [...]

  4. David Colbourn says:

    Even Code Monkeys have a contact list.

    Is there anyone who could discus the general approaches that Watson has taken toward unstructured data?

    Additionally is there anyone you could recommend in the AI field that has a good grasp of the different approaches?

    This is a wildly interesting field.

  5. dale says:

    Quora looks like it might be a good place to discuss things. If you want something more official, the place to start would be ibmwatson.com.