3 ms·

This is an ad for Partyline. The thesis, sadly, is poorly demonstrated. This should have been called "The present of UI input is text for this particular use c

This is an ad for Partyline. The thesis, sadly, is poorly demonstrated. This should have been called "The present of UI input is text for this particular use case". Text (or speech) as an input method is alright when you know exactly what you want. However, this is not a very good exploration interface, where you need to interact with things that are presented in a 2D/3D space. Text also isn't very expressive when it comes to consuming content. The future of UI, in most cases, is no UI. Let some AI study and understand you, and let it make decisions on your behalf.

meerita 11y ago
I thought exactly the same. There are poor examples and for a strong title there is little substance.
ygra 11y ago
Wolfram Alpha suffers a bit from that, I think: You can query a great lot of things but for many of them it requires you to a) know about that you can query that, b) knowing how to query for that, c) how to give options, parameters. The natural language interface sounds great and makes for awesome demos (like »Wow, they can actually understand that?«), but it's horrible for discoverability, especially in cases where you can try five different ways of saying the same thing but only the sixth would actually yield a result.
- miguelrochefort 11y ago
  The shortcomings of natural language interfaces are obvious, and I cannot understand why so many people believe it to be the future. Humanity needs a formal language, which requires formal thinking. Ambiguity is not a problem when a language is used to browse, explore and clarify thoughts, but this kind of interaction needs to be feedback-driven. You can't just input an ambiguous natural language query and expects accurate results. However, you can start with an ambiguous natural language query, and quickly make it into a formal one thanks to smart auto-completion. Only when the query reach a formal threshold should it be executed.
  - juliendorra 11y ago
    Formalization could be done in interaction with the computer. For a basic example today, we can look at how Google correct your query and suggest other queries, or any other form of auto-complete and auto-correct
brudgers 11y ago
Knowing what I want is hard. Knowing what someone else wants is even harder, and the assumption that it's easy underpins graphic interfaces. Under the traditional model, the menu is only populated with things that a design committee thought I might like to do. I can't compose compound commands or automate multi-step tasks or integrate external tools in front of or after an arbitrary menu pick. Traditional GUI's are usually "See Spot Run" or the entire Oxford English Dictionary. They embody query rather than search. Their static iterative rather than lazy possible. Words are better than pictures of hamburgers and floopy disks because we think in words not icons.
Chris_Newton 11y ago
The future of UI, in most cases, is no UI. Let some AI study and understand you, and let it make decisions on your behalf. Assuming for the moment that this is a desirable outcome, how would you train that AI? Today we have a vast programming industry and software supports a huge amount of what we do, yet we still struggle to write non-trivial software that behaves correctly in all cases. That’s partly because even with precise and unambiguous programming languages, specifying exactly the behaviour we want turns out to be hard. It’s also partly because we sometimes don’t even specify what behaviour we want in advance, because we forgot to consider a certain case or didn’t anticipate a certain interaction. What you’re talking about would require advancing beyond programming with precise technical instructions and into creating systems that can handle complicated but unanticipated situations in the way the humans running the system would choose themselves or at least consider an acceptable or desirable alternative. But how will such a system “study” and “understand” what is required of it in an unanticipated situation? I’m not saying no-one will ever have better ideas than we have today for dealing with these issues, but I suspect we’re a very long way from achieving the kind of general purpose AIs you’re talking about. I can believe that within some of our lifetimes we will see systems that make educated guesses but are on balance more reliable than humans in certain specific cases, perhaps if the people working on self-driving cars continue to make good progress. On the other hand, that’s also a great example of something where some very smart people have been working for several years and with vast resources to handle a single situation that we teach kids to deal with after a few days of training, yet they’re still a long way from the kind of futuristic AI we’re talking about, which perhaps gives a sense of how hard this kind of problem really is.

meerita 11y ago

I thought exactly the same. There are poor examples and for a strong title there is little substance.

ygra 11y ago

Wolfram Alpha suffers a bit from that, I think: You can query a great lot of things but for many of them it requires you to a) know about that you can query that, b) knowing how to query for that, c) how to give options, parameters. The natural language interface sounds great and makes for awesome demos (like »Wow, they can actually understand that?«), but it's horrible for discoverability, especially in cases where you can try five different ways of saying the same thing but only the sixth would actually yield a result.

miguelrochefort 11y ago
The shortcomings of natural language interfaces are obvious, and I cannot understand why so many people believe it to be the future. Humanity needs a formal language, which requires formal thinking. Ambiguity is not a problem when a language is used to browse, explore and clarify thoughts, but this kind of interaction needs to be feedback-driven. You can't just input an ambiguous natural language query and expects accurate results. However, you can start with an ambiguous natural language query, and quickly make it into a formal one thanks to smart auto-completion. Only when the query reach a formal threshold should it be executed.
- juliendorra 11y ago
  Formalization could be done in interaction with the computer. For a basic example today, we can look at how Google correct your query and suggest other queries, or any other form of auto-complete and auto-correct

brudgers 11y ago

Knowing what I want is hard. Knowing what someone else wants is even harder, and the assumption that it's easy underpins graphic interfaces. Under the traditional model, the menu is only populated with things that a design committee thought I might like to do. I can't compose compound commands or automate multi-step tasks or integrate external tools in front of or after an arbitrary menu pick. Traditional GUI's are usually "See Spot Run" or the entire Oxford English Dictionary. They embody query rather than search. Their static iterative rather than lazy possible. Words are better than pictures of hamburgers and floopy disks because we think in words not icons.

Chris_Newton 11y ago

The future of UI, in most cases, is no UI. Let some AI study and understand you, and let it make decisions on your behalf. Assuming for the moment that this is a desirable outcome, how would you train that AI? Today we have a vast programming industry and software supports a huge amount of what we do, yet we still struggle to write non-trivial software that behaves correctly in all cases. That’s partly because even with precise and unambiguous programming languages, specifying exactly the behaviour we want turns out to be hard. It’s also partly because we sometimes don’t even specify what behaviour we want in advance, because we forgot to consider a certain case or didn’t anticipate a certain interaction. What you’re talking about would require advancing beyond programming with precise technical instructions and into creating systems that can handle complicated but unanticipated situations in the way the humans running the system would choose themselves or at least consider an acceptable or desirable alternative. But how will such a system “study” and “understand” what is required of it in an unanticipated situation? I’m not saying no-one will ever have better ideas than we have today for dealing with these issues, but I suspect we’re a very long way from achieving the kind of general purpose AIs you’re talking about. I can believe that within some of our lifetimes we will see systems that make educated guesses but are on balance more reliable than humans in certain specific cases, perhaps if the people working on self-driving cars continue to make good progress. On the other hand, that’s also a great example of something where some very smart people have been working for several years and with vast resources to handle a single situation that we teach kids to deal with after a few days of training, yet they’re still a long way from the kind of futuristic AI we’re talking about, which perhaps gives a sense of how hard this kind of problem really is.