Updated SwiftKey keyboard update gives access to Microsoft's new Bing AI inside any app

Comments

• Reply 1 of 6

  mike1 Posts: 3,478member

  April 13, 2023 8:24PM

  Does the app send still send every keystroke, not just those for search, back to MS?
  

  

   2Likes 0Dislikes 0Informatives
• Reply 2 of 6

  sxygayjust Posts: 38member

  April 13, 2023 8:28PM

  Didn't they initially announce they were discontinuing it a few months ago, only to change their mind? Maybe they realized they could do this and gather more info..

   0Likes 0Dislikes 0Informatives
• Reply 3 of 6

  anonconformist Posts: 202member

  April 14, 2023 3:58AM

  JP234 said:

  This seems on the face of it to be nothing more than a slightly more intrusive form of autocorrect.

  If this provides full access to Bing Chat, that’s an absurdly grossly overly-simplistic extreme understatement of what can be done.
  
  I’ve created a working console version (in Swift) of my personal invention of a form of Solitaire by describing the rules and requirements for how it behaves. That took several sequential prompts for the first attempt. It included as close to an “Attract Mode” of it playing itself as feasible in a console application with Swift not having a standard way to check if a key has been pressed without waiting for input. I prompted for the Attract Mode, but it alerted me to the limitations with explanation. I asked it to specially print the possible valid moves during the turns.
  
  I didn’t write a single line of code to do this, I did this all in English. This is a solitaire game that has run on only my own iOS and MacOS devices in the past, implemented in Objective-C, but not released to the world because I hadn’t spent the time to make it nice and shiny.
  
  So, to a limited extent, calling it autocorrect on steroids is partially true, but in any case, it has generative capabilities that far exceed anything you’ve used for autocorrect before. Hell, I’ve had it create song parodies in the style of various characters, human and otherwise, I’ve prompted it to create 4-person love songs for a fictional race with 4 genders, and it did it. Yesterday I had it create a parody of “Thriller” incorporating “The Hokey Pokey” without it using the original “Thriller” lyrics explicitly. And no, it’s not limited to 4 characters/voices in a song, I’ve had it do more.
  
  Within certain intentional constraints for not offending people somehow (guardrails) and other legalities, while it has an upper limit for the number of tokens (roughly equal to words, not always) in the context, if it has been trained via data how to do something, you can have it do that. That includes teaching it via prompts what to do: for example, I didn’t like the code formatting style it generated, so I asked it first to reformat the Swift code according to the WebKit C++ standard that it looked up online, and it told me Swift has certain requirements for syntax that’d be affected, so after I prompted a bit different, and it neatly fulfilled that code formatting style while (not what I had in mind!) translating the Swift code into C++ code, which I had to clarify wasn’t what I wanted, so it retranslated back into Swift. There were still some formatting style rule changes I wanted implemented, as that wasn’t completely the style I wanted: it did it.
  
  no autocorrect is that capable: it was following directions as specified in English to generate code and reformat it in a way I’d hope a human would be able to do, but it was faster, no typos, and filled in details I didn’t painstakingly explain as well as the ones I did.

  

   1Like 0Dislikes 1Informative
• Reply 4 of 6

  lowededwookie Posts: 1,196member

  April 15, 2023 8:31PM

  JP234 said:

  anonconformist said:

  JP234 said:

  This seems on the face of it to be nothing more than a slightly more intrusive form of autocorrect.

  If this provides full access to Bing Chat, that’s an absurdly grossly overly-simplistic extreme understatement of what can be done.
  
  I’ve created a working console version (in Swift) of my personal invention of a form of Solitaire by describing the rules and requirements for how it behaves. That took several sequential prompts for the first attempt. It included as close to an “Attract Mode” of it playing itself as feasible in a console application with Swift not having a standard way to check if a key has been pressed without waiting for input. I prompted for the Attract Mode, but it alerted me to the limitations with explanation. I asked it to specially print the possible valid moves during the turns.
  
  I didn’t write a single line of code to do this, I did this all in English. This is a solitaire game that has run on only my own iOS and MacOS devices in the past, implemented in Objective-C, but not released to the world because I hadn’t spent the time to make it nice and shiny.
  
  So, to a limited extent, calling it autocorrect on steroids is partially true, but in any case, it has generative capabilities that far exceed anything you’ve used for autocorrect before. Hell, I’ve had it create song parodies in the style of various characters, human and otherwise, I’ve prompted it to create 4-person love songs for a fictional race with 4 genders, and it did it. Yesterday I had it create a parody of “Thriller” incorporating “The Hokey Pokey” without it using the original “Thriller” lyrics explicitly. And no, it’s not limited to 4 characters/voices in a song, I’ve had it do more.
  
  Within certain intentional constraints for not offending people somehow (guardrails) and other legalities, while it has an upper limit for the number of tokens (roughly equal to words, not always) in the context, if it has been trained via data how to do something, you can have it do that. That includes teaching it via prompts what to do: for example, I didn’t like the code formatting style it generated, so I asked it first to reformat the Swift code according to the WebKit C++ standard that it looked up online, and it told me Swift has certain requirements for syntax that’d be affected, so after I prompted a bit different, and it neatly fulfilled that code formatting style while (not what I had in mind!) translating the Swift code into C++ code, which I had to clarify wasn’t what I wanted, so it retranslated back into Swift. There were still some formatting style rule changes I wanted implemented, as that wasn’t completely the style I wanted: it did it.
  
  no autocorrect is that capable: it was following directions as specified in English to generate code and reformat it in a way I’d hope a human would be able to do, but it was faster, no typos, and filled in details I didn’t painstakingly explain as well as the ones I did.

  If only I had the mental capacity to understand what you write. Hell, I can't even finish reading it!

  He told it to reformat code in a way that was more visually appealing to him. It wasn’t quite right but ended up close enough.
  
  He had written a Solitaire game but there were certain aspects that weren’t quite right.
  
  In each case AI solved his problems for better or worse but it was something far more achievable than Autocorrect.

   0Likes 0Dislikes 0Informatives
• Reply 5 of 6

  anonconformist Posts: 202member

  April 16, 2023 12:30AM

  lowededwookie said:

  JP234 said:

  anonconformist said:

  JP234 said:

  This seems on the face of it to be nothing more than a slightly more intrusive form of autocorrect.

  If this provides full access to Bing Chat, that’s an absurdly grossly overly-simplistic extreme understatement of what can be done.
  
  I’ve created a working console version (in Swift) of my personal invention of a form of Solitaire by describing the rules and requirements for how it behaves. That took several sequential prompts for the first attempt. It included as close to an “Attract Mode” of it playing itself as feasible in a console application with Swift not having a standard way to check if a key has been pressed without waiting for input. I prompted for the Attract Mode, but it alerted me to the limitations with explanation. I asked it to specially print the possible valid moves during the turns.
  
  I didn’t write a single line of code to do this, I did this all in English. This is a solitaire game that has run on only my own iOS and MacOS devices in the past, implemented in Objective-C, but not released to the world because I hadn’t spent the time to make it nice and shiny.
  
  So, to a limited extent, calling it autocorrect on steroids is partially true, but in any case, it has generative capabilities that far exceed anything you’ve used for autocorrect before. Hell, I’ve had it create song parodies in the style of various characters, human and otherwise, I’ve prompted it to create 4-person love songs for a fictional race with 4 genders, and it did it. Yesterday I had it create a parody of “Thriller” incorporating “The Hokey Pokey” without it using the original “Thriller” lyrics explicitly. And no, it’s not limited to 4 characters/voices in a song, I’ve had it do more.
  
  Within certain intentional constraints for not offending people somehow (guardrails) and other legalities, while it has an upper limit for the number of tokens (roughly equal to words, not always) in the context, if it has been trained via data how to do something, you can have it do that. That includes teaching it via prompts what to do: for example, I didn’t like the code formatting style it generated, so I asked it first to reformat the Swift code according to the WebKit C++ standard that it looked up online, and it told me Swift has certain requirements for syntax that’d be affected, so after I prompted a bit different, and it neatly fulfilled that code formatting style while (not what I had in mind!) translating the Swift code into C++ code, which I had to clarify wasn’t what I wanted, so it retranslated back into Swift. There were still some formatting style rule changes I wanted implemented, as that wasn’t completely the style I wanted: it did it.
  
  no autocorrect is that capable: it was following directions as specified in English to generate code and reformat it in a way I’d hope a human would be able to do, but it was faster, no typos, and filled in details I didn’t painstakingly explain as well as the ones I did.

  If only I had the mental capacity to understand what you write. Hell, I can't even finish reading it!

  He told it to reformat code in a way that was more visually appealing to him. It wasn’t quite right but ended up close enough.
  
  He had written a Solitaire game but there were certain aspects that weren’t quite right.
  
  In each case AI solved his problems for better or worse but it was something far more achievable than Autocorrect.
  
  I've done notably more than that, but that's a good start. It's largely limited by how precise you are in directions and how much memory (context) it has to work with.
  
  I've thought about how to simplify what it really is: as autocomplete is to words, Bing Chat using GPT-4 is autocomplete for complex thoughts once you've provided enough context.
  
  "I know, let's make a parody of a Michael Jackson song, in the voice of Weird Al!."
  
  Name the song, you can let it choose the direction of the parody based just on that alone and related words. It'll create something that fits amazingly well. Add more context, that gets added into the generated output: Thursday I instructed Bing Chat to create a parody of "Thriller" incorporating "The Hokey Pokey" and it did it quite well, asking it to explicitly not reuse any lyrics from "Thriller" to comical results that made sense.
  
  Here is what is important to understand: autocomplete and Auto-correct for your personal usage only has the information from you, while Bing Chat and GPT directly are trained on a huge portion of the public internet.
  
  The result of having all that data for training has surprising results. I'm studying how it works bit by bit using what I have available to study, which leaves quite a few details missing. Here's the thing: the creators, researchers don't fully understand how it does all the things it does, as quite a bit of the functionality is what's called "emergent behavior" as it wasn't there by explicit design, it just exists with no clear idea how that happened.
  
  There is a reason many have signed a petition calling for the halt on creating anything more powerful than GPT-4: GPT-4 released to the public has actually been weakened as a side-effect of trying to reduce the odds it offends people.
  
  GPT-5 is slated to be even more powerful, trained with more data. There are projections it'll develop more emergent behavior. What it actually turns out to be could be interesting.

   0Likes 0Dislikes 0Informatives