3 ms·

We do have a lot of hype and buzz. There's a giant AI hype bubble right now and it's going to pop and bring us back to reality. But at the same time, people a

by retrac 4y ago

We do have a lot of hype and buzz. There's a giant AI hype bubble right now and it's going to pop and bring us back to reality. But at the same time, people are flipping out because it is, well, machine learning. I was following AI, mostly as a hobby interest, though I did some undergrad courses, before it got all buzz-wordy. Just ten, even five years ago, some of the stuff being demonstrated, were the sort of thing I thought would take many decades to develop, and that I may not see in my lifetime. Speech recognition, machine translation, machine vision, for example. For more than half a century, good machine translation and speech recognition were the holy grail of AI. An incredible amount of work went into them. Specialized systems, grammars and parsers, programmed general knowledge databases. Collecting and correlating enormous aligned bilingual text corpuses with phonetic annotation. After several decades of such work, it got us to mediocre-quality Google Translate style services that were the norm until recently. Now, without any of that specialized work, with a much more general learning algorithm, there are translation systems that are superior to anything humans designed to specifically be machine translators. It just falls out accidentally from the large language models -- they can also, incidentally, translate, and do better than anything specifically designed to translate ever did.

matheweis 4y ago
> Speech recognition, machine translation, machine vision, for example. All basically solved problems in the last few years. Is this true or is there just a massive gulf in application at the moment? For one trivial example take something like indexing historical newspapers that you would really expect to be in the class of “solved problems” because it’s largely typeset, but the commercial offerings [1,2,3] are just chock full of errors and there is absolutely nothing I can apply off the shelf that gets better results without a ton of extra effort and compute. 1. Newspapers.com 2. Newspaperarchive.com 3. Genealogybank.com
- alophawen 4y ago
  > Is this true or is there just a massive gulf in application at the moment? It is only true if you drink the cool aid. Speech recognition - Siri still have major issues undestanding me. Youtube text2speech routinely mistranscribes because it simply have no understanding of the language. Machine translation is hilarious at best and dangerously wrong at worst. Machine vision still could not spot the difference between pastry and furry animals last time I checked. All of these are examples of non-working over hyped tech. It is not a list of "basically" solved problems.
  - pixl97 4y ago
    You have some odd issue here... You're thinking that any technology is going to be perfect, and it's not, humans are not either. Don't put your base line as perfect, but at the rate of human failure. Siri isn't even the latest models that have a much lower error rate.
  - OtmaneBenazzou 4y ago
    Whisper is better at speech recognition than humans. Learn about the SOTA instead of mentioning bad mainstream products made years ago.
  - HyperSane 4y ago
    My company has replaced hundreds of human transcribers with a speech to text model. DeepL is actually really good a lot of the time. When was the last time you checked the status of machine vision, because the problem you suggested is not hard for it anymore.
    - alophawen 4y ago
      Your company can only pull the rug like this because the public is not very picky it seems. A real human transcriber still outperforms any automated system in existence.
      - HyperSane 4y ago
        > Your company can only pull the rug like this because the public is not very picky it seems. My company performed extensive accuracy testing and lets people choose to use a human transcriber if they want. Most people are perfectly happy to use the machine transcription. You are just seem bitter. Did an AI steal your wife?
  - brabel 4y ago
    I must have drank the cool aid because when I talk to my AI assistant, it bloody well understands me a lot better than my wife :D. Machine translation still makes a few mistakes but hardly more than a human. Machine vision: I worked on a factory where a machine would approve/reject products passing through at a ridiculously high speed, and it never got one wrong. This was 15 years ago. Your experience is completely the opposite of mine.

by retrac 4y ago

matheweis 4y ago

> Speech recognition, machine translation, machine vision, for example. All basically solved problems in the last few years. Is this true or is there just a massive gulf in application at the moment? For one trivial example take something like indexing historical newspapers that you would really expect to be in the class of “solved problems” because it’s largely typeset, but the commercial offerings [1,2,3] are just chock full of errors and there is absolutely nothing I can apply off the shelf that gets better results without a ton of extra effort and compute. 1. Newspapers.com 2. Newspaperarchive.com 3. Genealogybank.com

alophawen 4y ago
> Is this true or is there just a massive gulf in application at the moment? It is only true if you drink the cool aid. Speech recognition - Siri still have major issues undestanding me. Youtube text2speech routinely mistranscribes because it simply have no understanding of the language. Machine translation is hilarious at best and dangerously wrong at worst. Machine vision still could not spot the difference between pastry and furry animals last time I checked. All of these are examples of non-working over hyped tech. It is not a list of "basically" solved problems.
- pixl97 4y ago
  You have some odd issue here... You're thinking that any technology is going to be perfect, and it's not, humans are not either. Don't put your base line as perfect, but at the rate of human failure. Siri isn't even the latest models that have a much lower error rate.
- OtmaneBenazzou 4y ago
  Whisper is better at speech recognition than humans. Learn about the SOTA instead of mentioning bad mainstream products made years ago.
- HyperSane 4y ago
  My company has replaced hundreds of human transcribers with a speech to text model. DeepL is actually really good a lot of the time. When was the last time you checked the status of machine vision, because the problem you suggested is not hard for it anymore.
  - alophawen 4y ago
    Your company can only pull the rug like this because the public is not very picky it seems. A real human transcriber still outperforms any automated system in existence.
    - HyperSane 4y ago
      > Your company can only pull the rug like this because the public is not very picky it seems. My company performed extensive accuracy testing and lets people choose to use a human transcriber if they want. Most people are perfectly happy to use the machine transcription. You are just seem bitter. Did an AI steal your wife?
- brabel 4y ago
  I must have drank the cool aid because when I talk to my AI assistant, it bloody well understands me a lot better than my wife :D. Machine translation still makes a few mistakes but hardly more than a human. Machine vision: I worked on a factory where a machine would approve/reject products passing through at a ridiculously high speed, and it never got one wrong. This was 15 years ago. Your experience is completely the opposite of mine.