3 ms·

[(partially) tongue-in-cheek] After working in AI research for about 10 years, I suspect that most of what you see (especially outside "scientific" literature)

[(partially) tongue-in-cheek] After working in AI research for about 10 years, I suspect that most of what you see (especially outside "scientific" literature) is... very sophisticated smoke and mirrors. After reading this and seeing the (impressive!) video, I'm starting to think it was smoke and mirrors and hype since the beginning, and even earlier than that.

nkoren 8y ago
Yeah, the way that voice commands were recognized via pattern-matching the syllabic rhythms rather than via the actual meaning of the commands was, in hindsight, kind of cheating. But consider the degree to which today's neural nets are simply engaging in (admittedly far more complex) pattern-matching, with little more understanding of the actual meaning than Elektro. A lot of today's AI researchers/enthusiasts/hypesters like to imagine that there isn't really anything fundamental missing from this picture, that it's just a matter of the pattern-matching getting better until it really becomes useable in a general-purpose sort of way. But I'm sure that Elektro's makers could have told themselves the same thing. "Look at how sophisticated Elektro is, with 48 electrical relays! Imagine what it will be able to do with hundreds of relays! Or thousands! The possibilities are virtually limitless!"
- perl4ever 8y ago
  Yes, but false hopes keep people going. And these days, your garden variety automaton has billions of tiny switches. There is a well-known saying that "quantity has a quality all its own". My uninformed opinion is that there is probably some crucial learning mechanism missing from current AI but I suspect the raw power of current electronics is completely sufficient. Perhaps we just need a way to couple one module to another to automate training in some way we haven't invented yet but will seem obvious when we do.
- gattilorenz 8y ago
  > consider the degree to which today's neural nets are simply engaging in (admittedly far more complex) pattern-matching, with little more understanding of the actual meaning than Elektro. Yes, that's one of the things I'm referring to...
PhantomGremlin 8y ago
After working in AI research for about 10 years, I suspect that most of what you see (especially outside "scientific" literature) is... very sophisticated smoke and mirrors. My favorite example of this is IBM's Watson. But people just ignore Watson's gaffes because they don't fit the narrative. To refresh people's recollections, here's the Final Jeopardy category of a man/machine contest from a number of years ago: U.S. Cities Clue: Its largest airport is named for a World War II hero; It's second largest, for a World War II battle Watson: What is Toronto????? The correct answer is: What is Chicago? The humans didn't have a problem with it. O'hare airport, Midway airport. It's quite obvious that Watson didn't understand the question at all. https://www.youtube.com/watch?v=C5Xnxjq63Zg https://www.youtube.com/watch?v=C5Xnxjq63Zg

nkoren 8y ago

Yeah, the way that voice commands were recognized via pattern-matching the syllabic rhythms rather than via the actual meaning of the commands was, in hindsight, kind of cheating. But consider the degree to which today's neural nets are simply engaging in (admittedly far more complex) pattern-matching, with little more understanding of the actual meaning than Elektro. A lot of today's AI researchers/enthusiasts/hypesters like to imagine that there isn't really anything fundamental missing from this picture, that it's just a matter of the pattern-matching getting better until it really becomes useable in a general-purpose sort of way. But I'm sure that Elektro's makers could have told themselves the same thing. "Look at how sophisticated Elektro is, with 48 electrical relays! Imagine what it will be able to do with hundreds of relays! Or thousands! The possibilities are virtually limitless!"

perl4ever 8y ago
Yes, but false hopes keep people going. And these days, your garden variety automaton has billions of tiny switches. There is a well-known saying that "quantity has a quality all its own". My uninformed opinion is that there is probably some crucial learning mechanism missing from current AI but I suspect the raw power of current electronics is completely sufficient. Perhaps we just need a way to couple one module to another to automate training in some way we haven't invented yet but will seem obvious when we do.
gattilorenz 8y ago
> consider the degree to which today's neural nets are simply engaging in (admittedly far more complex) pattern-matching, with little more understanding of the actual meaning than Elektro. Yes, that's one of the things I'm referring to...

PhantomGremlin 8y ago

After working in AI research for about 10 years, I suspect that most of what you see (especially outside "scientific" literature) is... very sophisticated smoke and mirrors. My favorite example of this is IBM's Watson. But people just ignore Watson's gaffes because they don't fit the narrative. To refresh people's recollections, here's the Final Jeopardy category of a man/machine contest from a number of years ago: U.S. Cities Clue: Its largest airport is named for a World War II hero; It's second largest, for a World War II battle Watson: What is Toronto????? The correct answer is: What is Chicago? The humans didn't have a problem with it. O'hare airport, Midway airport. It's quite obvious that Watson didn't understand the question at all. https://www.youtube.com/watch?v=C5Xnxjq63Zg https://www.youtube.com/watch?v=C5Xnxjq63Zg