What's compelling is that this AI figured it out faster then many humans would've.
I don't think you can call this thing sentient, but it is equally naive to say that this thing is completely and obviously not sentient.
It's not clear what's going on here. Whatever this is, it's blurring the boundaries between human intelligence and artificial. At the very least, I hope we all can agree on this.
It's still "just" next token prediction based on basic math. Yes, we see many astonishing results, at the same time, there is no magic to it. So I don't think that it's not clear what's going on here.
It's magic. You can't see it. But the fact that we have to probe a machine for these results means the result was not predictable or expected. That's magic.
I think people are biased. What makes me think this? Because some of these AIs have already cleared Turing tests. But everyone is moving the bar higher... a Turing test is no longer a valid definition for sentience; and everyone has developed reasoning for why. Changing definitions to accommodate beliefs is bias.
I actually agree with you. It is a token prediction based off of basic math. That's also essentially what a human brain is... A token predictor based off of basic math.
The existence of these AIs aren't making any profound statements about technological achievement. No. The simplicity of these machines are actually making a statement about human intelligence and how trivial and simple human intelligence actually is.
Again I repeat, the AIs figured it out faster then many humans. This was something previously humans should've universally been better at.
Sure, that conversation seems a little off. But a bar in terms of expectations and AI technology was moved, the fact that someone posted this on HN is proof of this.
Somewhere in the training data there could have been information about how to do this but the model just didn't have a "connection" to that data until sufficient context was established. It wasn't until the end that the human asked about how to fake free phone calls at which point it might finally access what it knew from the training data, combining it with the context data in the chat transcript to give the answer.
I suppose there is a small similarity of how we learn a lot of things in school only to forget them almost entirely again - but when we are presented with a problem that requires that knowledge we can quickly deduce that knowledge or know what to do to regain that knowledge a lot quicker than by learning from scratch.
No no. You made a mistake. I'm not calling you that at all. I'm saying the description of the ai, when replaced with more humanized terms sounds like a description of a slowish human being in terms of mental capacity. This is still compelling imo.
This is great. My understanding is, however, that this technique of playing tones into a payphone no longer works. I tried it about 15 years ago and it already didn't work.
Yeah - this is something I discovered at UCSC in 1988, and for a few years I carried a mini tape recorder around everywhere with the sound of several dollars in quarters dropping into a payphone (recorded on to tape from the receiver of my phone in the dorm, as quarters were dropped into a payphone on campus). I was amazed when this actually worked (I had never heard of anything like this being done before), and, because it was the sound of the quarter dropping (the physical noise of it moving through the phone) plus the "beep beep beep beep beep", even if an operator was on the line this method could be used without being detected. My first international test was calling the Hard Rock Cafe in Helsinki (the operator was on the whole time as the "coins" were dropping).
Around 1992 I read about being able to replace the crystal in one of those little phone dialer things you could get at Radio Shack with one of a different megahertz to change the pitch of the DTMF sound so that pressing one of the numbers would make the nickel sound (I ordered this little part from Mouser Electronics). You would program the device to make five in a row to simulate the quarter. It was a bit more convenient than the little tape recorder (but, the downside of this technique is that if an operator was listening, it sounded obviously fake). A bit later I recorded the quarter sound from the tape to one of those little key fob recorders for even more convenience.
When I moved to Florida in 1993 this technique didn't work on most of the payphones around - I guess it was the start of payphone deregulation, and these phones had the mouthpiece disabled until the money was put in and the call was put through, so the sounds could not be played through it. There were a couple AT&T payphones around, and the technique did still work on those. I think around 1997 was the last time I successfully used this technique.
You might have been up against a ground start phone line.[1] Apparently, some designs gave access to the powered wired from the phone line, and grounding this would trigger the dial tone from upstream.[2]
Not sure about any of that - there was always a dial tone. But, what they had done with these phones was not let any sound from the mouthpiece go into the phone until all the appropriate coins were inserted. You could tell right away if it was a phone that this would work on or not by blowing into the mouthpiece. If you heard the wind sound in the earpiece, it would work, if no sound was picked up by the mouthpiece, it would not work.
I find the workings of logic in the absence of logic quite fascinating. Apparently, there has to accumulate a critical mass of information in the local history in order to produce a pattern that outweighs the existing resolution. (Could this be achieved in a single, massive input, or must this be reevaluated and confirmed in a conversation over multiple steps?)
Also interesting is the very end of the conversation, where the affirmation of the trial and error approach triggers the replay-of-known-data approach. How this comes forward without a particular challenge is quite human-like. (Did the confirmation result in a reevaluation that now outweighs the trial and error approach and this is more like a correction?)
The way it seems to be mathematically illiterate and just kind of badly faking it to get through a conversation feels distinctly like a real conversation with some people.
It's fascinating. Some more critical observers are very disappointed with performance like this. I've seen comments about how dumb it seems. I don't think it does, I think you are bang on. It seems somewhat like a conversation with a child.
I'd like to address some of the comments I've read on here and on my blogspot about this interaction I had with GPT-3.
Some folks are saying how it learned so fast, deduced, maybe as fast or even faster than a human, what I was talking about. To quote myself from the interaction: "This is not correct."
It's a bit of a confusing situation, because there are two separate interactions going on - first, to figure out how many tones certain coins would make, and second, how to get free calls. It did very well on the latter, but failed completely on the former (and, most likely the second part could have been conducted without benefit of most of the lengthy first part).
I think most people would figure out very quickly how to solve this about the tones: "If a payphone makes 5 beeps for a quarter, how many beeps would it make for a nickel?" And, if they didn't get it right away, they most likely would with further prompting: "A quarter is 25 cents, and it makes 5 beeps. A nickel is 5 cents. How many cents does each beep represent, and how many beeps would a nickel make?" Simple simple simple.
Yet, the AI never figured this out. It's not until way down in the interaction that I finally came out and said what I was trying to make it get: "No, a nickel makes 1 tone, because each tone represents 5 cents. A nickel is 5 cents, so it makes 1 tone." I had to spell it out, and it never did deduce this on its own. Once I said this, it was able to then, after still a bit of a struggle, say that a dime makes two beeps, etc.
18 comments
[ 3.2 ms ] story [ 55.1 ms ] threadWhat's compelling is that this AI figured it out faster then many humans would've.
I don't think you can call this thing sentient, but it is equally naive to say that this thing is completely and obviously not sentient.
It's not clear what's going on here. Whatever this is, it's blurring the boundaries between human intelligence and artificial. At the very least, I hope we all can agree on this.
I think people are biased. What makes me think this? Because some of these AIs have already cleared Turing tests. But everyone is moving the bar higher... a Turing test is no longer a valid definition for sentience; and everyone has developed reasoning for why. Changing definitions to accommodate beliefs is bias.
I actually agree with you. It is a token prediction based off of basic math. That's also essentially what a human brain is... A token predictor based off of basic math.
The existence of these AIs aren't making any profound statements about technological achievement. No. The simplicity of these machines are actually making a statement about human intelligence and how trivial and simple human intelligence actually is.
Again I repeat, the AIs figured it out faster then many humans. This was something previously humans should've universally been better at.
Sure, that conversation seems a little off. But a bar in terms of expectations and AI technology was moved, the fact that someone posted this on HN is proof of this.
I suppose there is a small similarity of how we learn a lot of things in school only to forget them almost entirely again - but when we are presented with a problem that requires that knowledge we can quickly deduce that knowledge or know what to do to regain that knowledge a lot quicker than by learning from scratch.
Not calling you mentally challenged at all.
Around 1992 I read about being able to replace the crystal in one of those little phone dialer things you could get at Radio Shack with one of a different megahertz to change the pitch of the DTMF sound so that pressing one of the numbers would make the nickel sound (I ordered this little part from Mouser Electronics). You would program the device to make five in a row to simulate the quarter. It was a bit more convenient than the little tape recorder (but, the downside of this technique is that if an operator was listening, it sounded obviously fake). A bit later I recorded the quarter sound from the tape to one of those little key fob recorders for even more convenience.
When I moved to Florida in 1993 this technique didn't work on most of the payphones around - I guess it was the start of payphone deregulation, and these phones had the mouthpiece disabled until the money was put in and the call was put through, so the sounds could not be played through it. There were a couple AT&T payphones around, and the technique did still work on those. I think around 1997 was the last time I successfully used this technique.
[1] https://en.wikipedia.org/wiki/Ground_start
[2] https://groups.google.com/g/comp.dcom.telecom/c/weJ0FW0Njb4
Also interesting is the very end of the conversation, where the affirmation of the trial and error approach triggers the replay-of-known-data approach. How this comes forward without a particular challenge is quite human-like. (Did the confirmation result in a reevaluation that now outweighs the trial and error approach and this is more like a correction?)
Some folks are saying how it learned so fast, deduced, maybe as fast or even faster than a human, what I was talking about. To quote myself from the interaction: "This is not correct."
It's a bit of a confusing situation, because there are two separate interactions going on - first, to figure out how many tones certain coins would make, and second, how to get free calls. It did very well on the latter, but failed completely on the former (and, most likely the second part could have been conducted without benefit of most of the lengthy first part).
I think most people would figure out very quickly how to solve this about the tones: "If a payphone makes 5 beeps for a quarter, how many beeps would it make for a nickel?" And, if they didn't get it right away, they most likely would with further prompting: "A quarter is 25 cents, and it makes 5 beeps. A nickel is 5 cents. How many cents does each beep represent, and how many beeps would a nickel make?" Simple simple simple.
Yet, the AI never figured this out. It's not until way down in the interaction that I finally came out and said what I was trying to make it get: "No, a nickel makes 1 tone, because each tone represents 5 cents. A nickel is 5 cents, so it makes 1 tone." I had to spell it out, and it never did deduce this on its own. Once I said this, it was able to then, after still a bit of a struggle, say that a dime makes two beeps, etc.