4 ms·

I’m just trying to “prove” that it isn’t just randomly statistically choosing the next logical word. It has to know context and have some level of “understandin

sramsay 4y ago
This. We have had a lot people -- including journalists and academics with big microphones -- learn for the first time what a Markov chain is, and then conclude that ChatGPT is a "just Markov chains" (or whatever similarly reductive concept). They really, really don't know what they're talking about it, and yet it's becoming a kind of truth through repetition. Pretty soon, the bots will start saying it!
- _2uwr 4y ago
  > They really, really don't know what they're talking about it, and yet it's becoming a kind of truth through repetition. Kind of like religion or that people working for the state are more trustworthy than people taking drugs or sleeping on the street or under the age of 18. >Pretty soon, the bots will start saying it! We are chemical based repetition machines, psychologists see this with kids using bobo dolls exposed to new ideas on tv or in books repeating learned behaviour on bobo dolls. I think some of the chemicals we make like https://en.wikipedia.org/wiki/N,N-Dimethyltryptamine https://en.wikipedia.org/wiki/N,N-Dimethyltryptamine actually help to create new idea's, as many people say they come up with solutions after some sleep. There appears to be a sub culture in silicon valley were microdosing lsd helps to maintain the creativity with coding. It would seem logical for the bots to start saying it. If the bots start amplifying flawed knowledge like a lot of Reddit content or Facebook content, the internet will need to deal with the corruption of the internet, like using Wikipedia as a source of reference. https://en.wikipedia.org/wiki/Wikipedia:List_of_hoaxes_on_Wikipedia/Less_than_one_year https://en.wikipedia.org/wiki/Wikipedia:List_of_hoaxes_on_Wi... https://en.wikipedia.org/wiki/Reliability_of_Wikipedia https://en.wikipedia.org/wiki/Reliability_of_Wikipedia https://en.wikipedia.org/wiki/List_of_Wikipedia_controversies https://en.wikipedia.org/wiki/List_of_Wikipedia_controversie... In practice, sources of reference will need cleaning up and the question about who should be responsible for that will probably ensue. It will call into question such entities like encyclopaedia companies who have been the gatekeeper of the knowledge they sell, in much the same way the Guinness World Record are the self declared arbitrators of world records. We already see Judges working in a semi cohesive manner, but the appeals process tells you its flawed and inefficient, ergo can the judiciary even be trusted to interpret law properly? Its also interesting how the financial system also reinforces this mob rule mentality and not always for the better, but then some people will no doubt suggest we should be grateful we have bungled our way to this point in history.
VectorLock 4y ago
If you don't think its just statistically choosing the next word, what do you think its doing? Do you think it understands the mathematical concepts you're questioning it about?
- scarface74 4y ago
  This is a better example I posted https://news.ycombinator.com/item?id=34811353 https://news.ycombinator.com/item?id=34811353
  - VectorLock 4y ago
    I haven't tried Bing but ask ChatGPT "what is 12348120348102341 + 10823401828482?" Bonus points: If you ask Google Search this, it knows the right answer.
    - scarface74 4y ago
      Now try this with Google: “Write an AppleSoft Basic program to do the computation” (Here it gives me a simplistic Basic program that wouldn’t work because the number would be too large) Me: “AppleSoft Basic can’t handle numbers that long” ChatGPT: You're correct that the integer arithmetic in AppleSoft BASIC has some limitations and cannot handle numbers as large as the sum of the two numbers you provided. AppleSoft BASIC is a 16-bit programming language, which means that it can handle integer values from -32,768 to 32,767. If you want to work with very large numbers in AppleSoft BASIC, you can use a string to store the numbers and write your own code to perform arithmetic operations on the strings. Here's an example program that demonstrates this approach: 10 A$ = "12348120348102341" 20 B$ = "10823401828482" 30 L = LEN(A$) 40 M = LEN(B$) 50 IF L < M THEN A$ = LEFT$("0", M - L) + A$ : L = M 60 IF M < L THEN B$ = LEFT$("0", L - M) + B$ : M = L 70 C$ = "" 80 R = 0 90 FOR I = L TO 1 STEP -1 100 D = ASC(MID$(A$, I, 1)) + ASC(MID$(B$, I, 1)) - 96 + R 110 C$ = CHR$(D MOD 10 + 48) + C$ 120 R = INT(D / 10) 130 NEXT I 140 IF R > 0 THEN C$ = CHR$(R + 48) + C$ 150 PRINT C$
      - VectorLock 4y ago
joe_the_user 4y ago
I’m just trying to “prove” that it isn’t just randomly statistically choosing the next logical word. It has to know context and have some level of “understanding” of other contexts. But you ... aren't. The statistically most likely words coming after problem X may well be solution X. Because it's following the pattern of humans using rules. And context is also part of a prediction. The only this is different from something just using rules is that it will also put in other random things from it's training - but only at the rate they occur, which for some things can be quite low. But only some things.
- scarface74 4y ago
  > it still doesn't show it's using the consistent rules that a person would expect. How is this different from humans? If you give me the same coding assignment on different days, I’m not going to write my code the exact same way or even structure it the same way. But I did once see a post on HN where someone ran an analysis on all HN posters and it was able to tell that I posted under two different names based on my writing style. Not that I was trying to hide anything. My other username is scarface_74 as opposed to Scarface74.
- dclowd9901 4y ago
  Don’t we have a problem then? By nature of effective communication, AI could never prove to you it understands something, since any sufficient understanding of a topic would be met with an answer that could be hand-waved as “Well that’s the most statistically likely answer.” Newsflash: this basically overlaps 100% with any human’s most effective answer. I think I’m beginning to understand the problem here. The folks here who keep poo-pooing these interactions don’t just see the AIs as unconscious robots. I think they see everyone that way.
  - danaris 4y ago
    No; what we need, in order to be willing to believe that understanding is happening, is to know that the underlying structures fundamentally allow that. ChatGPT's underlying structures do not. What it does, effectively, is look at the totality of the conversation thus far, and use the characters and words in it, combined with its training data, to predict, purely statistically, what characters would constitute an appropriate response. I know that some people like to argue that what humans do cannot be meaningfully distinguished from this, but I reject this notion utterly. I know that my own thought processes do not resemble this procedure, and I believe that other people's are similar.
lelanthran 4y ago
> I’m just trying to “prove” that it isn’t just randomly statistically choosing the next logical word. It has to know context and have some level of “understanding” of other contexts. FCOL, you can't use "complex output" as proof that the process has any intelligence directing it. If you could, we would take the Intelligent Design argument seriously. We don't. We never did. We need a good clear argument to convince us now why it is a good idea to accept Intelligent Design as an argument.
- naasking 4y ago
  It's not just complex output, it's output that's relevant to the prompt including considerable nuance. If that's not bordering on intelligence, then you shouldn't consider humans intelligent either.
  - lelanthran 4y ago
    > it's output that's relevant to the prompt including considerable nuance. You can say the same thing about Intelligent Design, and yet we dismiss it anyway.
    - naasking 4y ago
      We didn't dismiss intelligent design, we replaced it with a more parsimonious theory that better explained the evidence. Big difference.

sramsay 4y ago

This. We have had a lot people -- including journalists and academics with big microphones -- learn for the first time what a Markov chain is, and then conclude that ChatGPT is a "just Markov chains" (or whatever similarly reductive concept). They really, really don't know what they're talking about it, and yet it's becoming a kind of truth through repetition. Pretty soon, the bots will start saying it!

_2uwr 4y ago
> They really, really don't know what they're talking about it, and yet it's becoming a kind of truth through repetition. Kind of like religion or that people working for the state are more trustworthy than people taking drugs or sleeping on the street or under the age of 18. >Pretty soon, the bots will start saying it! We are chemical based repetition machines, psychologists see this with kids using bobo dolls exposed to new ideas on tv or in books repeating learned behaviour on bobo dolls. I think some of the chemicals we make like https://en.wikipedia.org/wiki/N,N-Dimethyltryptamine https://en.wikipedia.org/wiki/N,N-Dimethyltryptamine actually help to create new idea's, as many people say they come up with solutions after some sleep. There appears to be a sub culture in silicon valley were microdosing lsd helps to maintain the creativity with coding. It would seem logical for the bots to start saying it. If the bots start amplifying flawed knowledge like a lot of Reddit content or Facebook content, the internet will need to deal with the corruption of the internet, like using Wikipedia as a source of reference. https://en.wikipedia.org/wiki/Wikipedia:List_of_hoaxes_on_Wikipedia/Less_than_one_year https://en.wikipedia.org/wiki/Wikipedia:List_of_hoaxes_on_Wi... https://en.wikipedia.org/wiki/Reliability_of_Wikipedia https://en.wikipedia.org/wiki/Reliability_of_Wikipedia https://en.wikipedia.org/wiki/List_of_Wikipedia_controversies https://en.wikipedia.org/wiki/List_of_Wikipedia_controversie... In practice, sources of reference will need cleaning up and the question about who should be responsible for that will probably ensue. It will call into question such entities like encyclopaedia companies who have been the gatekeeper of the knowledge they sell, in much the same way the Guinness World Record are the self declared arbitrators of world records. We already see Judges working in a semi cohesive manner, but the appeals process tells you its flawed and inefficient, ergo can the judiciary even be trusted to interpret law properly? Its also interesting how the financial system also reinforces this mob rule mentality and not always for the better, but then some people will no doubt suggest we should be grateful we have bungled our way to this point in history.

VectorLock 4y ago

If you don't think its just statistically choosing the next word, what do you think its doing? Do you think it understands the mathematical concepts you're questioning it about?

scarface74 4y ago
This is a better example I posted https://news.ycombinator.com/item?id=34811353 https://news.ycombinator.com/item?id=34811353
- VectorLock 4y ago
  I haven't tried Bing but ask ChatGPT "what is 12348120348102341 + 10823401828482?" Bonus points: If you ask Google Search this, it knows the right answer.
  - scarface74 4y ago
    Now try this with Google: “Write an AppleSoft Basic program to do the computation” (Here it gives me a simplistic Basic program that wouldn’t work because the number would be too large) Me: “AppleSoft Basic can’t handle numbers that long” ChatGPT: You're correct that the integer arithmetic in AppleSoft BASIC has some limitations and cannot handle numbers as large as the sum of the two numbers you provided. AppleSoft BASIC is a 16-bit programming language, which means that it can handle integer values from -32,768 to 32,767. If you want to work with very large numbers in AppleSoft BASIC, you can use a string to store the numbers and write your own code to perform arithmetic operations on the strings. Here's an example program that demonstrates this approach: 10 A$ = "12348120348102341" 20 B$ = "10823401828482" 30 L = LEN(A$) 40 M = LEN(B$) 50 IF L < M THEN A$ = LEFT$("0", M - L) + A$ : L = M 60 IF M < L THEN B$ = LEFT$("0", L - M) + B$ : M = L 70 C$ = "" 80 R = 0 90 FOR I = L TO 1 STEP -1 100 D = ASC(MID$(A$, I, 1)) + ASC(MID$(B$, I, 1)) - 96 + R 110 C$ = CHR$(D MOD 10 + 48) + C$ 120 R = INT(D / 10) 130 NEXT I 140 IF R > 0 THEN C$ = CHR$(R + 48) + C$ 150 PRINT C$
    - VectorLock 4y ago

joe_the_user 4y ago

I’m just trying to “prove” that it isn’t just randomly statistically choosing the next logical word. It has to know context and have some level of “understanding” of other contexts. But you ... aren't. The statistically most likely words coming after problem X may well be solution X. Because it's following the pattern of humans using rules. And context is also part of a prediction. The only this is different from something just using rules is that it will also put in other random things from it's training - but only at the rate they occur, which for some things can be quite low. But only some things.

scarface74 4y ago
> it still doesn't show it's using the consistent rules that a person would expect. How is this different from humans? If you give me the same coding assignment on different days, I’m not going to write my code the exact same way or even structure it the same way. But I did once see a post on HN where someone ran an analysis on all HN posters and it was able to tell that I posted under two different names based on my writing style. Not that I was trying to hide anything. My other username is scarface_74 as opposed to Scarface74.
dclowd9901 4y ago
Don’t we have a problem then? By nature of effective communication, AI could never prove to you it understands something, since any sufficient understanding of a topic would be met with an answer that could be hand-waved as “Well that’s the most statistically likely answer.” Newsflash: this basically overlaps 100% with any human’s most effective answer. I think I’m beginning to understand the problem here. The folks here who keep poo-pooing these interactions don’t just see the AIs as unconscious robots. I think they see everyone that way.
- danaris 4y ago
  No; what we need, in order to be willing to believe that understanding is happening, is to know that the underlying structures fundamentally allow that. ChatGPT's underlying structures do not. What it does, effectively, is look at the totality of the conversation thus far, and use the characters and words in it, combined with its training data, to predict, purely statistically, what characters would constitute an appropriate response. I know that some people like to argue that what humans do cannot be meaningfully distinguished from this, but I reject this notion utterly. I know that my own thought processes do not resemble this procedure, and I believe that other people's are similar.

lelanthran 4y ago

> I’m just trying to “prove” that it isn’t just randomly statistically choosing the next logical word. It has to know context and have some level of “understanding” of other contexts. FCOL, you can't use "complex output" as proof that the process has any intelligence directing it. If you could, we would take the Intelligent Design argument seriously. We don't. We never did. We need a good clear argument to convince us now why it is a good idea to accept Intelligent Design as an argument.

naasking 4y ago
It's not just complex output, it's output that's relevant to the prompt including considerable nuance. If that's not bordering on intelligence, then you shouldn't consider humans intelligent either.
- lelanthran 4y ago
  > it's output that's relevant to the prompt including considerable nuance. You can say the same thing about Intelligent Design, and yet we dismiss it anyway.
  - naasking 4y ago
    We didn't dismiss intelligent design, we replaced it with a more parsimonious theory that better explained the evidence. Big difference.