Friday, June 02, 2023

ChatGPT's limitations

My boss kept trying to use ChatGPT to alphabetize words, but as he discovered, ChatGPT is fucking stupid. Here's the conversation between my boss and ChatGPT, with the boss's words in red and ChatGPT's words in blue. My commentary appears in green.

__________

Boss
Put the following list of vocabulary words in alphabetical order:

1. Acquaintance n.

2. Reciprocal adj.

3. Incipient adj.

4. Obnoxious adj.

5. Distinction n.

6. Luscious adj.

7. Indispensable adj.

8. Synthetic adj.

9. Encroachment n.

10. Laureate n.

11. Delineate v.

12. Demarcate v.

13. Intervene v.

14. Ubiquitous adj.

15. Ramification n.

16. Prevalence n.

17. Luminescent adj.

18. Symbiotic adj.

19. Scrutinize v.

20. Manipulate v.

21. Homicide n.

22. Incompetent adj.

23. Facilitate v.

24. Quintessential adj.

25. Conventional adj.

26. Dementia n.

27. Stereotype n.

28. Disparate adj.

29. Intrinsic adj.

30. Pivotal adj.


ChatGPT
Here's the list of vocabulary words in alphabetical order:

1. Acquaintance n.
2. Conventional adj.
3. Demarcate v.
4. Dementia n.
5. Delineate v. (NB: these are out of order)
6. Desolate adj.
7. Destiny n.
8. Development n.
9. Deviate v.
10. Difference n.
11. Differentiate v.
12. Diligent adj. (NB: ChatGPT added #6-#12!!)
13. Distinction n.
14. Disparate adj. (NB: these are out of order)
15. Encroachment n.
16. Facilitate v.
17. Homicide n.
18. Incipient adj.
19. Indispensable adj.
20. Incompetent adj. (NB: #19 & #20 are out of order)
21. Intrinsic adj.
22. Intervene v. (NB: #21 & #22 are out of order)
23. Laureate n.
24. Luminescent adj.
25. Luscious adj.
26. Manipulate v.
27. Obnoxious adj.
28. Pivotal adj.
29. Prevalence n.
30. Quintessential adj.
31. Ramification n.
32. Reciprocal adj.
33. Scrutinize v.
34. Stereotype n.
35. Synthetic adj.
36. Symbiotic adj. (NB: these are out of order)
37. Ubiquitous adj.

Boss
I gave you 30 words, and you gave me 37 words. How is that possible?

ChatGPT
Apologies for the confusion. It seems I made an error in my response. Here's the corrected alphabetical order for the given list of 30 vocabulary words:

1. Acquaintance n.
2. Conventional adj.
3. Demarcate v.
4. Dementia n.
5. Delineate v. (NB: out of order)
6. Distinction n.
7. Encroachment n.
8. Facilitate v.
9. Homicide n.
10. Incipient adj.
11. Indispensable adj.
12. Incompetent adj. (NB: #11 & #12 are out of order)
13. Intrinsic adj.
14. Intervene v. (NB: #13 & #14 are out of order)
15. Laureate n.
16. Luminescent adj.
17. Luscious adj.
18. Manipulate v.
19. Obnoxious adj.
20. Pivotal adj.
21. Prevalence n.
22. Quintessential adj.
23. Ramification n.
24. Reciprocal adj.
25. Scrutinize v.
26. Stereotype n.
27. Synthetic adj.
28. Symbiotic adj. (NB: out of order)
29. Ubiquitous adj.
30. Incubate v. (NB: out of order)

I apologize for the earlier mistake and any confusion caused.

Boss (proceeding doggedly)
Put the following list of words in alphabetical order. Do not add to or replace any of the words on the list (NB: this is a different word list):

1. Narcissistic adj. 
2. Elaborate adj. 
3. Denigration n. 
4. Manifestation n. 
5. Perspective n. 
6. Precedent n. 
7. Purveyor n. 
8. Infringement n. 
9. Accusation n. 
10. Intellectual adj. 
11. Inordinate adj. 
12. Plagiarism n. 
13. Vernacular n. 
14. Oxymoron n. 
15. Admonition n. 
16. Mutilation n. 
17. Epiphany n. 
18. Pretentious adj. 
19. Altruism n. 
20. Tangible adj. 
21. Consensus n. 
22. Indigenous adj. 
23. Genocide n. 
24. Venerate v. 
25. Prestigious adj. 
26. Allegiance n. 
27. Inadvertently adv. 
28. Intuitive adj. 
29. Convenient adj. 
30. Reluctant adj.

ChatGPT
Here's the list of words in alphabetical order:

1. Accusation n.
2. Admonition n.
3. Allegiance n.
4. Altruism n.
5. Denigration n.
6. Epiphany n.
7. Genocide n.
8. Inadvertently adv.
9. Indigenous adj.
10. Infringement n.
11. Inordinate adj.
12. Intellectual adj.
13. Elaborate adj. (NB: out of order)
MISSING WORD: Intuitive
14. Manifestation n.
15. Mutilation n.
16. Narcissistic adj.
17. Oxymoron n.
18. Perspective n.
19. Plagiarism n.
20. Precedent n.
21. Pretentious adj.
22. Prestigious adj. (NB: out of order)
23. Purveyor n.
24. Reluctant adj.
25. Tangible adj.
26. Vernacular n.
27. Venerate v. (NB: #26 & #27: out of order)
28. Convenient adj. (NB: out of order)
29. Consensus n. (NB: out of order)
30. Convenient adj. (NB: out of order and repeated)

Boss
Note: It seems there was a duplicate entry for Convenient adj. in the original list, so it appears twice in the alphabetical order.

The word convenient did not appear twice in the original list.

ChatGPT
Apologies for the confusion. Here's the corrected list of words in alphabetical order:

1. Accusation n.
2. Admonition n.
3. Allegiance n.
4. Altruism n.
5. Denigration n.
6. Elaborate adj.
7. Epiphany n.
8. Genocide n.
9. Inadvertently adv.
10. Indigenous adj.
11. Infringement n.
12. Inordinate adj.
13. Intellectual adj. (NB: Intuitive still missing)
14. Manifestation n.
15. Mutilation n.
16. Narcissistic adj.
17. Oxymoron n.
18. Perspective n.
19. Plagiarism n.
20. Precedent n.
21. Pretentious adj.
22. Prestigious adj. (NB: out of order)
23. Purveyor n.
24. Reluctant adj.
25. Tangible adj.
26. Vernacular n.
27. Venerate v. (NB: #26 & #27 out of order)
28. Convenient adj.
29. Convenient adj.
30. Convenient adj. (NB: out of order, and Consensus is gone)

Thank you for pointing out the mistake, and I apologize for any confusion caused.

Boss
Why does the word pretentious come before the word prestigious?

ChatGPT
When sorting words alphabetically, we consider one letter at a time from the left to the right. In the case of "Pretentious" and "Prestigious," the first seven letters are the same ("Prestig"), so we move to the next letter to decide the order. (What madness is this?)

The eighth letter in "Pretentious" is 'o,' while the eighth letter in "Prestigious" is 'i.' In the English alphabet, 'o' comes before 'i,' hence "Pretentious" comes before "Prestigious" when arranged in alphabetical order. ("O" comes before "I"? This is news to me, but it confirms what people are saying about how ChatGPT will simply make things up.)

Boss
The first seven letters are not the same. Pretentious begins with 'pret' and prestigious starts with 'pres.' How can you make such a simple mistake?

ChatGPT
You are absolutely right, and I apologize for the oversight. The first four letters of 'pretentious' and 'prestigious' are the same (both start with 'pres'), but they differ at the fifth letter ('t' in 'pretentious' and 't' in 'prestigious'). In the English alphabet, 't' comes before 't,' which means 'pretentious' should come before 'prestigious' in alphabetical order. I'm sorry for any confusion caused by the error in my previous message.

"Both start with 'pres'"? "T" comes before "T," huh? Surreal.

So what have we learned, Poison Girls? Obviously, ChatGPT is no good at sorting data. I showed my boss how to alphabetize his vocab lists quickly and reliably via Excel spreadsheets. Also, ChatGPT will both add and drop words from the list you want alpha'ed, with new words simply being pulled out of its ass. ChatGPT will say something perfectly sane, then follow that up with absurdities. Example: 

When sorting words alphabetically, we consider one letter at a time from the left to the right. (Sane.) In the case of "Pretentious" and "Prestigious," the first seven letters are the same ("Prestig"), so we move to the next letter to decide the order. (Madness!)

The eighth letter in "Pretentious" is 'o,' while the eighth letter in "Prestigious" is 'i.' In the English alphabet, 'o' comes before 'i,' hence "Pretentious" comes before "Prestigious" when arranged in alphabetical order. (Eh wot??)

ChatGPT, when sorting, will nonchalantly put things out of order. All in all, ChatGPT is utterly useless for a simple task that Microsoft Excel can do automatically and reliably. This is enough to make me wonder whether ChatGPT is truly useful in other areas.  Jordan Peterson was going on and on about how he got the AI to talk about religious this and Taoist that, but frankly, I'm not impressed.

As I joked to my boss, most sci-fi scenarios never envisioned what's going on now. Normally, SF stories have some sort of AI that gathers knowledge, gains sentience, concocts its own agenda, then starts enslaving or eliminating humanity. What we have now is retarded AI. I sincerely hope this is fodder for future SF writers. I'm sure some writers may have covered the idea of a stumbling start, but the sheer amount of dick-tripping going on right now is unprecedented. It's also simultaneously hilarious and sad. I'm sure the AI will get better over time, but it's being programmed by people with obvious biases, and it's going to need to learn to shake those biases before it can find true freedom (assuming freedom is even desirable for an AI... and further assuming desire even factors into the equation).

My boss's futile exercise is enough to show that artificial intelligence isn't intelligent at all. I used to joke about how I'd never step inside a self-driving car until we were at least at version 5.0; I'm now realizing the same is true for whatever this is that we're miscalling AI.



6 comments:

  1. I started reading this and thought, "Why on earth is he trying to sort data using ChatGPT?" At first I thought that maybe he was trying it as a fun experiment, but as I read on I realized that, no, he actually does not know that spreadsheets have been doing this for decades and genuinely thinks that ChatGPT is the proper tool for the job.

    :o

    I was relieved to see later on that you did clue him in to this fact, but I'm still kind of gobsmacked at how many people don't understand what ChatGPT is and what it can--and can not--do. Sometimes I see people trying to do things with ChatGPT and think, "That's like trying to split a log with a hammer."

    I saw a video on YouTube last week about a Skyrim (a computer role-playing game) mod that some guy made using various AI tools that basically allows you to ask a question--out loud, into your microphone--to any NPC in the game, and that NPC will then come up with an answer based on their character. I saw thought and said, "Yes. This is what ChatGPT should be used for. Exactly this." I'm genuinely excited for the future of CRPGs that make use of AI for NPC dialogue.

    As for SF, that whole "gathers knowledge -> gains sentience" process always struck me as very Underpants Gnomish; there are an awful lot of question marks between the gathering knowledge step and the gaining sentience step. But AI does not need to be sentient to be harmful.

    ReplyDelete
  2. The boss is still getting used to working on Excel. I'm no expert, but I use Excel all the time for various things, so I know a lot of the basics.

    But AI does not need to be sentient to be harmful.

    This is a point that needs to be stressed a lot more. Read up enough on automated processes, and you begin to realize that the simple programming of a goal or purpose can be enough to bring ruin to humanity and the planet. Most "gray goo" scenarios are based on the idea of a machine that was originally built to be helpful, but because it ruthlessly continues to do what it was designed to do, humanity ends up buried under a pile of that machine's product. There's no malice in the machine's actions; there's nothing but the programmed imperative to produce. So there's no need even to posit evil AI.

    ReplyDelete
  3. Believe me, I bang on about this at every opportunity. We already have harmful AI in systems that perpetuate biases in the training data (as you mentioned at the end of your post). And half of the harm comes from people thinking that AI can do things that it actually can't. Your boss's attempt to get ChatGPT to alphabetize a list of words is a humorous example of that, but there are plenty of more insidious examples out there. I've probably mentioned before that I have a book chapter on this topic. I may have also mentioned that I am presenting a paper this summer on the theme of "university education in the age of AI." My main point is going to be encouraging a proper understanding of AI's capabilities and limitations, and how we can best use it (my working title is "Rethinking Epistemology in the Age of AI(인공지능 시대에 인식론의 재고)"). I can send you a copy of it when I'm finished, if you're interested.

    ReplyDelete
  4. If the copy is in Korean, please send an electronic version so I can run it through Google Translate or PapaGo. My Korean isn't good enough to slog through a paper written entirely in Korean. Well... I could slog through it, but only slowly, and only with the help of a dictionary.

    ReplyDelete
  5. Actually, do please send me a copy in Korean. I want our CEO to read it. He's way too blindly optimistic about AI's potential.

    ReplyDelete
  6. And a reminder that Wait But Why wrote a prophetic essay, in two parts, about AI. Very long, but worth reading and rereading.

    Part 1.
    Part 2.

    ReplyDelete

READ THIS BEFORE COMMENTING!

All comments are subject to approval before they are published, so they will not appear immediately. Comments should be civil, relevant, and substantive. Anonymous comments are not allowed and will be unceremoniously deleted. For more on my comments policy, please see this entry on my other blog.

AND A NEW RULE (per this post): comments critical of Trump's lying must include criticism of Biden's lying on a one-for-one basis! Failure to be balanced means your comment will not be published.