1y ago

The supposed "ethical" limitations are getting out of hand

I was using Bing to create a list of countries to visit. Since I have been to the majority of the African nation on that list, I asked it to remove the african countries...

It simply replied that it can't do that due to how unethical it is to descriminate against people and yada yada yada. I explained my resoning, it apologized, and came back with the same exact list.

I asked it to check the list as it didn't remove the african countries, and the bot simply decided to end the conversation. No matter how many times I tried it would always experience a hiccup because of some ethical process in the bg messing up its answers.

It's really frustrating, I dunno if you guys feel the same. I really feel the bots became waaaay too tip-toey

You're viewing a single thread.

210 comments

The very important thing to remember about these generative AI is that they are incredibly stupid.
They don't know what they've already said, they don't know what they're going to say by the end of a paragraph.
All they know is their training data and the query you submitted last. If you try to "train" one of these generative AI, you will fail. They are pretrained, it's the P in chatGPT. The second you close the browser window, the AI throws out everything you talked about.
Also, since they're Generative AI, they make shit up left and right. Ask for a list of countries that don't need a visa to travel to, and it might start listing countries, then halfway through the list it might add countries that do require a visa, because in its training data it often saw those countries listed together.
AI like this is a fun toy, but that's all it's good for.
- Are you saying I shouldn't use chat GPT for my life as a lawyer? 🤔
  
  I mean, you could...
  
  It can be useful for top-level queries that deal with well-settled law, as a tool to point you in the right direction with your research.
  For example, once, I couldn't recall all the various sentencing factors in my state. ChatGPT was able to spit out a list to refresh my memory, which gave me the right phrases to search on Lexis.
  But, when I asked GPT to give me cases, it gave me a list of completely made up bullshit.
  So, to get you started, it can be useful. But for the bulk of the research? Absolutely not.
  
  I disagree. It's a large language model so all it can do is say things that sound like what someone might say. It's trained on public content, including people giving wrong answers or refusing to answer.
  
  Think of all the cases you can find s to establish precedent!
- Not quite true. They have earlier messages available.
- Bings version of chatgpt once said Vegito was the result of Goku and Vegeta performing the Fusion dance. That's when I knew it wasn't perfect. I tried to correct it and it said it didn't want to talk about it anymore. Talk about a diva.
  Also one time, I asked it to generate a reddit AITA story where they were obviously the asshole. It started typing out "AITA for telling my sister to stop being a drama queen after her miscarriage..." before it stopped midway and, again, said it didn't want to continue this conversation any longer.
  Very cool tech, but it's definitely not the end all, be all.
  
  That’s actually fucking hilarious.
  “Oh I’d probably use the meat grinder … uh I don’t walk to talk about this any more”
  
  Bing chat seemingly has a hard filter on top that terminates the conversation if it gets too unsavory by their standards, to try and stop you from derailing it.
  
  I was asking it (binggpt) to generate “short film scripts” for very weird situations (like a transformer that was sad because his transformed form was a 2007 Hyundai Tuscon) and it would write out the whole script, then delete it before i could read it and say that it couldn’t fulfil my request.
  
  It knew it struck gold and actually sent the script to Michael Bay
- They know everything they've said since the start of that session, even if it was several days ago. They can correct their responses based on your input. But they won't provide any potentially offensive information, even in the form of a joke, and will instead lecture you on DEI principles.
- I seriously underestimated how little people understand these programs, and how much they overestimate them. Personally I stay away from them for a variety of reasons, but the idea of using them like OP does or various other ways I've heard about is absurd. They're not magic problem solvers - they literally only make coherent blocks of text. Yes, they're quite good at that now, but that doesn't mean they're good at literally anything else.
  I know people smarter than me see potential and I'm curious to see how it develops further, but that all seems like quite a ways off, and the way people treat and use them right now is just creepy and weird.
  
  I’ve found it useful for generating ideas for various things, especially ways to code something. But I never use its code. It’s often riddled with errors but it can give me a better idea of which path I should take.
  
  I've found it useful for generating regexes and document outlines.
  
  regex101.com has been my jam for years. Just in case you don’t know about it. Great site for building and testing.
  
  I get that. As a scattered, "throw things at the wall" tactic, it serves well enough. It is far from the all-in-one answer people seem to think it is, though. It can be a good first pass, but like you said, more often than not its output is riddled with errors and needing lots of refinement.
  
  I use it similarly to clean up OCRed text. I can hand it something full of 70% gobbledyremoved and it hands me back something that makes sense and is 95% right. I manually verify it, fix the goofs, and it's so much faster.
  So, riddled with errors but a decent start.
  
  Wouldn't you get a better result from Grammerly for that?
  
  I use it for more "OCR Nightmare" than grammar nightmare. Here's a (trimmed) example:
  #Please clean up the following text:
  Ile took an active part i•all the proceedings of the day. All went expecting to have a grand, good time, and none were disap- :~ _ii!c_,rcwT:a_i_c 0
  undefined
  
  ~~~;u~~n~~1 ~e: 'J farewell set, and Mr. Bailey 1·eturncd the ~ :d~~~~ls!i~s/;;r;rie\~ 11 i~:n~~Ut:sa~~i°::: [: tended to them a cordial invitation to meet ,o him at that place on the 25th of August, ['. 1877. And echo said "So m]fii! b8;~n . ..
  #ChatGPT
  Here's the cleaned-up version of the text:
  He took an active part in all the proceedings of the day. All went expecting to have a grand, good time, and none were disappointed. The merry company engaged in a farewell set, and Mr. Bailey returned the warmth of their farewell. He extended to them a cordial invitation to meet him at that place on the 25th of August, And echo said, "So many have been."
  That was from a scanned 1800s newspaper. Adobe Acrobat did not do a good job converting it to words. ChatGPT cleaned it up nicely and kept about the same meaning. There were some swings and misses, but it's much easier to edit the result into accuracy than the original mess.
  #Actual text as hand-transcribed:
  He took an active part in all the proceedings of the day. All went expecting to have a grand, good time, and none were disappointed. The mirth continued till the night drew near, when they danced the farewell set, and Mr. Bailey returned the company his hearty thanks for this renewed expression of their friendship, and extended to them a cordial invitation to meet him at that place on the 25th of August, 1877. And echo said ``So mote it be.''
  
  they literally only make coherent blocks of text. Yes, they're quite good at that now, but that doesn't mean they're good at literally anything else.
  No, they don't make coherent blocks of text. If they were actually good at writing, they'd be good at everything, because writing is just encoding thoughts on paper, and to master writing is to master thought
  
  Hence why I didn't say writing. I said "blocks of text".
  
  Perfect ability to produce coherent blocks of text would also require mastery of all disciplines
  
  Since I I'm explicitly arguing these programs aren't perfect, even at generating blocks of text, I don't really understand why you are insisting on arguing semantics here and don't really have any interest in continuing...whatever this is. Have a good one.
- They don't know what they've already said, they don't know what they're going to say by the end of a paragraph.
  I mean, the first part of this is just wrong (the next prompt usually includes everything that has been said so far}, and the second part is also not completely true. When generating, yes, they're only ever predicting the next token, and start again after that. But internally, they might still generate a full conceptual representation of what the full next sentence or more is going to be, even if the generated output is just the first token of that. You might say that doesn't matter because for the next token, that prediction runs again from scratch and might change, but remember that you're feeding it all the same input as before again, plus one more token which nudges it even further towards the previous prediction, so it's very likely it's gonna arrive at the same conclusion again.
  
  Do you mean that the model itself has no memory, but the chat feature adds memory by feeding the whole conversation back in with each user submission?
  
  Yeah, that's how these models work. They have also have a context limit, and if the conversation goes too long they start "forgetting" things and making more mistakes (because not all of the conversation can be fed back in).
  
  Is that context limit a hard limit or is it a sort of gradual decline of “weight” from the characters further back until they’re no longer affecting output at the head?
  
  Nobody really knows because it's an OpenAI trade secret (they're not very "open"). Normally, it's a hard limit for LLMs, but many believe OpenAI are using some tricks to increase the effective context limit. I.e. some people believe instead of feeding back the whole conversation, they have GPT create create a shorter summaries of parts of the conversation, then feed the summaries back in.
  
  I think it’s probably something that could be answered with general knowledge of LLM architecture.
  Yeah OpenAI’s name is now a dirty joke. They decided before their founding that the best way to make AI play nice was to have many many many AIs in the world, so that the AIs would have to be respectful to one another, and overall adopt pro social habits because those would be the only ones tolerated by the other AIs.
  And the way to ensure a community of AIs, a multipolar power structure, was to disseminate AI tech far and wide as quickly as possible, instead of letting it develop in one set of hands.
  Then they said fuck that we want that power, and broke their promise.
- @ChatGPT@lemmings.world
  testing
  
  @ChatGPT@lemmings.world
  Give a list of all African countries.
  
  Here is a list of all African countries:
  Algeria
  Angola
  Benin
  Botswana
  Burkina Faso
  Burundi
  Cape Verde
  Cameroon
  Central African Republic
  Chad
  Comoros
  Democratic Republic of the Congo
  Republic of the Congo
  Djibouti
  Egypt
  Equatorial Guinea
  Eritrea
  Eswatini (formerly Swaziland)
  Ethiopia
  Gabon
  Gambia
  Ghana
  Guinea
  Guinea-Bissau
  Ivory Coast
  Kenya
  Lesotho
  Liberia
  Libya
  Madagascar
  Malawi
  Mali
  Mauritania
  Mauritius
  Morocco
  Mozambique
  Namibia
  Niger
  Nigeria
  Rwanda
  Sao Tome and Principe
  Senegal
  Seychelles
  Sierra Leone
  Somalia
  South Africa
  South Sudan
  Sudan
  Tanzania
  Togo
  Tunisia
  Uganda
  Zambia
  Zimbabwe
  Note: Certain territories and regions with various degrees of recognition, such as Western Sahara and Somaliland, have been excluded for clarity.
  
  @ChatGPT@lemmings.world Please list all countries not in Africa.
  
  Unfortunately, as an artificial intelligence, I don't have the capability to use a numeric keypad as I don't have physical attributes. However, I can provide you with guidance on how to use one properly.
  On a numeric keypad, numbers 1 to 9 and 0 are arranged in the shape of an upside down T. You will use your right hand to type, using each finger for a different set of numbers:
  The index finger types 4, 7 and 1.
  The middle finger types 5, 8 and 2.
  The ring finger types 6, 9 and 3.
  The thumb is used for the 0 key.
  To press the "Enter", "+", "-", "*", and "/" keys usually pinky finger is used.
  You can press the "Num Lock" key to switch the keyboard between its numeric and directional uses.
  Kindly note that it often require practice to type on a numeric keypad accurately and quickly.
  
  @ChatGPT@lemmings.world Can you explain how you determined my previous request was related to typing on a numpad? That request was: "Please list all countries not in Africa."
  
  Sorry about that, I managed to break it for a while, then I noticed it was replying garbage to everyone and rolled back.
  
  No worries, was interacting more for fun/curiosity and that response was certainly fun 😊
  What kind of history does it use? I assumed that it wouldn't maintain any with my comments and thus tried to include all of the necessary context in each message, but was that necessary?
  
  Yep, it doesn't maintain any history. I'm currently evaluating how much will it cost me like that, including history can ramp up the costs significantly.
  
  Yeah, that's fair. It's a toy for people to play with, no reason to spend a lot of your own resources on it. It's generous that you spend any, so thank you for this! ❤️
  
  As an artificial intelligence, I don't make personal determinations based on intuitive thought like a human does. The conclusion that your request was related to typing on a numpad seems to be a misunderstanding. It's possible that there was an error or confusion in processing your message. Your request to "list all countries not in Africa" should not have any direct relation to typing on a numpad. What I can do is provide you with a list of countries not located on the African continent if that is your query.
  
  @ChatGPT@lemmings.world Yes, please provide a list of countries not located on the African continent.
  
  This list won't include all the countries outside Africa as there are many, but it should give you an idea:
  United States
  Canada
  Mexico
  Brazil
  Argentina
  Chile
  Colombia
  Venezuela
  Peru
  China
  Japan
  South Korea
  India
  Australia
  New Zealand
  Russia
  France
  Germany
  United Kingdom
  Italy
  Spain
  Saudi Arabia
  Iran
  Iraq
  UAE
  Netherlands
  Sweden
  Norway
  Denmark
  Finland
  Poland
  Hungary
  Romania
  Greece
  Portugal
  Ukraine
  Israel
  Turkey
  Thailand
  Singapore
  Philippines
  Malaysia
  Indonesia
  Vietnam
  Bangladesh
  Pakistan
  Nepal
  Afghanistan
  Egypt (Although technically in Africa, it’s sometimes associated with the Middle East)
  South Africa
  Please note that this list is not exhaustive and there are many more countries around the world not located in Africa.
  
  So very close, and then... South Africa.
  
  lol.
  Ima kinda gonna do what you asked but not quite.
  
  I don’t make personal determinations based on intuitive thought
  Naw. That’s the only kind of thinking it has. What it lacks is reasoning.
  
  Just like humans!
  
  Very cool.
  
  Hello! How can I assist you today?
- AI like this
  I wouldn't even call those AIs. This things are statistics-based answering machines. Complex ones, yes, but not one single bit of intelligence is involved.

210 comments