Reason for ChatGPT becoming 'dumber'

[GoingToSleep]

I’m not sure if this has been said before or not but I thought this might be important enough to make it its own thread. I have come up with a few reasons why ChatGPT has supposedly become dumber or has not come to answer questions like how it used to.

Profit for customer satisfaction: The people in charge of ChatGPT probably decided that they were not making enough money and wanted to have some more. So with the release of ‘GPT4’, they decided to just make the current one dumber and create a paid version of ChatGPT so people have to buy it. Basically, they are trading customer satisfaction for sales which I think was a huge risk.

Outdated information: ChatGPT’s information is only capped to September 2021 which means its skill level will be not as high as human intelligence because it can not learn new things so it will have a fixed knowledge capacity. This means that ChatGPT will not be able to do things that would have changed from September 2019 leading it to be ‘dumber’.

I have basically outlined the reasons why ChatGPT has gotten ‘dumber’. If you think there is something else, please let me know in the replies.

 

Dislike ads? Remove them and support the forum: Subscribe to Fastlane Insiders.

[tigerbalm]

I don't think that it is just that. As an engineer who has paid for chatGPT since March I can unequivocally say that GPT 4 has gotten dumber as well. I used to use it to build fully fledged APIs and such, and it would remember the context from over 100 messages prior. Now I struggle to get it to retain context from 3 messages prior. My suspicion is that they cut corners to facilitate more requests as when I started using it they capped you at 10 per 3 hours, but now I think it's 50 per 3 hours.

 

[Fox]

The paid version sucks now also.

Likely the problem is this...

ChatGPT was too open and this lead to issues.
What if someone asks a forbidden question like "How many genders are there?".

So now you have to bring in rules.

It can't just give the right answer - it also needs to give the correct right answer.

For LOTS of topics, questions, and ways of doing things.

So they likely had to load it up with 1,000 extra rules since there are many forbidden questions out there.

*You can't say this*
*Don't mention this*
*This person is a bad person*
*This idea is wrong*

So now instead of just giving you an answer - it needs to spend most of the time making sure it ticks all the right boxes and you don't get any dangerous ideas that might actually get you thinking.

Kinda funny to see the whole arc in a short period of time... "wow this is good" to "ahhh this is kinda dumb".

If you see who is calling the shots on all of this it is not too surprising.

 

[ensoniq2k]

So now instead of just giving you an answer - it needs to spend most of the time making sure it ticks all the right boxes and you don't get any dangerous ideas that might actually get you thinking.

This is most probably the case. Remember it gave pretty spot on legal advice? They can't have it do that, they might get held responsibly.

Plus if you're into conspiracy stuff lawyers probably don't like to be replaced by an AI but there is no proof that anything like that actually happened.

 

Dislike ads? Remove them and support the forum: Subscribe to Fastlane Insiders.

[Sandboarder]

The paid version sucks now also.

Likely the problem is this...

ChatGPT was too open and this lead to issues.
What if someone asks a forbidden question like "How many genders are there?".

So now you have to bring in rules.

It can't just give the right answer - it also needs to give the correct right answer.

For LOTS of topics, questions, and ways of doing things.

So they likely had to load it up with 1,000 extra rules since there are many forbidden questions out there.

*You can't say this*
*Don't mention this*
*This person is a bad person*
*This idea is wrong*

So now instead of just giving you an answer - it needs to spend most of the time making sure it ticks all the right boxes and you don't get any dangerous ideas that might actually get you thinking.

Kinda funny to see the whole arc in a short period of time... "wow this is good" to "ahhh this is kinda dumb".

If you see who is calling the shots on all of this it is not too surprising.

I also think bombing the model with unnecessary data is a problem here. Rn I'm working on my software and it incorporates AI models. I've tested some of them and uncensored ones, i.e. those retrained on filtered dataset, are visibly more powerful.

 

[Jrjohnny]

It’s probably because ChatGPT has been blowing up, computational costs are going up.

It’s probably expensive to run an AI to millions of people a day.

Its also probably the millions of stupid questions like “how many genders are there?”

Or “what gender are you?”

People ask the weirdest questions…

 

[ensoniq2k]

It’s probably because ChatGPT has been blowing up, computational costs are going up.

It’s probably expensive to run an AI to millions of people a day.

Its also probably the millions of stupid questions like “how many genders are there?”

Or “what gender are you?”

People ask the weirdest questions…

If the AI model is learning from all the questions it's being asked it would certainly explain why it's getting dumber.

 

Dislike ads? Remove them and support the forum: Subscribe to Fastlane Insiders.

[Aidan04]

GPT (In my opinion) is fantastic, but only in a few limited cases. I've been using it consistently these past few months to help me on some work, and is basically a helpful advisor.

I have noticed it has become limited in memory recently witch can be frustrating at times but you can always correct it. The advantages are better than the drawbacks.

To use a tool, you need to understand it. People are treating it like a magic box.

Things GPT is fantastic at:
- Writing specific code
- Explaining technical details and concepts
- Teaching you hard concepts like economics or physics
- Talking in abstracts or hypotheticals
- Understanding philosophical questions
- Comparing topics/expanding on topics
- Writing copy

It has been immensely helpful. Treat it more like a guide than anything and take things with a grain of salt.

 

[The-J]

ChatGPT is 100% getting worse and you basically need to use it with the Internet for it to be useful. (WebChatGPT is free but you can buy ChatGPT Plus for $20/mo and use plugins with it)

I maintain that the best way to use GPT is to slap it onto an existing base of knowledge & use it to answer questions within that realm of knowledge. Program it to say "I don't know, ask someone else" to anything that isn't within its own base of knowledge.

I'm not convinced that general AI will ever be any good. Specialized AI (expert system) is game changing.

Every LLM programmed in a certain way will have its biases. Liberal, conservative, communist, fascist, you can make it talk however you want. Since it's Silicon Valley liberals running the show, ChatGPT outputs reflect their views. Right now, it's unfeasible to self-host your own LLM so that you can program it with your own worldview, but it won't be that way forever.

 

[Andy Black]

I pay for it so it doesn't tell me it's too busy.

 

Dislike ads? Remove them and support the forum: Subscribe to Fastlane Insiders.

[RicardoGrande]

If the AI model is learning from all the questions it's being asked it would certainly explain why it's getting dumber.

"Dear chat gpt: How 2 BIG smma make money no work outsource SEO optimize my life and and Andrew Tate in porto reeko???!?
(plz respond )"

 

[Aidan04]

ChatGPT is 100% getting worse and you basically need to use it with the Internet for it to be useful. (WebChatGPT is free but you can buy ChatGPT Plus for $20/mo and use plugins with it)

I maintain that the best way to use GPT is to slap it onto an existing base of knowledge & use it to answer questions within that realm of knowledge. Program it to say "I don't know, ask someone else" to anything that isn't within its own base of knowledge.

I'm not convinced that general AI will ever be any good. Specialized AI (expert system) is game changing.

Every LLM programmed in a certain way will have its biases. Liberal, conservative, communist, fascist, you can make it talk however you want. Since it's Silicon Valley liberals running the show, ChatGPT outputs reflect their views. Right now, it's unfeasible to self-host your own LLM so that you can program it with your own worldview, but it won't be that way forever.

Self-hosted/open source AI will be the beginning of a total shift in how work is done. It will be wild to see what kind of applications it will be used for.

Definitely agree with the knowledge base thing. It's like NZT from Limitless: "It works better if you're already smart".

 

[Caminsky]

"Dear chat gpt: How 2 BIG smma make money no work outsource SEO optimize my life and and Andrew Tate in porto reeko???!?
(plz respond )"

"how manifest fire rizz no cap"

 

Dislike ads? Remove them and support the forum: Subscribe to Fastlane Insiders.

[GoingToSleep]

The paid version sucks now also.

Likely the problem is this...

ChatGPT was too open and this lead to issues.
What if someone asks a forbidden question like "How many genders are there?".

So now you have to bring in rules.

It can't just give the right answer - it also needs to give the correct right answer.

For LOTS of topics, questions, and ways of doing things.

So they likely had to load it up with 1,000 extra rules since there are many forbidden questions out there.

*You can't say this*
*Don't mention this*
*This person is a bad person*
*This idea is wrong*

So now instead of just giving you an answer - it needs to spend most of the time making sure it ticks all the right boxes and you don't get any dangerous ideas that might actually get you thinking.

Kinda funny to see the whole arc in a short period of time... "wow this is good" to "ahhh this is kinda dumb".

If you see who is calling the shots on all of this it is not too surprising.

I think they are just trying not to get on any communities bad side.