Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

There is a new version of GPT-4 Turbo now live in ChatGPT

twitter.com

Post reply on HN
Print view View original

61–70 of 91 posts

Re: There is a new version of GPT-4 Turbo now live in ChatGPT

#61

Post by gumballindie »

MattRix wrote: post #35
gumballindie wrote: post #10

Earlier quoted context omitted.

What value does a procedural text generator provide?

I’m a little perplexed at why you feel the need to refer to it in reductionist terms like this? It seems to provide a lot of value considering that it has 100 million weekly active users. I regularly find myself using it a dozen times a day. I feel like the only reason someone wouldn’t find it useful is if they weren’t ever doing anything new.

Loads of useless products have hundreds of millions of users. Threads, Twitter, Facebook, and so on. Just spamming forums to promote this product wont increase its utility.

Re: There is a new version of GPT-4 Turbo now live in ChatGPT

#62

Post by apwell23 »

danielvaughn wrote: post #55
apwell23 wrote: post #50

Earlier quoted context omitted.

> It's one of those things where, the more you think about it, the more confused you get. > it successfully interpreted the picture Sorry, but how did you know that interpretation was a success?

Because it explained what the exercise was. I just tried sharing but it said "sharing chats with images is not yet supported", so I'll paste what it said: 'Assignment 6 on the homework sheet you've provided is a grammar exercise that asks the student to think of two adjectives (describing words) that describe what each of the three pictures can do. The pictures are labeled "bubble gum," "porcupine," and "milk." It se…

My interactions with chatgpt have been frustrating. It cannot seem to process even basic interactions.

Here is an example https://chat.openai.com/share/65c8cbd7-8556-4e02-a726-32eaa8...

What am I doing wrong here ?

Re: There is a new version of GPT-4 Turbo now live in ChatGPT

#63

Post by apwell23 »

terhechte wrote: post #6
retskrad wrote: post #3

At $10/month, ChatGPT-4 is reasonable. At $20, that's a no no!

I use it usually 10-20 times per day. 5-10 when I'm not working and like 20-30 when I'm working. It saves me a huge amount of time. It would already make sense at a wage of $20 but I earn more and so it is so much worth it. Even if I earn no wage (say a indie dev working on a game) it would still be useful because I save time which I can invest in other things (or play video games). I'd honestly pay $50-$100 for it a…

You seem like a power user .

Can you tell me why it can't process basic questions

https://chat.openai.com/c/19a2280c-1dc6-4f77-9313-c61ff4be4d...

What am I doing wrong here ?

Re: There is a new version of GPT-4 Turbo now live in ChatGPT

#64

Post by TeMPOraL »

apwell23 wrote: post #62
danielvaughn wrote: post #55

Earlier quoted context omitted.

Because it explained what the exercise was. I just tried sharing but it said "sharing chats with images is not yet supported", so I'll paste what it said: 'Assignment 6 on the homework sheet you've provided is a grammar exercise that asks the student to think of two adjectives (describing words) that describe what each of the three pictures can do. The pictures are labeled "bubble gum," "porcupine," and "milk." It se…

My interactions with chatgpt have been frustrating. It cannot seem to process even basic interactions. Here is an example https://chat.openai.com/share/65c8cbd7-8556-4e02-a726-32eaa8... What am I doing wrong here ?

Not sure, but I'd check two things:

1) Make sure you're chatting with GPT-4, not GPT-3;

2) Try something other than recipes and lyrics. It so happens that ChatGPT models have been instructed (via system prompt) to specifically not reproduce recipes and lyrics verbatim[0]. Those two categories are verboten - ChatGPT is literally instructed to not repeat recipes found on-line, but rather invent its own.

I suspect you're hitting 2) - sure, you provided your own recipe, but from the POV of ChatGPT, this isn't clearly distinguishable from how "web search" works (run web search, inject results into temporary context, let the model generate reply, send user reply without the temporary context), so it may be acting dumb because it's over-eager to apply the instructions from its system prompt.

(That's a good reason to prefer using API directly via code, or a third-party chat frontend: you get to control the system prompt this way, and can cut out some of OpenAI's restrictions.)

--

[0] - Yes, they literally mention those two things in their instructions. The linked tweet has a reply showing the reconstructed system prompt of newest GPT-4 iteration, where you can spot this, but this also applied to earlier iterations for some time now.

Re: There is a new version of GPT-4 Turbo now live in ChatGPT

#65

Post by apwell23 »

TeMPOraL wrote: post #64
apwell23 wrote: post #62

Earlier quoted context omitted.

My interactions with chatgpt have been frustrating. It cannot seem to process even basic interactions. Here is an example https://chat.openai.com/share/65c8cbd7-8556-4e02-a726-32eaa8... What am I doing wrong here ?

Not sure, but I'd check two things: 1) Make sure you're chatting with GPT-4, not GPT-3; 2) Try something other than recipes and lyrics. It so happens that ChatGPT models have been instructed (via system prompt) to specifically not reproduce recipes and lyrics verbatim[0]. Those two categories are verboten - ChatGPT is literally instructed to not repeat recipes found on-line, but rather invent its own. I suspect you'r…

ok. How about this

https://chat.openai.com/c/312cdb5f-8674-4302-ad8a-e91486c053...

Chatgpt gave me some useless answer.

I asked google and it gave me much much better answer on the search page

"It's difficult to define dream pop as a genre; it's more of a sound and aesthetic than a list of limitations. That's partially why Beach House mastered it so well; over the course of nine albums, the group hardly deviates from the gauzy sonics and vocals that characterize its music."

This probably has something to do with point 2 you mentioned.

Most of my iteractions with chatgpt are useless like these. I would rather use google and correctly sourced human written text.

Maybe I am just using chatgpt wrong?

I know it shines in things like 'write poem about crow and cat for a 3 yr old' but i am surprised by ppl on this thread saying they use it ask questions about economics/sql ect.

Re: There is a new version of GPT-4 Turbo now live in ChatGPT

#66

Post by appplication »

DonHopkins wrote: post #56
jerpint wrote: post #29

Not sure if it’s related or not but I noticed the UX has changed on the web interface and when I ask for code, it now defaults to running it through an interpreter with a hidden view by default. This is very annoying to me as I want to mostly read the code, not have it be executed without any context

Is it smart enough to solve the halting problem and not bother running code with infinite loops?

Infinite loops and the halting problem are separate issues. You can detect infinite loops with basic static analysis.

Re: There is a new version of GPT-4 Turbo now live in ChatGPT

#67

Post by gumballindie »

MattRix wrote: post #42
gumballindie wrote: post #20

Earlier quoted context omitted.

> GPT-4 is definitely closer to the latter than to the former. That’s a seriously dubious statement.

Have you used it? A GPT4 chatbot could convince 90% of people it was a human, right now. It could pass the majority of knowledge-based tests in all kinds of fields. 100 million people find it useful enough to use weekly.

It could convince 90% of the people that are ready and willing to think it was human. Even when it does, it's only a piece of software designed specifically to emulate, therefore it's fake. Regardless of how much people want to claim it is intelligent, it's not. Sure it has load of data at its dispoal to generate well mimicked output, but to some of us it's so obvious and buggy that it's a waste of time.

Re: There is a new version of GPT-4 Turbo now live in ChatGPT

#68

Post by wenc »

apwell23 wrote: post #63
terhechte wrote: post #6

Earlier quoted context omitted.

I use it usually 10-20 times per day. 5-10 when I'm not working and like 20-30 when I'm working. It saves me a huge amount of time. It would already make sense at a wage of $20 but I earn more and so it is so much worth it. Even if I earn no wage (say a indie dev working on a game) it would still be useful because I save time which I can invest in other things (or play video games). I'd honestly pay $50-$100 for it a…

You seem like a power user . Can you tell me why it can't process basic questions https://chat.openai.com/c/19a2280c-1dc6-4f77-9313-c61ff4be4d... What am I doing wrong here ?

That link doesn't open. (Unable to load conversation)

Re: There is a new version of GPT-4 Turbo now live in ChatGPT

#69

Post by wenc »

apwell23 wrote: post #62
danielvaughn wrote: post #55

Earlier quoted context omitted.

Because it explained what the exercise was. I just tried sharing but it said "sharing chats with images is not yet supported", so I'll paste what it said: 'Assignment 6 on the homework sheet you've provided is a grammar exercise that asks the student to think of two adjectives (describing words) that describe what each of the three pictures can do. The pictures are labeled "bubble gum," "porcupine," and "milk." It se…

My interactions with chatgpt have been frustrating. It cannot seem to process even basic interactions. Here is an example https://chat.openai.com/share/65c8cbd7-8556-4e02-a726-32eaa8... What am I doing wrong here ?

I tried your prompts and it gave me the right answers. Also onions, chilies and tomatoes are not vegetables. Your response was actually not correct -- GPT actually gave the correct answer.

Re: There is a new version of GPT-4 Turbo now live in ChatGPT

#70

Post by ben_w »

gumballindie wrote: post #67
MattRix wrote: post #42

Earlier quoted context omitted.

Have you used it? A GPT4 chatbot could convince 90% of people it was a human, right now. It could pass the majority of knowledge-based tests in all kinds of fields. 100 million people find it useful enough to use weekly.

It could convince 90% of the people that are ready and willing to think it was human. Even when it does, it's only a piece of software designed specifically to emulate, therefore it's fake. Regardless of how much people want to claim it is intelligent, it's not. Sure it has load of data at its dispoal to generate well mimicked output, but to some of us it's so obvious and buggy that it's a waste of time.

If it is convincing most people, then surely by definition it's closer to us than to a Markov chain.
Post reply on HN