There is a new version of GPT-4 Turbo now live in ChatGPT
twitter.com
twitter.com
https://www.reddit.com/r/singularity/comments/11vwlct/humor_...
So you could refuse to run any code with a loop whose termination isn't easily proven to be guaranteed[0], without violating the halting problem.
[0] And, in practice, terminating quickly. I doubt they want people to start mining bitcoin on all of those expensive H100s.
My favorite is Meme Machine, where you just specify a concept like "return to office, safari animals" and it generates an image with DALL-E and then composites text over it using Python:
https://www.reddit.com/r/ChatGPT/comments/17tb7rm/mememachin...
I noticed all my GPTs started working better yesterday, must have been the new model rollout.
Try it here: https://chat.openai.com/g/g-MP5Cx7F7W-meme-machine
>>there is a new version of GPT-4 Turbo now live in ChatGPT; you should hopefully find it a lot better!
please let us know what you think.<<
ChatGPT in this case is most likely referring to the implementation accessible via https://chat.openai.com/, once again raising questions for myself whether and how this change may impact the model accessible via their API gpt-4-1106-preview. Does "new version" mean what has been accessible via the chat website was less capable than gpt-4-1106-preview and if so, in what manner, or is gpt-4-1106-preview going to receive certain improvements that are currently being tested via the chat website?
I'd love to see more clarity on that front.
I do feel morally conflicted about how humanity built the entire internet basically for free, and OpenAI is just regurgitating it back to us and getting rich off of our collective data. But the tool is so goddamn useful that I can't help myself regardless.
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 ?
I also use it to be build multi linguistic tasks lists in English, Polish and Ukrainian for Google sheets.
I also just used it to help create crop planting plan for a farm in Kenya which has just had massive floods, with a constraint that local produce reaches the highest value in December and January. The other constraint is that some of the crops have survived but not all. It is an interesting yield management experiment
It is very useful in providing similar examples too. It's way better than a thesaurus.
I pay for but i only use it handful of times after i remind myself to use it .
In the last hour? It told me how to connect an ODB reader to my car and reset the error codes.
(I used to do image searches for those, and there's plenty of images on any simple topic you can think of, but DALL-E 3 lets me quickly make ad-hoc images based on ideas like "child apprehensive about a robot vacuum coming into the room to eat its toys" or whatever they were joking about 10 minutes earlier.)
Two recent examples that comes to mind:
1. I was helping my 6 year old niece with her homework, and the exercise was extremely confusing. It had 3 pictures - a gumball machine, an owl, and something else (can't remember). The instructions said "for each picture, write 3 adjectives that describe what the picture can do." It's one of those things where, the more you think about it, the more confused you get. I thought I must be going crazy. I took a screenshot and asked ChatGPT if it could understand what the exercise was asking for, and it successfully interpreted the picture but agreed that the exercise as described didn't make sense, because adverbs describe actions, not adjectives.
2. A more trivial example, but I quickly needed to know the time and temperature for cooking bacon in the oven. Stuff like this is incredibly irritating with Google, especially if you're in the kitchen with food particles all over your hand. I don't want to read some backstory about the recipe author growing up on a small farm, intermixed with 15 different ads, just so I can figure out how many teaspoons of salt I need.
> it successfully interpreted the picture
Sorry, but how did you know that interpretation was a success?
'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 seems your niece has already filled out the adjectives for the bubble gum as "chew" and "blow," which are actions rather than descriptive adjectives. Adjectives for bubble gum might be "sticky" and "sweet." For the porcupine, adjectives like "spiky" and "prickly" would be appropriate, and for the milk, "cold" and "creamy" could work. The task is to use descriptive words, not actions or verbs.'
So then I said:
'What’s confusing is that it says “an adjective describing what each picture can do”. Isn’t that a verb?'
And it responded:
'You're correct; there seems to be a misunderstanding in the instructions. Adjectives are words that describe the qualities or states of being of nouns—like 'red,' 'quick,' 'tall,' 'sleepy,' etc., while verbs are action words like 'run,' 'talk,' 'eat,' 'do.'
The phrase "describing what each picture can do" indeed suggests action, which would typically be described by a verb. However, since the instructions are asking specifically for adjectives, it seems they want descriptive words that can indirectly suggest an action or capability. For instance:
For bubble gum: "stretchy" (since it can be stretched) and "flavorful" (since it has a taste). For the porcupine: "sharp" (since it has quills that can prick) and "defensive" (since it can use its quills to defend itself). For the milk: "refreshing" (since it can be refreshing to drink) and "nutritious" (since it provides nutrition). These adjectives describe the qualities of the nouns that relate to what they can do without being action words themselves.'
Here is an example https://chat.openai.com/share/65c8cbd7-8556-4e02-a726-32eaa8...
What am I doing wrong here ?
If you look at the first response. First item in there was onions. Second was tomatoes.
> GPT actually gave the correct answer.
what? You just said saying onions is wrong ?
It gave the correct answer but you insisted it missed onions and tomatoes. Yes they were part of the recipe but those aren’t vegetables. Why did you expect it to return onions in its answer?
Do you consider onions and tomatoes to be vegetables? (They are not)
(I also make biryani and there are usually no vegetables in biryani and there was definitely none in the recipe you pasted)
Because Chatgpt said onions and tomatoes are 'commonly used vegetables' ? That was my first question. Can you please reread chatgpt answer carefully where it says onions and tomatoes are 'commonly used vegetables' .
Its not relevant what i consider vegetables. I built my question on what chatgpt said it considers as vegetables.
Unrelated:
> Do you consider onions and tomatoes to be vegetables? (They are not)
https://en.wikipedia.org/wiki/Onion
"An onion (Allium cepa L., from Latin cepa meaning "onion"), also known as the bulb onion or common onion, is a vegetable that is the most widely cultivated species of the genus Allium. "
(And of course the whole issue is bullshit - it stems from us using words that are kinda, but not quite, genotype classifiers, in contexts where we want phenotype classifiers. Tomato is a fruit, but does not go with fruit salad. A whale is a mammal, but you use a fishing ship to hunt for it, etc. There are grasses you cut with a chain saw, and trees you cut with a lawn mower. Plenty of legacy descriptors in common language.)
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.
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.
For example I don't know much about electronics so I asked it how design a circuit for a problem I had and it came up with a useful answer.
It is the accessibility. I can go and ask an esoteric question about economics and it an answer immediately. Along with that answer comes enough information that I can dig more myself if I want.
Can you give me a specific example of esoteric economics question you referred to. Would love to compare UX with google and chatgpt.
The other effect at work here is anchoring. OpenAI wants people to be prepared to pay more for AI. Setting a higher price anchors that price in people’s minds and makes it feel normal. This will serve them better in the future since the long-term objective is to provide services that replace employees, who are surely paid more than $20/month.
Cost is their business, not yours (unless you're a competitor to them).
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.
It can write boring code, like the OpenStreetMap query to find all supermarkets within 50x50km of Berlin.
It can do code and literature reviews of my solo projects and stories.
It can take a list of ingredients and suggest things I can cook with them, and unlike "real" recipe websites it can do this without telling me how their grandmother used to comfort them with this on a beautiful autumn weekend, when… [1000 words and four ad breaks later] … and then when the website actually gets to the recipe it turns out to be unsuitable anyway.
It can write form letters
It can help role-play difficult real-life scenarios.
-
I don't expect perfection in any of these things. If it was perfect, us software developers would already be redundant, as would all the lawyers, the accountants, and anyone else whose job can be reduced to images and words.
Information can be abundant yet hidden in obtuse or misleading writing. LLMs produce writing that is quite accessible.
That’s a seriously dubious statement.
https://openai.com/enterprise-privacy
Though I'm not sure if this covers other analysis of the data, or if it's something only for Enterprise API users.