>>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'm guessing the reason the model is listed as "preview" is so that they have this leeway to tweak it and optimize it as needed, for example in terms of the balance between performance/cost and capability which seems to be the case here.
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 as it saves me so much time and gives me so much value.
I think it's because AI is a crutch for junior engineers, and a rocket for senior engineers. It takes some degree of wisdom to know how to use it effectively, and the idea that you're going to lean entirely on it to make your program work is...misguided at best.
I went for a walk and had a conversation about Rademacher complexity and VC-dimension of a classifier, then I asked it to quiz my understanding based on what we talked about. It's almost at the level on "Her" but you have a very knowledgeable tutor to ask who won't judge and is always available.
Until ChatGPT can store and reference all past conversations, I'd suggest asking it to generate flashcards for e.g. Anki. I've mainly used it when I want to understand or deepen my understanding of something pretty specific.
I use it for several purposes. I most recently used it to find a recèpe for Limecreate for some Ukrainian builders and translate the recipe to Ukrainian.
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
My team was facing an issue with their microservice not recovering after a Kafka outage, which we only observed once and we couldn't reproduce, and didnt understand why not. Googling the symptoms did not yield anything. ChatGPT gave us a theory which happened to be correct, which was that the outage needed to be long enough for the kerberos ticket needed to expire. I estimate without this answer from chatgpt we would have wasted at least a day of our own or other people's time in investigation, but it was solved in minutes.
Same, I'd pay way more than $20/month. It saves me so much time it's insane. For the past several months I'd been trying to write a tree sitter parser, and I had a really hard time with the docs for some reason. I sat down with ChatGPT and had a working V1 parser in like 3 hours. And it wasn't just feeding me instructions, I was actively conversing with it, going off into tangents when I didn't understand something, etc. It's just...magic. I love it.
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.
Depends on where it sits on a spectrum between, say, a dumb Markov chain, and a human like you and me. GPT-4 is definitely closer to the latter than to the former.
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.
No level of capability is enough to convince some people it's not just a stochastic parrot. You're completely right on the Turing test. At this point it's just a matter of how long it could fool us. You have to remember it's not specifically trained to fool us either, it's trained through RLHF to be verbose and "helpful". If you train it right and add in some RAG (vector databases) I think you could fool most people for a pretty long time, some gullible people forever.
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.
I don't work in tech and I use LLaMa. It is great for writing overviews of a subject which I use to make sure I have most of my bases covered (and helping me see which of my ideas are more original), getting more examples, along with other creative tasks. Getting it to compare two subjects to see what they have in common is usually quite helpful.
The issue is i tried using these types of tools for a while and each response is inaccurate. It becomes annoying to constantly ask it to correct it or having to doubt the output. Works well for what is statistically likely to be abundant in information but you know that’s usually of little value.
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.
It can be an infinitely patient teacher for physics and German, that never gets frustrated by me getting the conjugations wrong or having whacky ideas like getting lift from plasma held in a magnetic field.
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.
You have to consider costs. If it costs them $15/month to provide the service (just a made up number) then $10/month would not be reasonable. The Plus users also subsidize the Free users to a certain extent.
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.
But therein lies the crucial question, does the actual business value of the service exceed the costs of running the service + reasonable profit margin? If not, then no amount of anchoring will help.
Nah, you have to consider how much time it saves for you. If your rate is $30/hr and ChatGPT saves you 1 hours per month more than free alternatives then why not?
Cost is their business, not yours (unless you're a competitor to them).
Even now that it has image generation via DALL-E 3, image comprehension via GPT-4V, and the Android app can hold a voice conversation with you, with unprecedented voice recognition accuracy?
Obscure command-line invocations, some specialist topic that's not discussed much online. Usually I Google something first and if I can't see any straightforward answer in a couple of minutes I'll ask ChatGPT.
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.
for me it’s really worth it, and use it as a starting point rather than spending time sifting through garbage google results. I am using it to study math, to ask about css frameworks, about api best practices, SQL dialects, and so many more things. Then i can drill in and go to source documents if I want to. It took me a while to see the value, however now i feel i have a “generalist” assistant and use it as much as i can.
I don’t think it has any knowledge that only it knows. That would be silly.
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.
There are all sorts of questions I've had that are either very hard, or very annoying, to use Google for. They're often technical, but really they could be anything.
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.
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 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.'
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.
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.
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.
Yes. The chat is gone now, but from what I saw, you explicitly asked about shopping for vegetables.
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)
"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. "
I apologize. There’s a lot of debate as to whether onions and tomatoes are fruits or vegetables and they can both be vegetables according to many sources.
As the saying goes, knowledge is knowing onions and tomatoes are fruits; wisdom is not putting them in a fruit salad.
(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.)
Curiosities may be plenty enough. I find the entertainment value of it to be worth $20 alone by now. A handy tool to ad-hoc generate coloring outlines for my kids based on whatever they're interested at the moment is a bonus.
(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.)
I use it constantly throughout the day. Queries that I would otherwise hit google for on esoteric topics. Software wise I like to use it when I am using a library I am very unfamiliar with. This tends to happen for me on FE work as I don’t typically do as much there. I can describe what I want and get a code example I can use on top.
In the last five minutes? I asked it what the model number of a door handle I had that I needed to get replaced. It was able to provide with a Home Depot link for the exact model.
In the last hour? It told me how to connect an ODB reader to my car and reset the error codes.
I use the API now, but I have to maintain my own "Chat" instance. I'd love to be able to just pay $20, but also keep my history. Right now I'm spending less than $10 a month on GPT-4 API costs, so maybe I don't need to upgrade.
The new model combined with the GPTs feature is incredible. I made 10 GPTs yesterday.
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:
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
On the bright side, in some cases at least it includes a test case, runs it, and if it fails then reads the error and tries to fix the code, all in a single run.
I think you can detect either all infinite loops or only infinite loops, but there are many loops which might or might not be infinite under any possible test?
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.
Now it also annoyingly searches almost anything on Bing. I was interested in some coffee machine recommendation/research and it started searching on Bing, which I would have done myself if I wanted to do that. Only after explicitly instructing it not to search for information online I could start asking it the interesting questions.
90 comments
[ 3.4 ms ] story [ 177 ms ] thread>>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.
It is very useful in providing similar examples too. It's way better than a thesaurus.
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
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 ?
That’s a seriously dubious statement.
Information can be abundant yet hidden in obtuse or misleading writing. LLMs produce writing that is quite accessible.
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.
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 pay for but i only use it handful of times after i remind myself to use it .
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.
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 ?
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.
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)
https://imgur.com/a/vzMdwWl
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.)
(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.)
In the last hour? It told me how to connect an ODB reader to my car and reset the error codes.
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.
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
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.
https://www.reddit.com/r/singularity/comments/11vwlct/humor_...