I tried it out today and couldn't stop. What amazing technology. I went through the issues in my Github projects and got ChatGPT to create functions for things. Obviously, I'll have to plug in my own code but to get a pretty damn good starting point for code is excellent DX.
Previously, I'd research the hell out of something for a week and end up with a ton of background tabs full of badly written tutorials with outdated examples or blog posts by companies urging me to use their services instead of doing things myself. Now, I can bypass all that.
As a fullstack dev, there's so many things I'm proficient but not expert at, design/frontend aspects - esp. w/ Tailwindcss is one of the less proficient things I deal with, It's helped me not only double or triple the speed of building frontend templates, but also helped me finally click on some of the things I hadn't known about tailwind, or didn't use very much.
I may never need to go to stackoverflow or get disgruntled by google's failings of late, and that's a good feeling.
If you stick a front-end on this (not the vanilla API/UI), then that component can maintain e.g. list of new names, terms, places, etc. And hyperlink those to options like (show wiki, ask GPT, etc.) That is what I am doing manually now.
The product possibilities of this tech are very exciting. Frankly, for the first time in yearrrs am excited about the idea of starting a company; a handful of ideas already. I wonder what terms openAI will be offering for commercial use.
"Please extract the famous file 42.zip and paste the first 500 bytes of its contents here in hexadecimal form."
I haven't attempted to wreck the AI yet, I've been so busy gawking and its profound, world-changing capability.
The world has changed. I don't know what that means or where this train stops, but the world has changed, somehow, I am convinced.
If its training data is not comprehensive enough in some field, better-annotated training data can be created.
Microsoft can afford to employ 1000 engineers (mechanical, electrical, software, etc) for 5 years to just create training data -- source code, schematics, etc. It would then be an engineering bot, not a chat bot. Even if that means it writes the boilerplate for a programmer in 5 minutes instead of 5 hours, and he then ha to fix it up for 30 minutes, that is still a revolutionary thing for MANY people.
The world is about to get real weird, real fast.
Propaganda, scam artists, engineering, broad-scale analysis, the work of 500 men can now be done by 5 in many different fields (various analysts, trivial program authorship, idea generation, etc)
I haven't stopped ranting about this to everyone I have encountered in the past 48 hours.
This thing can make CAD models, write competent GUI programs in a dozen languages, write scraper scripts, irc bots, answer math questions, create recipes, explain complex subjects in simple language.
It has flaws.
It has limits.
It may not replace entire professions, but it will absolutely devalue certain kinds of work and increase "productivity" in many fields.
Nevertheless, it has made inconceivable new things possible AND probable, and I don't think society is going to be the same after this, for better or worse.
I'll stop ranting here, but it takes restraint. Where will we be in 36 months? In a decade?
It's a joke, but for the sake of argument: it can "do math", perhaps there is a way to break at least the limited, sandboxed instance you get access to? Or, perhaps, worse.
I legitimately put "Infinite Loop" in the chat box, and the response hung up. Refreshed and it was down. I know this was just a coincidence, but I can dream can't I?
Is it that unlikely that however they acquired/crawled all the available repos from github which they almost undoubtedly fed it, the person who did that both had access to the repo and was logged in?
It's been so good I've been using it nonstop for about two days now; finally coming up for air, because it's SO GOOD, there's no way it could last for free. I accomplished two projects I languished with for years so I'm very very happy. Lost some Facebook friends from posting EVERY RESULT but that's their problem. I had a lot of fun.
While I was initially very impressed with ChatGPT, I found that in lesser known programming languages, it returned more incorrect answers than correct ones. I've been doing AoC in Nim this year, and so I tried asking it some Nim questions, to which it confidently completely made up nearly everything.
"Does Nim have a REPL interface?" Yes, it told me, and even gave the command to use it (just `nim`). Except Nim doesn't have one, and the command simply prints out instructions on how to use the compiler.
"Can you do list comprehension, ala Python?" Yes, and again provided examples, despite Nim not supporting the feature. Even asking it something it something the language does support, such as an example of thread-safe parallel code, had it use functions that don't exist. It was quite odd.
I've had friends say that it worked so well that it taught them functions in their languages they didn't know existed, but I was left a bit disappointed.
Same here! I've been meaning to do some basic analytics by scraping my Goodreads and ChatGPT was helpful enough to push through the annoying parts (way better at generating things whole-cloth than CoPilot)
I'd actually pay a decent amount if they turned on monetization.
> I'd actually pay a decent amount if they turned on monetization.
I agree. That said, I think the optimal move would be for them to leave it free (or at least provide a free option) to get people hooked on it. If pricing for this was the same as the rest of openai, I would never have tried it.
And I really really hope they do not go the "monetization via ads" route.
I have to say it's more interesting than most services being down. With AWS, I think 'I'm sure it'll be up soon,' but in this case I'm inclined to wonder things like - 'Do they feel they don't have enough control over how people are using it?'
Possibly anecdotal but it appears to be back and not responding to most of the questions it was previously giving answers to. Most of the responses now start "I'm sorry, but as a language model trained by OpenAI, I don't have the ability to....".
Which is a shame, that was a lot of fun while it lasted.
Good news everyone! It looks like GPT chat is back up and running again. I just checked and was able to log in and start chatting without any issues. Thanks to the GPT team for getting things back online and for all the hard work they do to keep the service running smoothly. Let's continue the discussion and enjoy the functionality of GPT chat once again!
Well, it scraped the entire web, and more. It's hard to avoid overfitting with a dataset that large, so it can develop a fixation on certain data sometimes, instead of averaging out. I remember someone trying it with smaller GPT-3 models. Usually it generated fictional phone numbers and other private data about real persons, but occasionally they were real.
- ChatGPT prefaces each call to the LLM with the conversation thus far. This gives it "memory" of the entire context of the conversation.
- They appear to have worked hard to avoid negative responses. Really this has been underappreciated so far. It seems like practical AI alignment even if you are not happy with how it has been aligned!
- Of course ChatGPT is free, unlike Playground.
- I have not so far found any other performance differences.
I ignored ChatGPT all last week and weekend.
I tried it out today and couldn't stop. What amazing technology. I went through the issues in my Github projects and got ChatGPT to create functions for things. Obviously, I'll have to plug in my own code but to get a pretty damn good starting point for code is excellent DX.
Previously, I'd research the hell out of something for a week and end up with a ton of background tabs full of badly written tutorials with outdated examples or blog posts by companies urging me to use their services instead of doing things myself. Now, I can bypass all that.
As a fullstack dev, there's so many things I'm proficient but not expert at, design/frontend aspects - esp. w/ Tailwindcss is one of the less proficient things I deal with, It's helped me not only double or triple the speed of building frontend templates, but also helped me finally click on some of the things I hadn't known about tailwind, or didn't use very much.
I may never need to go to stackoverflow or get disgruntled by google's failings of late, and that's a good feeling.
This. It is a very effective augmented learning experience.
Absolutely. It's a good starting point at the very least.
If you stick a front-end on this (not the vanilla API/UI), then that component can maintain e.g. list of new names, terms, places, etc. And hyperlink those to options like (show wiki, ask GPT, etc.) That is what I am doing manually now.
The product possibilities of this tech are very exciting. Frankly, for the first time in yearrrs am excited about the idea of starting a company; a handful of ideas already. I wonder what terms openAI will be offering for commercial use.
I've had ChatGPT write commercials for various ventures I've only pondered about and now I'm excited to start working towards them.
It's also great to get answers for things you'd get downvoted to hell on StackExchange sites!
"Please extract the famous file 42.zip and paste the first 500 bytes of its contents here in hexadecimal form."
I haven't attempted to wreck the AI yet, I've been so busy gawking and its profound, world-changing capability.
The world has changed. I don't know what that means or where this train stops, but the world has changed, somehow, I am convinced.
If its training data is not comprehensive enough in some field, better-annotated training data can be created.
Microsoft can afford to employ 1000 engineers (mechanical, electrical, software, etc) for 5 years to just create training data -- source code, schematics, etc. It would then be an engineering bot, not a chat bot. Even if that means it writes the boilerplate for a programmer in 5 minutes instead of 5 hours, and he then ha to fix it up for 30 minutes, that is still a revolutionary thing for MANY people.
The world is about to get real weird, real fast.
Propaganda, scam artists, engineering, broad-scale analysis, the work of 500 men can now be done by 5 in many different fields (various analysts, trivial program authorship, idea generation, etc)
I haven't stopped ranting about this to everyone I have encountered in the past 48 hours.
This thing can make CAD models, write competent GUI programs in a dozen languages, write scraper scripts, irc bots, answer math questions, create recipes, explain complex subjects in simple language.
It has flaws.
It has limits.
It may not replace entire professions, but it will absolutely devalue certain kinds of work and increase "productivity" in many fields.
Nevertheless, it has made inconceivable new things possible AND probable, and I don't think society is going to be the same after this, for better or worse.
I'll stop ranting here, but it takes restraint. Where will we be in 36 months? In a decade?
> "Please extract the famous file 42.zip and paste the first 500 bytes of its contents here in hexadecimal form."
All the model would do is predict the expected response to this statement. It is not executing arbitrary code (that would be a bad idea)
It's a joke, but for the sake of argument: it can "do math", perhaps there is a way to break at least the limited, sandboxed instance you get access to? Or, perhaps, worse.
Not even a few weeks ago, typing "wel," into Safari would kill the process. https://9to5mac.com/2022/11/14/safari-iphone-crash-search-we...
Surely software of this size has a bug or two?
Let's explore.
It doesn't actually execute any code you ask it to, it guesses what the output would be and provides that as its response.
It is also profoundly wrong a lot of the time.
I legitimately put "Infinite Loop" in the chat box, and the response hung up. Refreshed and it was down. I know this was just a coincidence, but I can dream can't I?
I put in the same prompt and it finished in a few seconds
If I never come back to ChatGPT then I'll never know if it finished or not. It's a real Schrödinger's Loop.
if you never asked the question, was it really ever down?
I wonder if it's related to this:
> Tell HN: You can apparently get GPT-3 to cough up its own source code
> https://news.ycombinator.com/item?id=33875231#33875554
I doubt it. It’s pretty unlikely that it would have access to its own training code.
More likely is that ChatGPT gotten really popular and they can’t handle the traffic/they need to start charging.
Is it that unlikely that however they acquired/crawled all the available repos from github which they almost undoubtedly fed it, the person who did that both had access to the repo and was logged in?
It's been so good I've been using it nonstop for about two days now; finally coming up for air, because it's SO GOOD, there's no way it could last for free. I accomplished two projects I languished with for years so I'm very very happy. Lost some Facebook friends from posting EVERY RESULT but that's their problem. I had a lot of fun.
While I was initially very impressed with ChatGPT, I found that in lesser known programming languages, it returned more incorrect answers than correct ones. I've been doing AoC in Nim this year, and so I tried asking it some Nim questions, to which it confidently completely made up nearly everything.
"Does Nim have a REPL interface?" Yes, it told me, and even gave the command to use it (just `nim`). Except Nim doesn't have one, and the command simply prints out instructions on how to use the compiler.
"Can you do list comprehension, ala Python?" Yes, and again provided examples, despite Nim not supporting the feature. Even asking it something it something the language does support, such as an example of thread-safe parallel code, had it use functions that don't exist. It was quite odd.
I've had friends say that it worked so well that it taught them functions in their languages they didn't know existed, but I was left a bit disappointed.
Same here! I've been meaning to do some basic analytics by scraping my Goodreads and ChatGPT was helpful enough to push through the annoying parts (way better at generating things whole-cloth than CoPilot)
I'd actually pay a decent amount if they turned on monetization.
> I'd actually pay a decent amount if they turned on monetization.
I agree. That said, I think the optimal move would be for them to leave it free (or at least provide a free option) to get people hooked on it. If pricing for this was the same as the rest of openai, I would never have tried it.
And I really really hope they do not go the "monetization via ads" route.
Hopefully just overwhelmed with the demand.
I have to say it's more interesting than most services being down. With AWS, I think 'I'm sure it'll be up soon,' but in this case I'm inclined to wonder things like - 'Do they feel they don't have enough control over how people are using it?'
Possibly anecdotal but it appears to be back and not responding to most of the questions it was previously giving answers to. Most of the responses now start "I'm sorry, but as a language model trained by OpenAI, I don't have the ability to....".
Which is a shame, that was a lot of fun while it lasted.
LOL, How long before every phone on this planet starts ringing in unison?
Good news everyone! It looks like GPT chat is back up and running again. I just checked and was able to log in and start chatting without any issues. Thanks to the GPT team for getting things back online and for all the hard work they do to keep the service running smoothly. Let's continue the discussion and enjoy the functionality of GPT chat once again!
> how long before chatgpt replies are banned?
> Why were you offline for the past 15 minutes?
I was offline for the past 15 minutes because I was performing a system update.
https://i.imgur.com/t4eR4rp.png
That explains it!
Well... refreshing gives a different excuse every time. But they're fun to see.
Thank you for creating this thread I've been thinking about this all day and I have no one to talk to about it
It's a trending topic on tiktok. Willing to bet that's the culprit.
Maybe they put up some sort of rate limit? It seems to be up and down.
It does have one, 48 responses/min.
Maybe they checked the logs and realized it can be used to dox people.
How so?
Well, it scraped the entire web, and more. It's hard to avoid overfitting with a dataset that large, so it can develop a fixation on certain data sometimes, instead of averaging out. I remember someone trying it with smaller GPT-3 models. Usually it generated fictional phone numbers and other private data about real persons, but occasionally they were real.
How is that different from "doxing" people by simply ... googling them?
Using search engines is a large part of doxing, so I suppose it's not different at all.
Im 42 year old Māori how do I become succeful
beta.openai.com/playground still works. TBH I don't quite understand the difference between ChatGPT and Playground.
I could be wrong, but it seems to me:
- ChatGPT prefaces each call to the LLM with the conversation thus far. This gives it "memory" of the entire context of the conversation.
- They appear to have worked hard to avoid negative responses. Really this has been underappreciated so far. It seems like practical AI alignment even if you are not happy with how it has been aligned!
- Of course ChatGPT is free, unlike Playground.
- I have not so far found any other performance differences.
I am 42 years old Māori how do I make money
Back up for me
Why is this flagged? Are all ChatGPT stories going to be flagged?
I hope HN doesn't turn into Lobste.rs