Announcement: AI generated answers are officially banned here
english.meta.stackexchange.com
english.meta.stackexchange.com
It’s trained to detect GPT-2.
Although approaches to predicting GPT-2 similarity should be somewhat comparable with ChatGPT since they share the same tokenizer.
Would this not be a better "middle" ground? Thoughts?
If people really want to be careless they can get AI generated code from copilot or chatgpt on their own already, I don't think this would be worse than that.
It's a pity, because you've also posted good comments and I think the proportion of good comments has been getting better over time, which is great, but that doesn't make things like the above ok. Also, you have a history of using this site for ideological battle and we don't want that here—it's not what this site is for, and destroys what it is for.
If you don't want to be banned, you're welcome to email hn@ycombinator.com and give us reason to believe that you'll follow the rules in the future. They're here: https://news.ycombinator.com/newsguidelines.html.
* Edit: on closer look, I can't tell if it might have been a bad joke instead.
[And don’t believe ChatGPT that claims it is an only a Language Model. It is not. It is a RL Agent trained with PPO.]
Does it apply to things that aren't programming? e.g. People are already using these AIs for legal work.
But, perhaps considering AI as an adversary is a bad idea. And alignment along the “Love is all you need” lines is the actual solution. Tricky problem…
So then the asker gets the wrong answer and no one else can correct it? Seems even worse.
If they stick to their guns, Stack Exchange won't be around in a decade. That's how fast the world is about to change.
Human evaluators can filter out the (for now) intermediate bad results. All of our AI models and products will climb a quality gradient year by year to the point where this will be increasingly less necessary.
To your other point, you can use machine data to bootstrap a "real" model. My team has done this to incredible effect.
ChatGPT (and presumably soon CoPilot) learn via * Reinforcement Learning from Human Feedback*[1] on top of the raw language model. This takes low numbers of high quality examples to improve the quality, and avoids the "junk training data" entropy problem you identify.
The OpenAI Codex text-davinci-002 and text-davinci-003 models have been trained for code generation using this human feedback process[2].
[1] https://huggingface.co/blog/rlhf
[2] https://beta.openai.com/docs/model-index-for-researchers
Technically, they aren't trained that way, their opposite recognition AI is trained that way. That's less than ideal when their original set of training data aren't drawn from their own output, only from a pool of humans influenced by their output. It becomes suicidal to the AI when it begins consuming its own output through what it considers to be trusted channels.
[edit] The ancient coder phrase "garbage in, garbage out" has never been more significant than now in the context of what's fed into neural nets and ML algos. I think it's basically silly hubris sprouting from a lack of underpinning technical understanding when people assert these models will train themselves without much more extreme means of assessing their output than can be provided by a few minimum wage workers somewhere.
No, they are trained that way. These aren't GANs or anything resembling that. There is a reinforcement model that is build from expert human guidance which is then used to fine-tune the LM.
From the HuggingFace explainer page I linked above: [there are] three core steps:
1. Pretraining a language model (LM),
2. gathering data and training a reward model, and
3. fine-tuning the LM with reinforcement learning.
Read https://arxiv.org/abs/2203.02155 if you want all the details (although the HF explainer is easier to follow). From the abstract of that paper:
> Starting with a set of labeler-written prompts and prompts submitted through the OpenAI API, we collect a dataset of labeler demonstrations of the desired model behavior, which we use to fine-tune GPT-3 using supervised learning. We then collect a dataset of rankings of model outputs, which we use to further fine-tune this supervised model using reinforcement learning from human feedback.
There is no "opposite recognition AI".
Maybe you could get some kind of browser extension malware that scrapes the things people are searching to use as input.
Give it a month before Google Search is overrun with AI-generated SEO spam (and a year before the next generation of AI has been trained on the aforementioned spam).
The obvious prediction is that the vast majority of sites will try to ban software generated text (while probably covertly allowing some for motivations from increasing user engagement to fulfilling government "requests"), but another equally obvious prediction is that such software will gradually become much more accessible, including being able to compile/tweak it at home, which means any effort to put the genie back in the bottle is certainly doomed to failure.
We may be living through the end of the era of being able to believe that a blurb, like this one, was actually written by some human somewhere. The implications of this seem nearly as impossible to imagine, as going back 30 years ago to trying to imagine the implications of us being able to post and exchange text/data with each other on a global network.
Interesting times we are living through, seemingly as always.
There was a discussion about that, recently: AI: Markets for Lemons, and the Great Logging Off (278 comment)
https://news.ycombinator.com/item?id=34169051
Before the machines being able to generate cr*p that looks authentic at glance, the assumption was that the screen we look at had human generated patterns. Since this is no longer the case, it's nowhere nearly interesting.
It actually had been losing value since these patters were generated for profit by army of people but with the arrival of the convincing AI generated text, the process has been greatly accelerated.
ChatGPT is very cool when it generates output upon my request, similar technology used for pretending being written by people is toxic. The problem with these AI is that they look convincing but they don't know what they are talking about(like the worst kind of a person). This is not only problem because often its not accurate(people are wrong too all the time) but because it doesn't contain the core reason we interact with people: To change their life and opinions or make them do something. AI can pretend to be different person but all AI out here is the same machine trained on the same things so it's like talking to the same annoying person everywhere all the time. AI can become interesting once it's trained individually like a human AND our input becomes very influential on them(for example if we can get an AI trained Cairo be interested in French literature by telling is something fascinating about it).
I guess the internet 2.0 will die off in spam and content will be created and shared in well moderated communities that can guarantee high ratio of people. WhatsApp or Telegram groups are very popular these days.
I like to farm karma as a hobby, and part of my work involves harvesting a lot of the most upvoted comments in discussions and using them as training data to generate new comments that have high potential for upvotes. Eventually this AI can be deployed to build up new accounts that have high karma.
This is a very weird and muddled description of what a language model does, let alone "AI" in general.
That said some professors have used the tool mentioned here to check for Chat GPT sourced content in their student's submissions https://medium.com/geekculture/how-to-detect-if-an-essay-was...
For a glance at the headlines that it has been inserting in the news, this seems plausible.
I have a sneaking suspicion that in a few years time, sites that explicitly ban AI content will either reverse their decision or become a thing of the past. AI tools are very quickly becoming accessible to the masses and that lets the masses create more and/or higher-quality content -- and, IMO, that's a very good thing.
But, obviously, established sites always struggle when they suddenly receive a large influx of new users/content, especially when they're at odds with (or completely oblivious to) the societal "norms" already established on those sites.
Yes in theory. In practise in this case specifically the problem is often that it is very difficult to judge the quality of the actual content in and of itself. This is already a problem for stackoverflow, and you will see many highly-upvoted incorrect answers to questions with correct answers languishing. Having AI-generated plausible answers to everything would likely further muddy the waters.
It's just another "StackExchange Mod Tool/Policy" with which to oppress the curious innocent masses. YoUr QuEsTiON is UnCLeAR--said the SPhynX. Then they press the button and you fall through the trap door into Jabba-the-Hutt's underthrone dungeon, of "closed as poorly worded / likely witchery" questions. Ugh...
Will the evil domino of "No Bots Allowed" fall at HN next? "You sounds like a bot. Off with yer head!"
You jest, but people are already asking/accusing others of ChatGPT writing their human-written comments here on HN.
I mean, unless somebody wants to manipulate people by training AI to spread crypto misinformation, but with the quality of ChatGPT at this point, it's going to be obvious and have the opposite effect.
You saw this maybe 5-10 years ago when the "50c army" was at its height spamming specific topics. The good comments were just impossible to find because of all the other junk in there.