Training from scratch and retraining is expensive. Also, they want to avoid training on ML outputs as samples, they want primarily human made works as samples, and after the initial public release of LLMs it has become harder to create large datasets without ML stuff in them
There was a good paper that came out recently saying that training on ml data will result in a collapse of cohesion. It’s going to be real interesting, I don’t know if they’ll be able to train as easily ever again
I recall spotting a few things about Image Generators having their training data contaminated using generated images, and the output becoming significantly worse. So yeah, I guess LLMs and IGA’s need natural sources, or it gets more inbred than the Habsburgs.
Hey, did you know your profile is set to appear as a bot and as a result many may be filtering your posts and comments? You can change this in your Lemmy settings.
Unless you are a bot… In which case where did you get your data?
I really don’t know, I’m speculating, but neither does openai know, that’s sure. So we have the most popular ML system used by millions based on…what exactly?
To be fair this tweet doesn’t say anything about training data but simply that it theoretically can use present day data if it looks it up online.
For gpt4 i think its was initially trained up to 2021 but it has gotten updates where data up to december 2023 was used in training. It “knows” this data and does not need to look ut up.
Whether they managed to further train the initial gpt4 model to do so or added something they trained separately is probably a trade secret.
Pretty sure that cutoff is often used because after that point ai generated content started to appear much more frequently, and the training data becomes corrupted.
Anyone know why most are a 2021 internet data cut off?
Training from scratch and retraining is expensive. Also, they want to avoid training on ML outputs as samples, they want primarily human made works as samples, and after the initial public release of LLMs it has become harder to create large datasets without ML stuff in them
There was a good paper that came out recently saying that training on ml data will result in a collapse of cohesion. It’s going to be real interesting, I don’t know if they’ll be able to train as easily ever again
I recall spotting a few things about Image Generators having their training data contaminated using generated images, and the output becoming significantly worse. So yeah, I guess LLMs and IGA’s need natural sources, or it gets more inbred than the Habsburgs.
I think it’s telling that they acknowledge that the stuff their bots churn out is often such garbage that training their bots on it would ruin them.
I think it’s just that most are based on chatgpt which cuts off at 2021.
Hey, did you know your profile is set to appear as a bot and as a result many may be filtering your posts and comments? You can change this in your Lemmy settings.
Unless you are a bot… In which case where did you get your data?
The data wasn’t stolen, I can at least assure you of that
You paid Hoffman?
Where do you get that from? At least ChatGPT isn’t limited to data from 2021. I haven’t researched about other models.
Gpt 3.5 is limited to 2021. Gpt 4; 4.5; the imaginary upcoming gpt 5 models are not, but that does not mean they aren’t limited in their own ways.
Are you sure those aren’t trained until 2021, frozen, and then fine tuned on later data?
I really don’t know, I’m speculating, but neither does openai know, that’s sure. So we have the most popular ML system used by millions based on…what exactly?
Yeah GPT 3.5 and some other FOSS models also say 2021
OpenAI stated in a tweet a few months ago that the limitation is no longer in place.
To be fair this tweet doesn’t say anything about training data but simply that it theoretically can use present day data if it looks it up online.
For gpt4 i think its was initially trained up to 2021 but it has gotten updates where data up to december 2023 was used in training. It “knows” this data and does not need to look ut up.
Whether they managed to further train the initial gpt4 model to do so or added something they trained separately is probably a trade secret.
Thanks!
Pretty sure that cutoff is often used because after that point ai generated content started to appear much more frequently, and the training data becomes corrupted.
For sure, this makes a lot of sense
Side note, might be useful to use poison pills on our data here, since this could be scraped in the future without consent