Kilcher emphasized the result that GPT-4chan slightly outperformed other existing language models on the TruthfulQA Benchmark, which involves picking the most truthful answer to a multiple choice question. An evaluation of the model on the Language Model Evaluation Harness.The model was released on Hugging Face, a hub for sharing trained AI models, along with the ‘playground’ feature allowing users to interact with it. An already trained instance of the model.
#Type o negative racist code#
The bots collectively wrote over 30,000 posts over the span of a few days, with 15,000 being posted over a span of 24 hours. The video also contains the following: a brief set of disclaimers, some discussion of bots on the internet, a high level explanation of how the model was developed, some other thoughts on how good the model is, and a description of how a number of bots powered by the model were deployed to post on the /pol/ message board anonymously.
The model thus learned to output all sorts of hate speech, leading Yannic to call it “The most horrible model on the internet” and to say this in his video: “The model was good, in a terrible sense … It perfectly encapsulated the mix of offensiveness, nihilism, trolling, and deep distrust of any information whatsoever that permeates most posts on /pol/.”
In this case, the model was made by fine-tuning GPT-J with a previously published dataset to mimic the users of 4chan’s /pol/ anonymous message board many of these users frequently express racist, white supremacist, antisemitic, anti-Muslim, misogynist, and anti-LGBT views. GPT-4chan is a large language model, and so is essentially trained to ‘autocomplete’ text - given some text as input, it predicts what text is likely to follow - by being optimized to mimic typical patterns of text in a bunch of files. On June 3rd of 2022, YouTuber and AI researcher Yannic Kilcher released a video about how he developed an AI model named ‘GPT-4chan’, and then deployed bots to pose as humans on the message board 4chan. His videos explaining AI papers are very educational, and I encourage you to check out his YouTube channel if you are not aware of them.
To be clear, this article presents criticisms of Yannic Kilcher's actions with respect to GPT-4chan specifically, and does not present a criticism or condemnation of him as a whole.This led to renewed discussions not covered in this piece. Update: On Ja statement titled " Condemning the deployment of GPT-4chan" was circulated by Percy Liang and Rob Reich, and was signed by hundreds of AI researchers and developers.If you are already aware of what happened, I recommend skipping the first two sections, but to still read the ‘Analysis’ and ‘Lessons’ sections.It is primarily for people in the AI community, but is accessible to those outside of it as well. As with my article on an older controversy related to AI, the intent of this is to provide a comprehensive summary of what happened, as well as what I consider to be valuable lessons that can be taken away from it all.This article contains an objective summary of a recent controversy related to an AI model named GPT-4chan, as well as a subjective commentary with my thoughts on it.
0 kommentar(er)
