Facebook's new language model has 'high propensity to generate toxic language and reinforce harmful stereotypes'
The researchers believe the technology is not yet mature enough for commercial deployment
Facebook parent company, Meta, last week released Open Pretrained Transformer (OPT-175B), a language model with 175 billion parameters trained on publicly available datasets.
Meta stated that the goal of the release is to increase community engagement in learning about large language models (LLMs).
However, according to the company's researchers, the system suffers the same problem as its predecessors PaLM and Davinci, and is terrible at avoiding results that reinforce sexist and racist prejudices.
In a paper [pdf] accompanying the release, the researchers cautioned that the system has an even higher risk of producing toxic results than Facebook's two previous language models.
Even when provided with a relatively innocuous prompt, the model has "a high propensity to generate toxic language and reinforce harmful stereotypes", according to the researchers.
In addition, the system is also vulnerable to "adversarial prompts", where small changes in phrasing can be used to circumvent the system's safeguards and generate offensive content.
The researchers suspect it's because the training data includes unfiltered text grabbed from social media chats, which increases the model's tendency to recognise and create hate speech.
The researchers believe "this technology is premature for commercial deployment" and "more scrutiny should be afforded to the training data with additional data characterisation and selection criteria in order to use data responsibly."
LLMs, which are sophisticated programs that can create paragraphs of text and simulate human conversation, have become one of the most popular AI concepts in recent years.
However, they come with multiple issues, such as generating misinformation, bigotry and toxic language.
Google, which is exploring the use of massive language models in its search offerings, caused a stir in 2020 when it fired the head of its AI ethics team after they published a research pointing out flaws in the technology.
Meta says their OPT-175B is the first 175-billion-parameter language model available to the larger AI research community, and that it will aid academics' understanding of how LLMs work.
"We believe the entire AIcommunity would benefit from working togetherto develop guidelines for responsible LLMs, andwe hope that broad access to these types of modelswill increase the diversity of voices defining theethical considerations of such technologies," the Meta researchers' paper concludes.
According to the company, academic researchers, persons linked with government, civic society, and academic organisations, as well as corporate research facilities, will have access to the model.
Meta says OPT-175B was trained on 992 Nvidia 80GB A100 GPUs, with each chip achieving a performance of 147 TFLOPS.
Meta also claims that OPT-175B is comparable to GPT-3 but only has a one-seventh of its carbon impact.