Artificial Intelligence: Meta has built a massive new AI language – and is giving it away for free

“It’s a great initiative,” says Thomas Wolf, chief scientist at Hugging Face, the AI ​​startup behind BigScience, a project in which more than 1,000 volunteers from around the world collaborate on a language model. in free access. “The more open models there are, the better,” he adds.

Big language models — powerful programs that can generate paragraphs of text and mimic human conversation — have become one of the hottest trends in AI over the past couple of years. But they have deep flaws and convey misinformation, bias and toxic language.

In theory, increasing the number of people working on this problem should help. Yet, because learning language models requires large amounts of data and computing power, they have so far remained projects for wealthy tech companies. The wider research community, including ethicists and social scientists who worry about their misuse, have had to stay on the sidelines.

Meta AI says it wants to change that. “A lot of us have been university researchers,” says Pineau. “We know the gap that exists between academia and industry in terms of the ability to build these models. Making it available to researchers was obvious. She hopes other people will look at her work, take it apart or be inspired by it. Breakthroughs come faster when more people are involved, she says.

Meta puts its model, called Open Pretrained Transformer (OPT), available for non-commercial use. It also publishes its code and a logbook that documents the training process. The logbook contains daily updates from team members regarding training data: how and when it was added to the model, what worked and what didn’t. In over 100 pages of notes, researchers log all bugs, crashes, and reboots that occurred during a three-month training process that ran non-stop from October 2021 to January 2022.

With 175 billion parameters (the values ​​of a neural network that are changed during training), the TPO is the same size as the TPG-3. It’s on purpose, explains Mr. Pineau. The team designed OPT to be as accurate as GPT-3 in language tasks and to be as toxic. OpenAI made GPT-3 available as a paid service, but did not share the model itself or its code. The idea was to provide researchers with a similar language model to study, says Pineau.

Google, which is exploring the use of large language models in its search products, has also been criticized for its lack of transparency. The company sparked controversy in 2020 when it ousted prominent members of its AI ethics team after they produced a study highlighting issues with the technology.

Culture shock

So why is Meta doing this? After all, Meta is a company that has said little about how the algorithms behind Facebook and Instagram work and has a reputation for burying unfavorable results from its own in-house research teams. See also: A roboticist explains why she loves working at Boston Dynamics. Meta AI’s different approach is largely explained by Ms. Pineau herself, who has been advocating for greater transparency in AI for several years.

Ms Pineau has helped change the way research is published at many of the biggest conferences, introducing a checklist of what researchers must submit with their results, including code and details on how experiments are carried out. Since joining Meta (then Facebook) in 2017, she has championed this culture in his AI lab.

“This commitment to open science is why I’m here,” she says. “I wouldn’t be here otherwise. »

Ultimately, Pineau wants to change the way we judge AI. “What we now call the state of the art cannot be limited to performance,” she says. “It has to be the state of the art in terms of liability as well”.

Nonetheless, giving away a significant language model is a bold move for Meta. “I can’t tell you there’s no risk that this model will produce language that we’re not proud of,” says Pineau. ” He will do it. »

Weigh the risks

Margaret Mitchell, one of the AI ​​ethics researchers whom Google forced out in 2020, and who now works for Hugging Face, sees the release of the OPT as a positive step. But she thinks there are limits to transparency. See the article: Smartphone: Samsung formalizes the launch of the Galaxy A series. Has the language model been tested with sufficient rigor? Do the foreseeable benefits outweigh the foreseeable disadvantages, such as the production of false information or racist and misogynistic language?

“The dissemination of a great linguistic model in the world, where a large audience is likely to use it or be affected by its results, involves responsibilities,” she explains. Ms. Mitchell notes that this model will be able to generate harmful content not only on its own, but also through downstream applications that researchers will build from it.

Meta AI has verified OPT to remove some harmful behaviors, but the goal is to release a model that researchers can learn from, with all of its flaws, Pineau says.

“We’ve had many discussions about how to do this in a way that allows us to sleep at night, knowing there’s non-zero risk in terms of reputation, non-zero risk in terms of harm,” she says. She rejects the idea that you shouldn’t release a model because it’s too dangerous, which is OpenAI’s reason for not releasing GPT-3’s predecessor, GPT-2. “I understand the weaknesses of these models, but it’s not a research mindset,” she says.

Ms Bender, who co-authored with Mr Mitchell the study at the center of the dispute with Google, is also concerned about how potential harms will be handled. “One thing that’s really key to mitigating the risks of any type of machine learning technology is to base assessments and explorations on specific use cases,” she says. “What will the system be used for? Who will use it, and how will the results of the system be presented to them? »

Some researchers wonder why large language models are built, given their potential for harm. For Ms. Pineau, these concerns must be addressed with more transparency, not less. “I believe the only way to build trust is extreme transparency,” she says.

“We have different opinions around the world about appropriate speech, and AI is part of that conversation,” she says. She doesn’t expect language models to say things that everyone agrees with. “But how to deal with this situation? You need many voices in this discussion. »

We would love to say thanks to the writer of this short article for this awesome content

Artificial Intelligence: Meta has built a massive new AI language – and is giving it away for free

We have our social media profiles here as well as additional related pages here.