“At first, we were thinking of this in a very long-term way,” says generative AI software engineer Emily D. “And then, suddenly, the dam broke.”

From a conference room in New York City, Emily recalls how just a few years ago, she was a research engineer in the Meta Fundamental AI Research organization (FAIR), where one of their primary goals was publishing papers for academic communities.

“I never imagined that just a few years later, our work would be touching billions of people through Meta's technologies. It's pretty amazing,” Emily says.

Now a tech lead for post-training capabilities, Emily gains a lot of insights from the product teams at Meta about what people care about, and where her team should be investing more in.

This culture of sharing and openness is by design.

“From my position in generative AI, exploratory research teams and product teams sit within the same org,” Emily explains. “We get this very cool connection between them, where research work flows into product, and product learnings can flow back into the research work and inform the kinds of things we work on.”

Mike L., a research scientist on the generative AI team, is solving a variety of scientific challenges, like how to build the next generation of language models across a multitude of areas. Similar to Emily, he also began his career at Meta on the Meta FAIR team and is enthusiastic about sharing what he learns. “We still publish our work in research papers when we have big, new models. It’s really nice to be able to tell everyone about them,” says Mike.

When senior research manager Melanie K. reflects on what it took to launch Meta Llama 3.1, she feels a lot of pride. Melanie leads the Llama text post-training research team, and is quick to mention the incredible amount of drive and discipline her team had.

“So much hard work went into this. For example, we put a tremendous amount of work into data cleaning and making sure that data quality was really high. I was really proud of the team for working really, really hard to get the overall results done on time at each training round,” says Melanie.

On the product side, Joe S., director of product management on the generative AI team, saw an explosion of creativity among startups and small businesses with the rollout of Llama 3 earlier this year, which makes the potential of this latest version of Llama so exciting.

“Using Llama as a foundational technology, people were building generative AI applications, chatbots and agents, and search engines,” Joe says. “They were extending the model using long context length, running on embedded devices and customizing state-of-the-art plugins.”

Seeing all these use cases answered a question Joe and his team had: When startups had access to a truly open model, what would they actually do with it? “They made Llama their lifeblood,” he says.

All this growth and possibility brings to mind a new twist on an old adage: With great potential comes great responsibility. “When we create a model or product that’s helpful, we also need to ensure that it’s safe,” Emily says. In other words, trust and safety is a prerequisite.

Research scientist Aston Z. says that open sourcing Llama responsibly and sharing findings in research papers shows the generative AI team’s commitment to transparency and collaboration. “When we design and evaluate methods, we try to reduce all sorts of biases,” he explains. “We emphasize safety measures, such as monitoring violation rate and false refusal rate. This ensures that our models are not only effective but also safe for people to interact with.”

These kinds of safeguards are standard at Meta, baked into everything AI teams touch. There’s also an extra layer of work done by the AI trust and safety team, helmed by director of product Vincent G.

“We have this effort of open sourcing tools like Meta Llama Guard 3,” Vincent says, “and as we open source those tools, we’re looking forward to the community improving on them for the benefit of all, and leveling the playing field on safety."

From Vincent’s perspective, it’s an exciting time to be working in AI at Meta, on teams where open source, responsibility and safety all converge into one.

“If people want to work at the cutting edge of open sourcing AI technology and making it available to the world,” says Vincent, “Meta is the company to do it at.”

Mike agrees. “I think Llama 3.1 is really exciting and really impactful. And there's no opportunity to do this anywhere else in the world,” he says. “That's why I'm here.”