Meta runs one of the most sophisticated content and ad targeting systems ever built. Every time someone opens Instagram or Facebook, an algorithm decides what they see in milliseconds. The system, known internally as Rankings and Recommendations, is the invisible driver of the company’s more than $160 billion in annual revenue. It also hasn’t yet been meaningfully powered by LLM, the same class of technology behind ChatGPT, Claude, and Gemini, according to Li.

“We’re still not using LLM architecture for ranking and recommendation work,” Lee said at the event. She added that LLMs are currently “not a big part of core rankings or recommendations.”

The hope is that LLM will eventually do so someday, as it not only optimizes what is already working, but can also infer content and context in ways that are fundamentally impossible with current systems. Today’s ranking engines are built on engagement signals like shares and watch time, and these signals require scale, and scale requires time. This is a feedback loop that works very well at the meta scale, but it has an upper limit. Because you can only optimize for what the user has already responded to, you’re inherently backward-looking and blind to content and context they’ve never encountered before.

LLM bypasses this completely and can infer in real time whether content is likely to be of interest to a particular user based on what the system already knows about the user, without having to learn from engagement history first. That’s a feature Meta wants to bring to its core product, because it could make today’s already formidable advertising business seem primitive.

“This is the result of a bit of a long-term research effort,” Lee said. “We don’t know exactly what that will look like, but we think it’s worth the investment.”

Until then, LLM exists in the meta in two relatively limited ways. The first is understanding the content. Run posts, videos, and other materials through language models to more accurately predict whether content will be interesting to a specific user without waiting for engagement signals to accumulate. The second concerns Threads, Meta’s text-based X rival, which Li says is “a little ahead of the curve” in applying an LLM-based approach to rankings. This is a natural fit given that language models are built for text.

“Traditional recommendation engines rely heavily on engagement signals, and it takes a lot of engagement to get engagement signals,” Li said. “LLM can determine in real time whether this is interesting content for you.”

infrastructure issues

The real reason for the delay is computing. Building and deploying LLM at the scale that Meta operates at (billions of users across Facebook, Instagram, WhatsApp, and Threads) requires infrastructure that doesn’t yet exist in sufficient quantities.

Mr Lee acknowledged that what the company thought it had sufficient capacity even 24 months ago has turned out to be largely wrong. Data centers have long lead times, and many of the ones Meta is building now won’t be operational until 2027 or later. In the meantime, the company has resorted to creative workarounds. Li said Meta has introduced industrial-grade tents that are rated to last 25 years and withstand tornadoes to house servers and get online capacity faster.

An even more difficult problem, Lee suggested, is predicting what will happen next. Training a model is a one-time cost. Running continuously in real time and at a scale of billions of daily users is continuous and far less predictable.

“I’m worried that we’ll underestimate it,” she continued.

When Meta integrates something new into its family of apps, it instantly reaches a user base that most technology companies would never reach. Infrastructure timelines cannot accommodate such demands.

What can you do with your current system and why is it important?

None of this is to say that Meta’s existing ranking system is static. If anything, Lee’s description of the current state of play was a reminder of how much leeway is left in the pre-LLM era.

Meta tracks ad performance through an internal metric called irev, and every six months the team arrives with a new list of experiments. Each team is incrementally pushing those numbers up, building on previous accomplishments. Lee called it “perhaps one of the wonders of the modern world.” In the fourth quarter alone, we saw a 7% increase in organic content views due to improved product rankings on Facebook. This was the company’s most revenue-impacting launch in the last two years. Each benefit improves ad performance, lowers costs for advertisers, puts more budget into the platform, and funds the next cycle of improvements.

It is this flywheel, not the LLM, that currently generates Meta’s profits. And the revenue from that flywheel will one day fund the research, human resources, and data centers that will replace it.