Social media giant Meta has laid off a scientific team that had previously developed an artificial intelligence (AI) method capable of predicting the structure of proteins, media reported Monday (Aug. 7), citing people familiar with the matter. This suggests that the company is abandoning purely scientific projects in favor of developing AI products that are easier to commercially liquidate.

Previously, Meta employed around 12 scientists on a project called ESMFold, which trained a large-scale language model capable of processing large amounts of biological data to predict protein structures, and used AI to create the first database containing over 600 million protein structures. The progress was once praised by those in medical circles involved in developing new drugs and treatments.

According to three people familiar with Meta's reorganization plans, the ESMFold team was disbanded this spring as part of the company's massive layoffs. But that had never been reported before.

Meta still employs thousands of AI scientists and engineers; the ESMFold team is small by comparison, the people familiar with the matter added. Nonetheless, the move to cancel the program suggests that Meta is looking to move away from blue-sky research (meaning engaging in basic science research without regard to the possibility of practical applications in the short term) in favor of AI projects that can generate revenue.

Yaniv Shmueli, a former research scientist and engineering manager at Meta AI who worked on ESMFold, said, "Meta is trying to restructure its research strategy to learn more about how to create advanced intelligence so that it becomes a business for Meta and not just some novelty project."

The ESMFold team

Meta established the Fundamental Artificial Intelligence Research (Fair) Lab in 2013, hiring leading scholars in the field to work on this area.

Last November, Fair researchers published a paper in Science detailing the results of ESMFold: a database of 617 million macrogenomic protein structures created by machine learning, known as the ESM Macrogenome Atlas. Macrogenomics is the study of little-known proteins from environmental samples from all over the planet, including microorganisms in the soil, ocean, and human body.

The ESMFold project first trained a large language model to learn evolutionary patterns and generate accurate structure predictions directly from the DNA sequences of proteins.

Meta also created an open-source database that allows scientists to easily retrieve specific protein structures relevant to their work, and expressed hope that the work will be able to "catalyze further scientific progress."

Meta's project is considered a competitor to DeepMind's protein folding prediction technology, AlphaFold, which was considered a scientific breakthrough for 2020 and has accuracy comparable to laboratory methods. ESMFold's language model, on the other hand, describes structures 60 times faster than AlphaFold, albeit with less accuracy.

Tim Hubbard, a professor of bioinformatics at King's College London, said large tech companies may have an advantage in deploying computing resources quickly and at scale, as well as providing computationally expensive services to scientists.

In the long run, however, the huge costs of keeping algorithmic services and databases running are an issue. meta has not confirmed whether it will continue to offer this service in the future, but for now the data will remain available to the research community. hubbard expects that academics will find a way to continue this type of work.

Full steam ahead in AI

Meta was one of the first large tech groups to invest in AI. Since establishing Fair Labs, it has published numerous papers and has been recognized by the scientific community for its advancements in AI.

However, as of now, the company has begun to lag behind competitors such as OpenAI, Microsoft, and Google, all of which have consumer-oriented generative AI (AIGC) chatbots.

The year 2023 has been dubbed the "year of efficiency" by CEO Mark Zuckerberg, and Meta has undergone a massive reorganization in recent months, restructuring its management and laying off around 20,000 employees.

Meta's new focus will be to leverage its long history of research and development in AI to create products centered around AIGC, a technology that generates human-sounding text passages, as well as images and videos.

In February of this year, Meta formed an AIGC team led by product director Chris Cox, which currently employs hundreds of people, including employees who transferred from Fair Labs, according to two people familiar with the matter. Meta is now reportedly trying to reconfigure Fair's research to match the GenAI team's goals.

Last week, it was reported that Meta plans to launch a series of chatbots with different character styles as early as September this year in a bid to catch up with its competitors.

Joelle Pineau, Meta's vice president of AI research, said, "Meta remains committed to conducting exploratory research based on open science, and the transfer of other projects from Fair's labs to our business has always been an integral part of how the team operates."