Meta's first model from Alexandr Wang's AI team is here
Meta Superintelligence Labs just released its first model, Muse Spark, after Mark Zuckerberg waged a months-long talent war to staff the unit.
Tom Williams/CQ-Roll Call, Inc via Getty Images
- Meta Superintelligence Labs released its first AI model: Muse Spark.
- Meta says it can deploy multiple agents at once, has better health and coding abilities, and a new shopping mode.
- Meta invested $14 billion into Scale AI and waged a months-long talent war to overhaul its AI efforts and create a new team.
We finally have our first glimpse at what Mark Zuckerberg paid $14 billion for.
Zuckerberg spent months staffing up its Superintelligence Labs division after he determined Meta needed an AI overhaul. It invested billions in Scale AI to bring Alexandr Wang in-house to lead the unit and engaged in a pricey talent war, poaching employees from its competitors with eye-watering pay packages.
Now, the AI model Wang's team has been working on is here: Muse Spark.
In a blog post, Meta announced that it was the first in the Muse family of AI models developed by Meta Superintelligence Labs.
"Nine months ago we rebuilt our AI stack from scratch," Wang wrote on X. "New infrastructure, new architecture, new data pipelines. Muse Spark is the result of that work."
Meta says Muse Spark includes a "contemplating mode" to orchestrate multiple agents simultaneously, improved health responses in collaboration with 1,000 physicians, and a shopping feature that turns creator and brand content across Meta platforms into recommendations.
Meta's stock jumped 8% after the announcement. Muse Spark is available on the Meta AI site and in the app.
Nicknamed "avocado" internally, previous reports indicated that the new model would be better at coding. In its blog post, Meta referred to coding workflows as an area with "current performance gaps" that it would continue to invest in.
Meta says that it conducted "extensive safety evaluations" on Muse Spark. The model shows "strong refusal behavior" for questions about high-risk subjects like chemical weaponry.
The company also conducted a third-party evaluation with Apollo Research, which found that it had "the highest rate of evaluation awareness of models they have observed," per the blog post.
Wang wrote that his team has more models on the way, some of which will be open-source. "This is step one," he wrote.
Zuckerberg posted on Threads that the model was the "first milestone" toward giving everyone personal superintelligence.
"We are building products that don't just answer your questions but act as agents that do things for you," Zuckerberg wrote. "I am optimistic that this will support a wave of creativity, entrepreneurship, growth, and health."
Read the original article on Business Insider