- Real-world impact is more important than benchmark scores, Meta AI VP says
- Song emphasizes security, trust and real-world benefits
- Meta's latest model prioritizes people, she says
Meta's newly appointed AI research chief, Dawn Song, is betting on agentic AI for the future of artificial intelligence, emphasizing they should augment humans rather than replace people altogether.
Song sees agents performing "economically valuable" tasks like repetitive and time-consuming work, ultimately freeing up humans to do more creative work.
While companies struggle to quantify AI's impacts and deliver a meaningful ROI, Song believes the focus should be real-world impact rather than benchmark scores.
AI agents should augment – not replace – humans
"The goal is not to replace humans," Song told the South China Morning Post. She confirmed she would be joining Meta Superintelligence Labs in a LinkedIn post, together with other team members from Virtue AI.
"[AI] must be secure, trustworthy, and beneficial," she added.
Song is also a professor in computer science at the University of California, Berkeley – a university that recently introduced Agents' Last Exam (ALE), a new type of benchmark that assesses whether AI agents can complete more than 1,500 economically valuable tasks across 55 different industries.
Meta itself launched its first new model, Muse, in April, which it says is designed to "prioritize people."
As MSL's Vice President of AI Research, Song will focus on AI safety, security and research, and will likely continue to emphasize the role of humans in an AI-first era.
Model capabilities are no longer a drawback for AI developers, with human, socioeconomic and geopolitical impacts now emerging as a major focus. Anthropic was recently forced by the White House to pull its latest frontier models over jailbreaking concerns.