Dawn Song joins Meta as Vice-President of AI Research at Superintelligence Labs. Her message is sharp: the next step is not a new foundation model, but agents that can deliver real economic value. The hire reshapes Mark Zuckerberg’s AI strategy.
Key Takeaways
- Dawn Song is leaving her UC Berkeley chair and Virtue AI to become VP AI Research at Meta Superintelligence Labs.
- In her view, AI agents are the next frontier, but only if they produce measurable economic value.
- Part of her Virtue AI team is joining her at Meta.
A heavyweight hire
Dawn Song is a computer science professor at UC Berkeley and co-director of the Center for Responsible, Decentralized Intelligence (RDI). She also co-founded Virtue AI, an enterprise AI safety startup. Her move to Meta was announced in late June 2026.
The hire comes right after her remarks at the World Economic Forum in Dalian, also known as Summer Davos. Dawn Song laid out her view of AI agents there, and the technical community read her words as a signal about what Meta wants to do next.
Several Virtue AI team members are moving with her. Meta is not just acquiring a researcher, but a built-out team, with agentic safety expertise that the Superintelligence Labs had been missing.
The move sits inside Meta’s broader AI rebuild after the Scale AI deal and the appointment of Alexandr Wang as Chief AI Officer. Superintelligence Labs is actively recruiting top academic profiles to compete with OpenAI, Anthropic, and Google DeepMind.
The internal budget layer fits. Andrew Bosworth imposed a strict cap on Meta engineers’ AI consumption back in May. The logic: spend less on internal compute, invest more in the talent that can build agents that actually move the needle.
The Dawn Song thesis: economic value first
Her Dalian message is precise. Per Dawn Song, the goal is not to replace humans. The goal is for AI agents to become more effective in the real-world domains that produce genuine economic value. The phrase is calibrated to reassure as much as to announce.
The yardstick already exists. UC Berkeley’s RDI Centre rolled out a benchmark in June called Agents’ Last Exam, or ALE. The test grades agents on more than 1,500 real-world economically valuable tasks across 55 industries.
The gap with current discourse matters. AI labs still measure progress on exam-style benchmarks or logic puzzles. Dawn Song moves the yardstick toward what companies actually pay for, sector by sector. That is the difference between an agent that answers a question and an agent that closes a case.
This grid also reshapes Meta’s product strategy. For an agent to generate revenue, it has to finish a task end to end: find the right document, execute the action, verify the result, and move on. Independent benchmarks show agents complete only 14% of WebArena tasks today, leaving an enormous gap to close.
The pivot Dawn Song lays out is as methodological as it is organizational. For Meta, that means teams working alongside pilot customers rather than closed labs maximizing a benchmark score. The line between research and product is fading.
Also on Horizon:
- Tencent Study: AI Agents Finish 14% of Real Tasks
- Coinbase Picks Chinese AI, AI Bill Drops 50%
- Anthropic Survey: AI Does Half Your Work
What changes short and medium term
In the short term, Dawn Song’s arrival rewires Meta’s research stack. Superintelligence Labs gains an agentic safety expert and a strong academic anchor. Analysts expect product announcements in the coming weeks focused on B2B agents, possibly around Hatch or a new enterprise tier.
Competitors feel the move. OpenAI already shipped ChatGPT Agent for consumers but lacks a mature enterprise offer with the same governance depth. Anthropic relies on Claude Tag and Mythos for business workflows, without an economic benchmark this explicit. The ALE grid could become a real reference for corporate buyers.
In the medium term, over three to six months, the question is enterprise budget allocation. CIOs want measurable return on investment. An agent that closes 14% of tasks no longer justifies a multi-million dollar contract. Meta promises to push that number toward 50% through Dawn Song’s team. The road stays bumpy, Zuckerberg admitting AI agents are slower than expected.
Safety remains central. The more autonomy an agent gets, the higher the cost of error. Dawn Song’s Virtue AI background gives her unusual credibility on that front. Meta can lean on her profile to engage European regulators finalizing AI agent guidelines this summer.
One thing to watch. Dawn Song has not announced a specific Meta model. Superintelligence Labs continues to work on foundation and agents in parallel. The bet is that task orchestration becomes the differentiator, not the parameter race. If she pulls it off, Meta changes category.
Follow the story on Horizon.
