Fireside Chat: Meta Superintelligence Labs + Vercel

Highlights

• Meta Superintelligence Labs ships 11 models and products starting from April through an organizational framework centered entirely on velocity.

• The Meta AI codebase scales from four engineers to ninety-seven contributors writing over two million lines of code in ninety-five days.

• Bring Your Own Compute enables running functions inside private AWS accounts while preserving Vercel deployment speeds and routing experiences.

• Muse Wonder integrates active memory and active focus to retain context over minutes, hours, and days.

• Muse Spark Open runs locally on Mac Studio hardware following a direct collaboration with Apple.

Timeline

1. 00:00 - 05:05

   Rebuilding Momentum Through Relentless Velocity

   Moving past initial skepticism in the AI frontier race requires removing internal friction and discarding legacy tech stacks. Operating with complete autonomy allows engineering teams to ship 11 distinct models and products starting from April. Distributing intelligence through products like Muse ensures personal agents reach broad user bases globally rather than remaining restricted to select entities.

   • Meta Superintelligence Labs achieves rapid progress through an absolute framework prioritizing velocity above all else.
   • Team structures remain intentionally small, pairing autonomous, high-agency engineers with streamlined infrastructure choices.
   • Open-weight models like Glimmer and Spark Open reflect a mission to distribute superintelligence broadly rather than centralize it.
2. 05:05 - 07:10

   Advanced Agent Capabilities and Local Execution

   Long-running agent workflows depend heavily on sustained context retention and persistent execution toward target outcomes. Isolating work trees prevents single agents from clogging development environments during complex coding tasks. Expanding availability beyond cloud infrastructure allows users to run powerful open models directly on local hardware.

   • Muse Wonder incorporates active memory and active focus to maintain context over hours and days of agentic work.
   • Muse Code utilizes isolated work trees to execute parallel tasks without bottlenecking terminal operations.
   • Collaboration with Apple enables Muse Spark Open to run locally on Mac Studio hardware.
3. 07:10 - 14:00

   Scaling Rapid Development with Vercel

   Transitioning to Vercel starts as an organic engineering experiment before becoming a core component of production deployments. Fast deploys and safe rollbacks encourage frequent experimentation and allow designers to ship code directly. Enterprise privacy requirements necessitate co-developing custom solutions to maintain impenetrable security barriers around internal data and infrastructure.

   • Initial trials with Vercel replace six-hour iteration cycles with rapid deployments and instant rollbacks.
   • A team of four engineers scales to ninety-seven contributors writing over two million lines of code in ninety-five days.
   • Strict security constraints require running code inside internal AWS accounts rather than external shared infrastructure.
4. 14:00 - 17:57

   Co-Developing Enterprise Infrastructure and Agent-First Codebases

   Collaborative design partnerships solve strict enterprise compliance walls by keeping user data secure under company accounts. Modern software development shifts toward agent-centric architectures capable of frequent, automated rewrites. Future infrastructure evolution must accommodate multi-hour stateful workloads rather than sticking solely to traditional web request-response models.

   • Bring Your Own Compute allows running functions inside internal AWS accounts while retaining Vercel routing and developer experience.
   • Codebases of the future require optimization for agents rather than human developers as primary demographic builders.
   • Long-running, stateful agent operations expose limitations in traditional request-response web infrastructure.