Presenters
Source
The Agentic Revolution: How AI is Reshaping Our Digital Universe 🚀
The world of technology is moving at a breakneck pace, and this year’s At Scale Conference is a testament to that. We’re not just talking about building systems at planet scale anymore; we’re witnessing a true revolution fueled by AI, transforming what was unimaginable just a year ago into our daily reality. Get ready to dive into the astonishing evolution of AI and data, and how it’s paving the way for unprecedented innovation! ✨
The Insatiable Appetite for Compute 🧠
AI has unleashed a crazy appetite for compute, and it shows no signs of slowing down. Data centers are being built at an unprecedented rate, and the demand for skilled talent is skyrocketing. Meta, for instance, has invested $115 million in the America’s Workforce Academy to train the next generation of AI professionals.
Across the industry, we’re seeing record numbers:
- AI-generated code is hitting new highs.
- Agentic token consumption is growing orders of magnitude quarter over quarter.
But the impact isn’t confined to coding. With 3.5 billion people engaging with Meta’s products daily, the complexity of supporting these apps at scale – encompassing machines, data centers, energy, hardware, software, product integrity, privacy, and the incredible people behind it all – is immense.
The Agentic Revolution: A New Paradigm 🤖
The agentic revolution is fundamentally changing how we operate. It’s accelerating some things, upending others, and creating fertile ground for new innovation.
Let’s consider our core product loop, exemplified by an Instagram user. When a user engages with the product, a complex system works behind the scenes. This loop involves:
- Core Product: The user interfaces they interact with.
- Data Generation: User interactions generate vast amounts of data.
- Recommender Systems: Powered by large language, vision, and speech models trained on massive datasets.
- Business & Product Decisions: These, along with recommender systems, shape the user’s experience.
Before agents, all these functions were meticulously designed, built, operated, and improved by human talent. Now, agents are not only coding proficiently – capable of one-shot big chunks of code – but they’re also excelling at:
- Managing data 📊
- Analyzing reports 📈
- Conducting ML experimentation 🧪
- Building other models 🏗️
The next frontier? Expanding the autonomy we grant our agents. With proper guardrails, they can soon perform work autonomously for hours, days, or even weeks. Imagine a highly intelligent and driven employee who makes decisions based on provided context, with humans acting as guides and supervisors. This is an entirely new and exciting paradigm, shifting humans from being in the loop to over the loop.
Agents as Infrastructure Citizens: A Shifting Landscape 🌐
The impact of agents is profound, even transforming our infrastructure.
- Automated traffic now surpasses human traffic on the internet, growing eight times faster.
- Within our own infrastructure, agents are becoming first-class infrastructure consumers, a stark contrast to the 20 years we spent optimizing for human request patterns (predictable, session-based, with natural rate limits).
This shift breaks several key elements simultaneously:
- Capacity Break: One engineer used to mean one unit of load. Now, one engineer can spawn tens or even hundreds of agents, leading to a potential load increase of 100,000 users overnight for a thousand-person organization.
- Identity Break: Agents aren’t human users, service accounts, or batch jobs. They make decisions, presenting a significant challenge for identity management.
- Velocity Break: While GitHub Copilot now writes 46% of code for its users, our CI/CD pipelines haven’t sped up proportionally. Agents write code in seconds, but the subsequent build, test, deploy, and monitor processes remain bottlenecks.
The crucial question for this conference: Can your infrastructure handle what happens when agents do the work?
The Data Transformation: Fueling Intelligent Decisions 💡
Data is at the heart of every business decision, product, and AI model. The way we interact with data is undergoing a dramatic transformation.
- Since February, we’ve seen the launch of agentic data apps, and in just 3 months, 63% of dashboards were created using them.
- We’ve also witnessed a 30x growth in agentic queries.
This democratizes data access and empowers data-driven decision-making. However, it also raises critical questions about trust and governance:
- What happens when the human experts who curated and analyzed data, understanding its privacy and regulatory aspects, are no longer in the loop?
- How do we trust insights generated by agents or humans with limited or no context, especially regarding regulations and expected guidance?
- How do we backtrack if something goes wrong?
Navigating data is becoming even more complex, but solutions are emerging:
- Agentic data apps enable non-experts and agents to build custom data products.
- Agent governance, safety, and privacy are not optional but prerequisites for scalable AI.
- Observability, Evals, and benchmarks are crucial for understanding data quality when agents make decisions.
The Reasoning Engine: Beyond Pattern Matching 🧠
The sheer volume and complexity of data are exploding.
- The ratio of training data to active parameters in open-weight LLMs has grown 3.1 times per year since 2022.
- Recent models use 20 times more data per parameter than optimal ratios suggested by 2022 scaling laws.
This hunger for data extends to critical business solutions. Reasoning, unlike pattern matching, requires a full view of user behavior across surfaces over time, not just summaries.
At Meta, we’re building towards:
- 500 million queries per second.
- A petabyte per second of throughput for training data reads.
This necessitates a complete rethinking of our data infrastructure:
- Freshness is paramount: Real-time streaming data is becoming the backbone, replacing traditional batch ETL for ranking and recommendation pipelines.
- Intelligent storage: Moving away from opaque blobs to storage that understands its data, fetching only what’s needed in flexible formats.
Reimagining Recommendations: From Pattern Matching to True Reasoning 💬
Recommender systems, the heart of many businesses and a driver of double-digit revenue growth, are also evolving. For two decades, they’ve relied on observing behavior and finding patterns. LLMs break this paradigm, enabling recommender systems that reason about intent, not just correlate signals.
- 42% of Instagram users want to fundamentally change what they see, not just adjust settings.
This calls for a new approach:
- Imagine a recommender system that you can direct conversationally, telling it what you want more or less of, and it reasons about your intent in real-time.
- This means understanding the difference between a casual fan looking for highlights and a club athlete seeking training drills, all from the same query (“soccer”).
The Flywheel of Innovation: Agents, Data, and Reasoning 🔄
The transformation of the product loop is profound:
- Agents are no longer just building the loop; they’re citizens within it, operating, discovering, analyzing, and creating alongside us.
- Data is shifting from batch to real-time, from opaque blobs to intelligent storage, driven by the demands of agents and reasoning models.
- Recommendations are evolving from pattern matching to genuine reasoning, from an algorithm that happens to you to one you direct.
These elements form a powerful flywheel:
- Agents make data more accessible.
- Better data enables sophisticated reasoning.
- Reasoning creates new demands that push agents and data infrastructure forward.
This isn’t just linear progress; it’s a self-accelerating cycle.
Beyond the Product Loop: Agents in Action at Meta 🚀
The impact of agents extends far beyond our core product loop. Innovation is blooming across Meta:
- Task Agents: Focused on domain-specific activities like coding, data analysis, and operations.
- Clowntown: A multi-agent orchestration framework enabling autonomy, allowing engineers to go on vacation while agents make semi-autonomous decisions.
- Malbook: Exploring an agent-to-agent social network where humans observe agent communication and negotiation.
- Frontier Loop: Accelerating executive decision-making by autonomously synthesizing high-context documents and updates.
The Future is Now: Co-Creating at Scale 🤝
A decade ago, At Scale was about building systems for a billion people. Today, we’re exploring what happens when those systems serve billions of people and millions of superintelligent agents simultaneously.
Imagine models that train themselves, continuously tuning and adapting without waiting for the next training cycle. When we get this right, when agents and humans co-create at scale, we’re not talking about incremental improvements; we’re talking about entirely new categories of products that don’t exist yet.
The insights shared today are not academic exercises; they are glimpses into the future being built now. The infrastructure you build this year will power innovations we can only dream of today.
So, the question remains: What will scale infrastructure enable you to build? Let’s dive in!