Databricks VP: Don’t Try to Speed AI Evolution through Brute Force

Key Concepts:

• AI Agents
• Deep Learning
• Neural Networks
• Large Language Models (LLMs)
• Hardware for AI (GPUs, specialized chips)
• Computational Efficiency
• Biomimicry
• Feature Engineering
• Reinforcement Learning
• Model Collapse
• Data Governance
• Medallion Architecture (Bronze, Silver, Gold data)
• Agentic AI
• Iterative Self-Improvement
• Neuromorphic Computing
• Out-of-Distribution vs. In-Distribution Data

Early Computing Experiences and Background

• Naveen Rao's introduction to computing began with a Texas Instruments 994A in 1978, influenced by his gadget-loving father and older brother.
• He learned to code in Logo and Basic at a young age (8-9 years old), highlighting the esoteric nature of computing at the time.
• His interest in building things led him to create circuits and logic systems, such as a buzzer system for quiz bowl, using components from Radio Shack.

Transition to AI and Deep Learning

• Rao's fascination with building intelligent machines led him to study AI as an undergraduate in the mid-1990s.
• He observed the shift from creative algorithms to regression methods, with neural networks initially gaining traction but later being overshadowed by Support Vector Machines (SVMs) due to ease of training and lower data requirements.
• The resurgence of neural networks in the mid-2000s was driven by increased data availability, which saturated the performance of regression methods.
• Neural networks' ability to automatically discover features from data, rather than relying on manual feature engineering, marked a significant shift.
• The ImageNet competition in 2012 served as an inflection point, demonstrating the superior performance of deep learning compared to regression methods.

Hardware Limitations and the Brain as Inspiration

• Rao emphasizes the brain's energy efficiency (20 watts) compared to the power consumption of modern data centers (gigawatts).
• He criticizes the industry's reliance on brute-force scaling of existing computer architectures (ALU, memory interface, caches) rather than exploring fundamentally more efficient designs.
• He points out that current GPUs consume enormous amounts of power (thousands of watts) and that this approach may not be sustainable or optimal for achieving true intelligence.
• He suggests that biology, constrained by energy, offers valuable insights into efficient computation.

The Plateau of Current AI Models

• Rao argues that current AI models are not improving as significantly as they once were, despite becoming more useful and economically valuable.
• He believes that these models are limited by their reliance on observational data and their lack of a true understanding of reality.
• He mentions that LLMs create a "latent representation" of the world based on observed data, which is inherently incomplete.

Iterative Self-Improvement and the Role of Reinforcement Learning

• Rao acknowledges the progress in iterative self-improvement of AI systems through automated reinforcement learning, citing his team's work on a system called "Towel."
• He draws a parallel between the shift from hand-tuned features to learned features and the current focus on learning the judgment of whether an output is good.
• He suggests that while these advancements will make AI systems better, it's still uncertain whether they will lead to superintelligence.

Nirvana: An Early AI Chip Company

• Rao founded Nirvana, an AI chip company, to build processors specifically designed for neural network computation.
• The goal was to create chips that could perform lower-precision matrix multiplication efficiently and scale effectively.
• Nirvana's first chip was a highly distributed processor with a mesh architecture, designed for scalability.

Rebuilding the Physical Substrate of Chips

• Rao criticizes the industry's continued reliance on zeros and ones, synchronous design, and finite bit-width representations of numbers.
• He argues that there is a need for more flexible ways to represent information, potentially trading off time for accuracy for power.
• He suggests exploring theoretical limits of computation, such as the Landau limit, to develop more efficient hardware.

AI's Impact on Software Development at Data Bricks

• Data Bricks is using AI co-pilots and working with model providers like Anthropic to enhance software development.
• AI tools are particularly effective at automating repetitive tasks, such as generating templates, headers, and project structures.
• Rao sees potential for AI to accelerate chip design and verification cycles.
• However, he believes that AI is unlikely to generate truly novel ideas or replace the need for human creativity in areas like designing new devices or physics.

The Economics of Large Language Models

• Data Bricks initially focused on making LLMs cheaper to train and serve, developing a mixture of experts architecture.
• The company later shifted its strategy to partnering with Meta and focusing on reinforcement learning at scale with customer data.

The Importance of the System Side of AI

• Rao emphasizes the importance of the system side of AI, which involves making AI models run efficiently in a reasonable amount of time.
• He notes that AI hardware is intrinsically tied to algorithmic development, with each iteration of deep learning reflecting the architecture of the hardware it was built on.

Copy and Paste as a Metaphor for AI's Impact

• Rao discusses how copy and paste exemplifies how AI is changing the work experience.
• He highlights the concept of agentic AI and how technologies like Neon are changing the way people think about relational databases.

AI's Impact on User Interfaces and Software Development

• Rao notes that AI is changing the mental model of how software works, with systems becoming more dynamic and capable of learning from usage.
• He envisions a future where the feedback loop between user behavior and software updates is tightened and automated.
• He suggests that AI can be thought of as a context that is continually evolving, rather than a static release.

The Future of Neural Interfaces

• Rao, who worked on neural interfaces during his PhD, acknowledges the challenges of interfacing with biology.
• He believes that while neural interfaces may have applications for injured individuals, it will take time to develop consumer-grade devices.

Developing for AI Agents

• Rao emphasizes the importance of monitoring AI agents to ensure they are doing what they are intended to do.
• He suggests that humans will move into more of an oversight function, using deterministic software to verify the behavior of AI agents.
• He notes that there are differences between building for consumption by an agent versus consumption by a human.

Preventing AI from Going Off the Rails

• Rao believes that AI can be prevented from going off the rails by implementing clear governance structures and permissioning tools.
• He emphasizes the need to explicitly define what AI models can and cannot do and to track the lineage of their actions.

Model Collapse and the Importance of Grounded Data

• Rao warns of the potential for model collapse, where LLMs train on machine-produced content rather than real-world data.
• He emphasizes the importance of grounding AI models in reality and leveraging curated data sources, such as the medallion architecture (bronze, silver, gold data).

Personal Use of AI Agents

• Rao uses coding agents and tools to clean up writing.
• He also benefits from AI features in email, such as suggestions for including recipients or documents.

Synthesis/Conclusion:

Naveen Rao provides a comprehensive perspective on the current state and future trajectory of AI, emphasizing the limitations of current hardware approaches and the need for more efficient and biologically inspired designs. He highlights the importance of data governance, the potential for model collapse, and the evolving role of humans in overseeing AI systems. He also discusses the impact of AI on software development and the need to develop new mental models for how software works. His insights offer a valuable framework for understanding the challenges and opportunities in the field of AI.