The New Stack Agents, Naveen Rao, Databricks

Key Concepts

• AI Agents
• Deep Learning
• Large Language Models (LLMs)
• Hardware Architecture for AI
• Computational Neurobiology
• Biomimicry
• Neuromorphic Computing
• Data Governance
• Model Collapse
• Reinforcement Learning
• Feature Engineering
• Data-centric AI
• Software Development with AI
• AI Infrastructure
• Economic Viability of AI
• The Medallion Architecture

Early Computing Experiences and the Spark of Curiosity

Naveen Rao's introduction to computing began with a Texas Instruments 994A in 1978, influenced by his gadget-enthusiast father and older brother. He learned to code in Logo, using the "turtle" to draw and create games, followed by basic programming at around age eight or nine. This early exposure sparked a fascination with building circuits and understanding how they could perform mathematical operations or store memory. He built a buzzer system for his quizbowl team using logic chips, designing the circuits on a computer beforehand. Science fiction also played a role, fueling his imagination about the possibilities of intelligent machines.

From Curiosity to AI Industry

Naveen has always been driven to build things. He studied AI as an undergrad in the mid-90s, a time when creative algorithms and the definition of learning were central topics. Neural networks had an initial wave of success in the early 90s, particularly with digit recognition, but regression methods like Support Vector Machines (SVMs) gained prominence due to their ease of training and efficiency in lower data regimes.

The Resurgence of Deep Learning

The mid-2000s marked a turning point as larger datasets became available, saturating the performance of regression methods. Neural networks re-emerged as a way to learn models directly from data without manual "feature engineering." The ImageNet competition in 2012 was a pivotal moment, where neural networks significantly outperformed regression methods, approaching human-level performance. This event led to a widespread belief that the future of computing might be based on neural networks.

The Brain as a Model for Efficient Computing

Naveen highlights the brain's efficiency, running on approximately 20 watts, in contrast to the gigawatt power consumption of modern data centers. He points out that biological systems are constrained by energy, leading to efficient computational solutions. Current AI hardware relies on brute force, using the same basic computer architecture (ALU, memory interface, caches) that has been around for 60 years, but with more transistors and power. He argues that this approach is unsustainable and that fundamentally more efficient and AI-aligned computing elements are needed.

Limitations of Current AI and the Plateau

Naveen believes that current AI models are reaching a plateau in terms of fundamental improvements. While they are becoming more useful and economically valuable, they still make "dumb errors" and lack a true understanding of reality. He suggests that these models are limited by their reliance on incomplete observational data and latent representations.

Self-Improvement in AI Systems

Naveen acknowledges that AI systems are capable of self-improvement through techniques like reinforcement learning. His team at Data Bricks has developed a method called "towel" for iterative self-improvement without the need for labeled data. This involves automating the judgment of whether an output is good, similar to the shift from hand-tuned features to learned features.

Nirvana: Building AI-Specific Hardware

Naveen founded Nirvana, an AI chip company, with the goal of creating a processor designed specifically for neural network computation. The chip was designed for low-precision matrix multiplication, energy efficiency, and scalability. The architecture was highly distributed, allowing for seamless scaling from one to many chips.

The Intel Acquisition: A Missed Opportunity

Naveen describes the acquisition of Nirvana by Intel as "not the right move." While the initial thesis was to combine Nirvana's architecture with Intel's process technology, several factors led to challenges. These included personal reasons, fear of market downturn, internal political battles within Intel, and delays in Intel's process technology. The first Nirvana chip was completed, but subsequent development was hampered by Intel's internal conflicts and process delays, leading to a scrapped design and a restart on TSMC's 16nm process.

Rethinking the Physical Substrate of Chips

Naveen argues that computer architecture has remained conceptually unchanged for 60 years, relying on zeros and ones, synchronous design, and finite bitwidth representations. He believes that new ways of representing information are needed, potentially trading off time for accuracy or power. He mentions the Landauer limit as a theoretical limit of minimum power of computation that has not been thoroughly explored.

AI's Impact on Software Development at Data Bricks

Data Bricks is using AI co-pilots and working closely with model providers to enhance software development. AI is particularly useful for automating repetitive tasks like generating templates, headers, and project structures. Naveen sees potential for AI to accelerate chip design and verification cycles. However, he believes that AI is unlikely to generate truly novel ideas that are "out of distribution."

The Economics of LLMs and Collaboration with Meta

Data Bricks (through Mosaic before the acquisition) focused on making LLMs cheaper to train and serve. They developed a mixture of experts architecture that was highly efficient. However, they decided to partner with Meta on the Llama models, recognizing Meta's significant investment in open-source LLMs. Data Bricks continues to conduct research on reinforcement learning at scale with customer data.

The Intrinsic Link Between Hardware and Algorithms

Naveen emphasizes the close relationship between AI hardware and algorithms. The architecture of the hardware influences the design of neural networks. He believes that system-level AI work is intrinsically tied to hardware considerations.

AI and the User Interface (UI)

Naveen discusses how AI is changing the user experience by making software more adaptive and iterative. He uses the example of copy and paste as a metaphor for how AI can change the work experience. AI agents can learn from usage patterns and continuously improve, tightening the feedback loop between users and software.

The Agentic AI and Continuous Evolution of Software

AI is shifting the paradigm from static software releases to continuously evolving systems. Software is becoming more like a person, constantly learning and adapting.

Data Governance and Oversight of AI

Naveen stresses the importance of monitoring AI systems to ensure they are doing what they are intended to do. Humans should maintain oversight, using deterministic software to verify the behavior of AI agents. Governance structures are needed to define the boundaries of what AI models can and cannot do, including access to data and function calls.

The Risk of Model Collapse

Naveen warns about the potential for "model collapse," where LLMs train on machine-generated content, leading to the propagation of errors and a decline in the quality of information. He emphasizes the importance of grounding AI models in real-world data and leveraging curated datasets like those in Data Bricks' medallion architecture.

Personal Use of AI Agents

Naveen uses coding agents and tools to clean up writing. He also notes the subtle integration of AI in email, such as suggestions to include people or documents.

Synthesis/Conclusion

Naveen Rao's insights highlight the need for a paradigm shift in AI hardware, moving beyond brute force and embracing more efficient, biologically-inspired architectures. He emphasizes the importance of data governance, oversight, and grounding AI models in reality to avoid the pitfalls of model collapse. While acknowledging the economic value and usefulness of current AI systems, he cautions against overestimating their intelligence and stresses the need for continuous innovation in both hardware and algorithms. He also underscores the transformative potential of AI in software development, enabling more adaptive and iterative systems.