Why AI Coding is More Dangerous Than You Think | Traversal, Anish Agarwal

Key Concepts:

• AI-driven code generation
• Software downtime and troubleshooting
• AI Site Reliability Engineer (AI SRE)
• Causal machine learning and reinforcement learning
• Large Language Models (LLMs)
• Observability systems
• Time to resolution (TTR)
• Minimum Viable Product (MVP) vs. Production System
• Reasoning models and detective story workflow
• Context loss in complex systems

1. The Problem: Increasing Software Complexity and Downtime

• AI-Generated Code: Code is increasingly written by AI systems like Cursor and Copilot, leading to a "Cambrian explosion of code."
• Lack of Context: No single person or team fully understands the entire codebase, making troubleshooting difficult. "No one has full context about what's happening because it's such a complex system."
• Cost of Downtime: Global IT outages are a major problem, costing enterprises approximately $400 billion annually. These outages affect critical infrastructure like hospitals, law enforcement, banks, and airlines.
• Troubleshooting Bottleneck: Traditional troubleshooting involves multiple teams, escalating from 5 to 80 people in a channel, leading to prolonged resolution times.

2. Traversal's Solution: AI Site Reliability Engineer (AI SRE)

• Core Function: Traversal is building an AI SRE to automatically troubleshoot and fix complex software systems when they break.
• Series A Funding: Traversal raised $48 million in Series A funding led by Sequoia and Kleiner Perkins.
• Perplexity Analogy: The AI SRE is similar to Perplexity in that it provides evidence and citations for its conclusions. However, instead of web links, the citations are links to the user's observability system.
• Digital Ocean Case Study: Traversal has been working with Digital Ocean (a large public cloud service provider with over 600,000 users) for six months.
• Improved Time to Resolution: Traversal has reduced Digital Ocean's time to resolution (TTR) by over 40% (37% to be exact).

3. Anish's Background and Motivation

• MIT and AI Research: Anish studied machine learning and AI at MIT, inspired by Google's AlphaGo and its ability to learn creatively through self-play.
• Causal Machine Learning: Anish's research focused on causal machine learning, which studies cause-and-effect relationships in data. Examples include A/B testing and clinical trials.
• Impact of ChatGPT: The emergence of ChatGPT felt like a "once in a lifetime thing" and a "religious experience," prompting Anish to apply AI to real-world problems.
• Uncertainty and Entrepreneurship: Anish enjoys uncertainty and creating something from scratch, drawing parallels between research and entrepreneurship.

4. Traversal's Development Process

• Problem Selection: Traversal started without a specific idea but focused on the intersection of causal machine learning, reinforcement learning, and AI agents.
• Incident Management: The problem of dealing with incidents was identified as a perfect fit, aligning with their research and the capabilities of LLMs and AI agents.
• MVP Development: An initial MVP was built within three months and achieved 90% accuracy with small companies.
• Challenges with Enterprise Scale: When tested with larger enterprises like Digital Ocean, the initial MVP's accuracy dropped to 0%.
• Re-architecting for Production: Traversal re-architected its system to leverage the computational power and inference capabilities of AI systems, rather than relying on human-like creativity. This led to a return to 90% accuracy.

5. Leveraging Reasoning Models

• Detective Story Analogy: Reasoning models are effective at "detective stories," connecting disparate pieces of evidence to arrive at a clear conclusion.
• Application to Incident Resolution: This detective story workflow aligns well with incident resolution, where the AI connects various symptoms to identify the root cause.

6. The Future of Software Engineering

• Context Loss and Debugging: The increasing complexity of AI-generated code leads to context loss, making debugging more challenging.
• Shift in Engineer Roles: Engineers are increasingly focused on validating, QAing, and troubleshooting code, rather than developing it.
• Reinventing Software Maintenance: Software maintenance needs to be reinvented to allow engineers to focus on more creative and architectural work.

7. Key Advice and Principles

• Perseverance: "If something goes wrong, that's okay... If it was easy, then everyone could do it."
• Focus on Long-Term Value: "Think about what's going to be important 10 years from now regardless of whatever happened in the world."
• Embrace Uncertainty: Acknowledge that in research and entrepreneurship, you often don't know where you are or where you're going.
• Resilience: Develop the ability to recover quickly from stress and setbacks.
• Surround Yourself with the Right People: "You live and die by the people you surround yourselves with."
• Deep Motivation: "I wouldn't just start a company for the sake of it. I think you should you should feel like some thing deep in you cuz it's not easy."

8. Traversal's Value Proposition

• AI-powered incident resolution: Traversal offers an AI-driven solution to automate the troubleshooting and resolution of software incidents.
• Reduced downtime: By significantly reducing time to resolution (TTR), Traversal helps companies minimize the financial impact of downtime.
• Improved engineer productivity: By automating troubleshooting, Traversal frees up engineers to focus on more creative and strategic tasks.

Conclusion:

Traversal is tackling the growing problem of software downtime caused by increasing complexity and the rise of AI-generated code. Their AI SRE leverages causal machine learning, reinforcement learning, and LLMs to automate incident resolution, significantly reducing time to resolution and freeing up engineers for more creative work. The company's success hinges on its ability to effectively leverage AI's strengths, particularly in reasoning and inference, and to build a production-ready system that can handle the complexities of enterprise environments.