Why AI Can Write Code But Can’t Keep It Running (Yet)

Shift from Syntax to Semantics in AI Development

• AI progress in software has evolved from simple autocomplete (syntax) to generating complex, semantically correct code that aligns with machine runtime behavior.
• "Semantics" refers to the actual behavior and properties of a system during execution, distinct from the written code (syntax), and remains the hardest barrier for AI to fully master.
• High-performing engineers leverage AI by training the models on their specific workflows, allowing the AI to adapt to complex constraints that generic models cannot yet predict.
• Current AI models excel at generating code but struggle with operational "heavy-tailed" distributions, where rare, exception-based states dominate system behavior.

Operational Complexity and the "Dirty Secret" of Infrastructure

• Historically, managing distributed clusters relied on human intuition and tribal knowledge rather than predictive algorithms, a practice the speaker identifies as a "dirty secret."
• The operational state space of software grows exponentially (factor of two per bit), creating a "universe" of states that is effectively infinite and impossible to fully enumerate or train on.
• Unlike code generation, AI cannot yet fully predict system correctness, availability, or performance trade-offs without grounding in real-world experiments or telemetry.
• The speaker argues that the history of systems engineering favors over-provisioning resources (e.g., RAM, network capacity) over complex predictive algorithms due to the unreliability of modeling all possible operational states.

Grepper's Approach to Observability and Operations

• Companies traditionally spent 5% to 20% of infrastructure budgets on observability tools to proxy for reliability, but marginal gains in reliability flattened as data volume grew.
• Grepper reduces observability costs by up to 90% by extracting signal from noise upstream in the data stream, rather than storing and processing all raw telemetry.
• The company's strategy involves using AI to analyze logs for new patterns and anomalies, then feeding these signals to a reasoning engine that hypothesizes and verifies causal chains.
• This "anomaly-to-causality" workflow replaces traditional root-cause analysis (which backtracks from lagging indicators) by focusing on the specific system changes that triggered a deviation.
• In security and operations, the goal is to have AI run specific experiments or queries to gather data, acknowledging that the AI cannot "think" the answer without simulating or observing the environment.

Engineering Productivity and Team Dynamics

• Using AI tools (specifically Claude Code with command-line agents and WSL), the speaker reduced a six-month engineering task to approximately ten days for a single person on a "auto-promotion" data lake feature.
• AI adoption creates a productivity divergence: top engineers become 10x more productive by effectively orchestrating sub-agents, while less experienced engineers do not see the same gains.
• The review process has shifted focus from code syntax verification to testing logic, golden expectations, and system behavior validation.
• The speaker utilizes a workflow where a main code agent spawns specialized command-line sub-agents that maintain persistent sessions to handle parallelized research and implementation tasks.

Market and Economic Perspectives on AI

• AI adoption in constrained digital domains (coding, image generation) is exploding because the distribution of solutions matches the training data distribution.
• Expansion into the "actual world" (e.g., self-driving cars, complex operations) faces significant hurdles due to economic equilibrium; the cost of solving these problems often exceeds the value they provide compared to simpler solutions.
• The speaker notes that self-driving cars required 20 years and $170 billion in investment yet still face worse unit economics than human-driven ride-sharing services.
• Venture Capital remains a human-driven field where success depends on relationship building, access, and trust, rather than just analytical capability.
• Future value creation will likely involve "constraining" the environment (e.g., designing systems specifically for AI interaction) rather than forcing AI to solve unconstrained, chaotic real-world problems.

Forward-Looking Statements on AI Capabilities

• The speaker remains skeptical that AI can soon replace human management of distributed clusters, viewing operations as a domain of "computational irreducibility" where simulation is required to get answers.
• Predictive modeling for operations will likely evolve to detect anomalies deterministically, triggering AI agents to run specific experiments to resolve the state.
• While code generation is nearing a "solved" state for many complex tasks, operations and security will take significantly longer due to the adversarial nature of the latter and the chaotic nature of the former.
• The ultimate limitation of AI is the gap between the "digital world" of training data and the "actual world" of physical execution, which may require a paradigm shift where humans engineer the world to fit AI's capabilities.