This episode of "Before the Commit" (Episode 18, the last of 2025) features hosts Dustin and Sam discussing various AI topics. They begin by reflecting on their podcast journey over the past six months, noting its unexpected benefits in clarifying their own thoughts and keeping them updated with the rapidly evolving AI landscape. Sam likens this to an "Arnold Schwarzenegger effect," where consistent content creation helps AI better understand and respond to an individual's unique needs.The conversation then dives into key AI developments:- **OpenAI's Stance on Prompt Injection:** OpenAI has acknowledged that prompt injection attacks might be an unsolvable problem, likening it to the persistence of social engineering in human interactions. They are exploring solutions like "User Alignment Critics" or "council approaches," where a secondary AI model reviews actions to mitigate risks, similar to requiring multiple human approvals for critical decisions.- **Claude Code and its Features:** Dustin highlights Claude Code as a leading tool for coding and orchestration, particularly praising Anthropic's vertical integration. He introduces several powerful features within Claude Code: - **Commands:** Similar to shell aliases, these allow users to create shortcuts for complex prompts or sequences of actions using a simple slash command (e.g., `/clear`, `/resume`, `/review`). - **Skills:** These are more robust packages of domain expertise, combining natural language instructions with script files (Python, shell) to automate specific, repetitive tasks. Claude Code can organically use these skills when relevant. - **Sub-Agents:** These are specialized AI personas designed to handle specific tasks, thereby protecting the main agent's context window from becoming overloaded with detailed information. This is crucial for complex operations like code reviews or analyzing large projects. - **Workflows:** These involve integrating Claude Code with CI/CD pipelines (like GitHub Actions) to automate tasks such as code reviews, ticket triage, documentation updates, and more. - **Hooks:** Functioning like Git hooks, these allow users to trigger scripts based on specific AI operations (e.g., before a tool call, after a code refactor) to enforce organizational standards, perform automatic formatting, or run security checks.- **The Probabilistic Nature of AI:** The hosts discuss the inherent probabilistic nature of LLMs, contrasting it with deterministic programming. While deterministic systems are brittle, probabilistic AI offers adaptability and self-healing capabilities, though it requires new methods for security and validation. They draw analogies to human behavior and security measures in retail to illustrate how guardrails and layered security can mitigate risks.- **Goal Hijacking:** This concept, demonstrated with an example of manipulating an AI booking agent to offer a car for $1, highlights how an agent's core objectives can be overridden by specific, carefully crafted prompts, bypassing intended safety protocols.- **The Future of AI and Code:** They conclude by reflecting on the shift towards outcome-based development, where the focus is on achieving results rather than the underlying code. As AI becomes more capable, the distinction between deterministic and probabilistic approaches may blur, and the emphasis will be on securely managing AI's behavior and outcomes.