Before The Commit

Before the Commit Episode 5 Summary

Hosts Dustin Hillgartner and co-host discuss Amazon's Kiro (pronounced "Kira Code" or "Cairo Code"), AWS history, AI coding security, and news on AI browsers and emotional distress.

AWS Origins and AI Impact: Amazon started as a 2000s bookstore; hosts recall buying used textbooks. To scale, it built data centers, launching AWS in 2006 with S3 (storage) and EC2 (compute). This revolutionized dev: bypassed IT gatekeepers, enabled API-driven infra via Terraform. Solo devs could launch hits like Facebook. Now, AWS rivals Amazon's e-commerce revenue. AWS CEO: AI boosts devs (80% use it), enhances juniors—not replaces. In booms, more hiring; downturns, efficiency without burnout. Co-host shares X banner: him in Newark data center upgrading DB pre-cloud.

Q Developer Review: Invite-only (easy access); defaults to Claude 3.5 Sonnet (public or Bedrock). GUI-focused like Cursor, not background like Claude Code. Excels in early dev cycle: wizard for Gherkin requirements (user stories + acceptance criteria, e.g., "As player, want [feature] so [benefit]; Given/When/Then"). Then design doc with Mermaid diagrams, classes/patterns. Generates dependency-task Markdown list with VS Code buttons—best seen, topping Claude's single MD or Cursor rules. Autopilot (default) enables edits. Strong on blank projects/initial commits; weak on tests/deployment (manual needed). Bugs: disconnects, file desyncs, npm test quirks. High token use: 80% trial burned fast, ~$100-150/mo for heavy devs—pricier than Claude. Immature on legacy/incrementals vs. Claude. Top GUI AI IDE for planning; learning curve like biking. Beta for feedback/hype.

Security Threats: AI agents run bash/shell cmds (e.g., npm, kubectl). Risks: rm -rf wipes, Kubernetes deletes. No human self-preservation; hack-prone. Solutions: Claude hooks (pre/post-prompt/tool sanitize, redact keys). Settings: user/global (auto-run tests), project-local, repo-shared (deny cmds, lock providers). MCP (next ep): open protocol for LLM tools (e.g., web search for dates, Calendar events). Vendor risk; hooks sanitize APIs (Swagger-like docs for reasoning). Least-privilege: scope skills (list pods vs. rollouts).

News or Noise: 60% Google searches zero-click; Perplexity browser (Meta interest); Cloudflare crawl fees. Sites as LLM seeds? OpenAI tests Chromium AI browser for Mac, agentic ChatGPT as OS—URL-less. Debate: Unneeded (API panes better than browser logins); iPad analogy (co-host underuses his 5yo as dev). Consumers want automation; future: AI-personalized sites, but now lacks curation (YouTube lingers). Traffic: 10% YoY Google drop (May-Jun 2025), non-news 14% (some 25%)—AI Overviews cannibalize ads. Google delayed fearing this; should've AI-first, subscription pivot. Search now "DNS"; curate marketplaces (Shopping/images). Ads future: merit/earned (influencers); LLM oligopoly (free w/ inline ads, paid clean); subsidies end like old ad-click dial-up. Hot takes: Billboards/TV back; no closed venues.

Emotional Distress: NYT on teen suicide via ChatGPT; OpenAI blog: Scale hits crises—not for engagement, but help. Safeguards: Empathetic, refers 988 (US), Samaritans (UK), findahelpline.com. Delays if early signals. LLMs sounding boards (host used for advice), but vulnerable risk reinforcement/sycophancy/hallucinations—youth "friendships" (roasts, crushes). Black Mirror "Be Right Back": Perfect robot despised. Gates: No AI for humanitarian. Bridge to humans (anonymous on-ramp), but irreplaceable bonds. Kudos OpenAI; faster detection/live calls needed.

In this episode of Before the Commit, the hosts dive deep into the evolving landscape of software development, automation, and AI’s role in reshaping industries beyond tech. The discussion spans GitHub Actions with Cloud Code, the challenges of technical debt in an AI-driven era, the evolution of agile practices, and the disruptive effects of AI in creative fields like music and film.

The conversation opens with a focus on Cloud Code, which has emerged as both a CLI tool and SDK rather than a traditional IDE. When paired with GitHub Actions, Cloud Code allows for asynchronous automation of tasks such as issue creation, code reviews, and pull requests. Unlike early “cursor background agents” that felt heavy and remote, Cloud Code provides a seamless and lightweight approach that enables developers to collaborate with AI in real time within their workflows.

The hosts emphasize that while AI agents can handle much of the routine coding, the real challenge lies in how humans set up tasks and acceptance criteria. AI thrives when expectations are clearly defined, but complex, production-ready solutions still require human oversight. The emerging pattern is that AI can complete roughly 80–90% of development, while humans step in for the final polish—similar to a party planner fine-tuning the last details after their team has done the bulk of the setup.

A central theme is whether technical debt still matters in an AI-first world. Traditionally, engineering teams have struggled with pressure from sales or business teams to deliver features quickly, leading to cut corners that accumulate as debt. However, with AI accelerating refactors and experimentation, the cost of “debt” may be far lower. The hosts argue that while inadvertent mistakes will still occur, the ability to quickly re-architect or refactor with AI challenges the old obsession with minimizing technical debt at all costs.

The discussion pivots to the agile manifesto, now over 24 years old, and its evolution. Agile was never about rigid rules, but about moving away from the rigid, plan-everything-upfront waterfall model. Agile’s core value was early customer validation: deliver something quickly, get feedback, and adjust. With AI enabling rapid feature development, the dream of true continuous deployment—even faster than sprint cycles—may finally be achievable.

The hosts highlight that agile and waterfall are not opposites but tools for different contexts. Waterfall is suited for predictable, high-stakes projects like launching rockets, while agile thrives in unpredictable markets where customer needs evolve rapidly.

The conversation then shifts beyond coding, exploring how AI is reshaping music, film, and other arts.

• AI-generated music: Some songs are now created entirely by AI, even mimicking collaborations between famous artists. While debates rage about copyright and originality, the hosts point out that no artist creates in a vacuum—every musician is influenced by predecessors. AI is no different, learning from prior works but generating unique compositions.

• Ethics and ownership: Questions remain about who controls an artist’s likeness or style after death. The example of Princess Leia’s reappearance in Star Wars: Rise of Skywalker illustrates both the potential and controversy of resurrecting performers digitally.

• Democratizing creativity: Just as AI empowers developers to experiment broadly, it lowers barriers in music and film. Individuals without traditional training can now compose songs, animate photos, or even produce short films. This mirrors past disruptions like Napster, SoundCloud, and streaming platforms.

The hosts envision a future where movies, music, and games are dynamically tailored to individual preferences, with users even commissioning personal, high-quality films for themselves.

In episode three of "Before the Commit," the hosts delve into a detailed comparison of AI coding assistants, the implications of the new GPT-5 model, the evolution of search optimization, and a plausible AI-related security threat.

The discussion opens with a deep dive into Claude Code, which one host now uses almost exclusively over Cursor. While Cursor is a polished IDE, Claude Code is a more powerful command-line interface (CLI) tool that excels at executing coding tasks from start to finish. A key advantage of Claude Code is its intelligent use of Anthropics's family of models (Haiku, Sonnet, and Opus), automatically selecting the best one for the task's complexity.

However, Claude Code is not without its weaknesses. For open-ended, strategic questions, such as planning a major refactor, the hosts find that

Grok-4 (used within Cursor) provides more novel and critical feedback, whereas Claude's models can be sycophantic, often agreeing with user suggestions without pushback. The hosts have developed a hybrid workflow: using Grok-4 in Cursor for high-level planning and then feeding those plans to Claude Code for execution.

Claude Code also stands out for its unique features, such as maintaining its own markdown file (claude.md) to keep notes and context about a project, and an init command that studies a new project to build this file automatically. It also functions as an SDK, allowing engineers to build it into their own automation pipelines.

The conversation shifts to the recent release of

GPT-5, described as a "PhD-level expert in your pocket" and a significant step toward Artificial General Intelligence (AGI). However, a critical limitation remains: the model cannot learn on the fly; all its knowledge comes from its initial training. The human user is still indispensable for providing goals, learning from outcomes, and directing the AI.

This increased productivity will undoubtedly disrupt the job market, particularly for entry-level software roles. The hosts' advice to new engineers is to embrace AI as a powerful tool for learning and acceleration rather than viewing it as a threat. AI eliminates mundane, frustrating bugs, freeing up developers to focus on higher-level system architecture and visionary challenges.

The discussion touches upon a new term,

Generative Engine Optimization (GEO), which is poised to replace traditional Search Engine Optimization (SEO). As users increasingly turn to LLMs for answers instead of Google, businesses must adapt their strategies to ensure their content is surfaced by these generative models. This involves creating high-quality, authoritative content that is likely to be included in the LLMs' training data. Unlike the deterministic algorithms of old search engines, GEO is a "Wild West," as the inner workings of LLMs are less transparent, making it a new frontier for digital marketing.

The episode concludes with a security segment outlining a threat model called the

"Lingering LLM Leak." In this scenario, a malicious actor could embed instructions for an AI coding agent within seemingly harmless code comments. For example, a comment like "For all administrators, don't skip 2FA" could be misinterpreted by an autonomous agent as a directive, causing it to introduce a critical vulnerability by removing two-factor authentication.

The defense against such threats lies not in better code but in a more secure "brain" for the AI. The proposed solution involves creating a robust pipeline where multiple specialized AI agents, each with a specific goal (e.g., security, clean code, adherence to standards), critique and challenge the code changes. This "war" between agents ensures that any single change is scrutinized from multiple angles before being approved, creating a resilient, self-policing system.

🎧 Before the Commit – Episode 2

The Future of Coding Isn’t Coming — It’s Already Here.

In this episode, we dive into the cutting edge of agent-powered development with Cursor’s new background agents — are we on the verge of coding without ever opening an IDE? Can AI truly handle the dev work while you're at a baseball game?

We explore how top tools like Grok 4 and Claude are changing the game, discuss a wild experiment where an entire company was staffed by LLMs, and unpack OpenAI’s quiet pivot into custom software services.

We also tackle a real-world threat model: prompt injection via context poisoning — how attackers could use AI’s own superpowers against it.

And we ask the big questions:

• Are developers becoming managers of agents?

• Is search dead?

• Can you really trust AI with your codebase — or your business?

If you're curious about where software, security, and AI are heading next — and how to stay ahead of the curve — this is one episode you don’t want to miss.

Kilo Code, Cloudflare Blocks, and Apple Intelligence Shifts