AI Breakfast #25

Executive Summary

At our 25th AI Breakfast, a group of software architects, fractional CTOs, and creative technologists discussed topics ranging from the ergonomics of coding agents and infrastructure choices for AI workloads to challenges in AI video consistency. The attendees also shared their latest work, including an "issue-to-PR" coding agent, a system for thermal physics simulations, and trading software utilizing AI for signal analysis.

Member Work

The host shared their progress on building an "issue-to-PR" online coding agent. They discussed their evolving toolset, noting that while they use Cursor and ChatGPT heavily, they have recently warmed up to Claude Code. They found Claude Code's default color themes and typography easier on the eyes during long sessions of reading generated text, and appreciated its more concise outputs compared to other models.

A software architect at a 3D printing company described their team's approach to maximizing human efficiency. They focus humans on "complex" issues (many variables, difficult to understand) while offloading "complicated" tasks (many moving parts, but straightforward) to agents. They implemented a "scars" policy: whenever a human has to manually fix code generated by an agent, they must also update the agent's instruction files. This ensures the agent learns from the mistake and doesn't repeat it, effectively building a repository of institutional knowledge.

A fractional CTO with a background in 3D and digital twins shared their recent experience learning Kotlin to build apps for Rokid AI glasses. They also discussed using AI to generate business plans for clients. Another attendee, a creative technologist and filmmaker, shared their frustrations with current AI video generation tools. They are working on a short film but finding it difficult to maintain character consistency—like clothing or ethnicity—across different shots, often resorting to generating massive amounts of footage to find "lucky shots."

Other members shared their projects as well: a portfolio manager is building AI-powered trading software to analyze market signals and is looking to build a UI for their Excel-based workflow; a data professional at a large beauty company is navigating the governance and scalability of internal agents; and a solopreneur is building a go-to-market automation platform and exploring workflow tools like Dify.

The State of Coding Agents

The group compared notes on their preferred coding environments. While Cursor is a popular choice, the host highlighted a specific "UI fatigue" from reading dense AI outputs all day. This led to a discussion about Zed, a high-performance code editor built in Rust. One attendee praised Zed for its configurability and speed, suggesting it as a strong alternative for those who find other editors sluggish or visually straining.

The conversation highlighted a trade-off: while modern AI-native editors offer powerful features, some developers still miss the robust debugging and file-handling capabilities of traditional IDEs like WebStorm. The group speculated on the future of these tools, with some hoping for better extension ecosystems to bridge the gap between AI capabilities and traditional development workflows.

Infrastructure for AI Workloads

A significant portion of the discussion focused on the "build vs. buy" decision for infrastructure, specifically Kubernetes versus Docker Compose. A fractional CTO argued that for most steady-state applications, Docker Compose is sufficient and less complex. However, the software architect explained why their use case—running transient thermal physics simulations—requires Kubernetes.

They detailed how their system spins up pods for specific simulation jobs and then shuts them down, a level of orchestration that Docker Compose cannot handle efficiently. They also shared how they use Carpenter, a Kubernetes auto-scaler, to manage AWS Spot instances. By automatically handling the interruption of cheap Spot instances and spinning up replacements, they achieve high performance at a fraction of the cost of on-demand servers. This setup allows them to run resource-intensive simulations economically.

Challenges in AI Video Consistency

The creative technologist in the group led a discussion on the limitations of current AI video models. Despite improvements in visual fidelity, models still struggle with temporal consistency and identity preservation. For narrative filmmaking, where a specific character needs to perform specific actions across multiple scenes, this is a dealbreaker.

The group discussed the "slot machine" nature of current workflows, where creators have to generate hundreds of clips to get one that matches their script. While some models allow for start and end frame interpolation, the middle frames often hallucinate unwanted changes, such as a character's suit changing color or their face morphing into a different ethnicity. This unpredictability currently limits AI video primarily to abstract or short-form content rather than serious narrative production.

Other Resources

• Claude Code: A coding interface praised by the host for its readable typography and concise, eye-friendly outputs.
• Zed: A high-performance, Rust-based code editor recommended for its speed and configurability.
• Streamlit: A Python library recommended for quickly turning data analysis scripts into interactive web apps.
• Superhuman: An email client mentioned for its speed and keyboard shortcuts, helping users process email faster.
• Carpenter (Karpenter): An open-source Kubernetes node provisioning project used to efficiently manage AWS Spot instances.
• Browser Use: A library mentioned for enabling AI agents to interact with web browsers to automate tasks.
• Rokid AR Glasses: Augmented reality glasses that one attendee is developing applications for using Kotlin.