RS368: Grok Bot & Hermes

The episode begins with Craig expressing a clear preference for Codex over Claude and other models, emphasizing the importance of building 'superpowered' AI systems that can do anything to extract real value. They discuss ongoing experiments with AI bots and agents, noting how companies like Rockbot and Codex are developing open versus closed platforms. Craig mentions the cost of hitting enter in AI interactions and how many users attempt to maximize experimentation within their limits, turning limitations into an opportunity to learn.

The hosts delve into the strategic choice between open and closed AI systems, debating the benefits of owning and controlling the infrastructure versus relying on proprietary platforms. Craig chose Hermes for his company's shared AI agent because of its configurability, security features, and control over data. They consider how sharing bots within teams enhances collaboration, but also highlight risks such as security and brittleness, especially with open systems like OpenClaw. Craig emphasizes security as the primary concern when giving AI agents significant power, stressing the importance of ownership and control.

Both Craig and Colin share their experiences with deploying AI for everyday tasks like CRM, task management, and reducing cognitive load. Colin describes how he uses AI to manage his CRM and integrates it with tools such as Notion and GitHub, enabling mobile and cross-platform workflows. Craig mentions how his team used AI to generate PRs and automate marketing, acknowledging that although current systems are not perfect, they are progressively improving. They discuss the ambitious vision of AI as a shared community assistant that can coordinate complex workflows across teams.

A significant part of the discussion centers on balancing AI model capabilities against cost and performance. Colin shares how he downgraded his workflows from expensive models like Anthropic’s to more affordable ones, noting that even lower-tier models like Sonnet can produce acceptable results for specific tasks. They debate whether the best models are always necessary, especially if users have existing skills and well-defined processes. Colin considers investing in powerful local hardware, such as a Mac Studio with substantial RAM, to run models privately, highlighting the trade-offs between cost, control, and convenience.

Colin describes his project of using AI to overhaul a 15-year-old WordPress site for his father’s lawn care business, leveraging AI agents to automate bug fixing, SEO improvements, and design updates. Craig reflects on how AI tools like Claude, Fable, and Astro enable creating more efficient content workflows, especially for tasks like research, article updates, and generating marketing materials. The conversation emphasizes moving away from traditional CMSs like WordPress toward modern stacks that integrate AI for rapid content production and experimentation, including building simple, purpose-specific tools and lead magnets.

The episode concludes with a reflection on the potential of AI agents to act as 'chief of staff' figures, helping manage priorities, prevent projects from stalling, and keep users focused. The hosts consider the importance of transparency and permissioning in multi-agent setups, as well as the necessity of systems that can maintain an overarching view of ongoing projects. Craig and Colin explore the idea of AI-powered tools that can be used for content, marketing, and technical work, underscoring the ongoing shift toward remote, automated, and private AI infrastructure.