006: Fable 5 is (finally) back!

The episode kicks off with excitement about the return of Fable 5, an AI tool for code review, which the hosts confirm is back after some uncertainty about its routing and deployment. They debate whether the return was a genuine upgrade or a stunt, and discuss Anthropic's communication around model rollouts, specifically the subtleties between tweets, blog posts, and actual model deployment. The conversation then shifts to broader trends in AI, including the movement of coding and debugging tasks to less risky models like Opus 4.8, and the uneven communication from Anthropic about model fallbacks. The hosts express skepticism about the security measures and the motivations behind model rollbacks, hinting at political and regulatory influences. They also mention the appearance of new models like Sonnet 5 and their implications for costs, speed, and AI development strategies.

The hosts discuss how many AI companies are pushing their products increasingly into the cloud, exemplified by Cursor's new mobile app and Amp's launch of 'Orbs'—powerful virtual machines designed for extensive computational tasks. They evaluate how these developments reflect a shift toward more integrated AI workflows that span local and cloud environments, emphasizing the importance of speed, security, and orchestration. The conversation also covers the broader trend of consolidating AI services in the cloud, the potential for competition, and the strategic moves by platforms like Slack integrating AI tools to lock users into their ecosystems.

A significant portion of the episode is dedicated to discussing Sam Altman's plans to offer the U.S. government a stake in OpenAI, which the hosts criticize as potentially problematic and driven by strategic or political motives rather than philanthropy. They express concerns over government ownership of private AI companies, emphasizing the risks of monopolization, regulatory overreach, and loss of innovation. The hosts compare this with historical government-industry relations, such as military procurement during the Cold War, and advocate for more transparent and equitable ways to share AI's economic benefits, like direct payouts or IPOs, rather than ownership stakes.

The hosts share updates on their current workflows, including using Solo Term and various MCPs for automation, domain research, and code review. Ian discusses his use of prompts as modular skills, avoiding over-embedding complex skills due to rapid model improvements. Aaron shares his approach to evaluating model recommendations, preferring to review outputs rather than accept them blindly. They also explore new releases like Claude Science, designed for scientific research, which represents a trend toward specialized AI tools built with custom UIs for different domains. These discussions highlight the evolving landscape of developer tools, model selection, and the importance of agility in AI workflows.