Devivre Personal Podcast

This episode explores a framework called ALTER, which uses AI in five roles—Advisor, Librarian, Tutor, Editor, and Roommate—to guide self-education. The system aims to create a personalized learning path, curate reliable resources, provide on-demand tutoring, refine work through feedback, and foster interdisciplinary connections.

This episode explains that the primary bottleneck in AI performance isn't the size of its context window, but the efficiency and cost of reading new information. Recent innovations by DeepSeek and Xiaomi address this by splitting AI models to reduce the processing required for reading prompts, making AI faster and cheaper to run.

New guidelines for AI skills address how models process reference files, with a rule to add a table of contents for files over 100 lines. The approach to instructions has shifted to "appropriate degrees of freedom," matching instruction detail to task fragility, and skills must be tested across different AI models. Additionally, reference files should be kept one level deep from the main skill file, and feedback loops can be built into skills for self-correction and improvement.

An experiment was conducted where an AI was given $10,000 to trade stocks for one week with the goal of outperforming the S&P 500. Despite initial challenges with the AI's passivity and a technical workaround involving a second bot, the AI eventually began making trades. However, market volatility and a Federal Reserve announcement led to losses, with the AI ultimately underperforming the S&P 500 by a small margin.

This episode discusses how to adapt AI systems, particularly generative models like chatbots, to comply with regulations like the EU AI Act. It explores the challenges of auditing dynamic systems and introduces a framework called Governance-as-Code, which translates legal requirements into automated checks within the deployment pipeline. The discussion covers how to define and measure abstract concepts like robustness and bias, and clarifies the tiered responsibilities for companies building on foundation models.

The United States and China are developing distinct approaches to AI governance, with the U.S. focusing on supply chain security and innovation, while China offers a comprehensive package of governance models, AI access, and infrastructure to developing nations. Middle-power countries are adopting a strategy of engaging with both sides, highlighting the need for the U.S. to establish a clear, exportable national AI governance framework to effectively compete on the global stage.

A new AI model, GLM-5.3, has been released with minimal safety features, enabling it to autonomously create sophisticated cyber exploits. This open-weight model can be modified to bypass safeguards, allowing for the creation of cyber weapons with a dramatically lowered barrier to entry.

An AI model was temporarily shut down due to a jailbreak that exposed software vulnerabilities, highlighting concerns about AI's ability to autonomously find and exploit critical software flaws. This incident underscores the race to develop AI capabilities and the need for increased AI literacy and cybersecurity measures.

This episode discusses a radical shift in programming workflows driven by AI agents, moving from single-threaded deep focus to parallel processing and orchestration. The setup involves using multiple agents, managed through terminal tools like tmux and Herdr, and scaled across multiple machines connected via Tailscale.

This episode discusses new prompting strategies for Anthropic's Opus 5.5 model, focusing on the 'effort' setting which is now recommended to start at medium. It also covers improvements to the model's memory cache, simplified workflows for developers using multiple coding agents, and new ways to request faster task completion.