This Week in AI for Ridiculously Busy People

The central theme of the week was token efficiency. The host argues that the AI industry has officially transitioned from a 'token subsidy era'—where per-seat pricing allowed users to consume thousands of dollars worth of compute for a fraction of the cost—into a 'token shortage era,' where usage-based billing is becoming the norm. Real-world signs of this shift included Uber capping employee AI usage at $1,500 per month, Walmart limiting access to its internal AI tool due to overwhelming demand, and TSMC signaling that the compute shortage could persist for years.

Despite the shortage, the market is actively responding with token-efficient architectures. Factory introduced native model routing to intelligently select cheaper or less capable models for simpler tasks, reportedly maintaining state-of-the-art performance while cutting costs by 25%. Perplexity launched a hybrid local-and-cloud inference system aimed at reducing both costs and privacy concerns. Harvey, in collaboration with Fireworks AI, built a 'worker-advisor' agent architecture where an open-weight model handles routine tasks and delegates only complex ones to a frontier model, outperforming the frontier model alone on legal tasks at a fraction of the cost. Microsoft demonstrated that post-training a model on McKinsey-specific tasks in collaboration with McKinsey resulted in GPT-5.5-level performance at one-tenth the cost.

On the product side, the host highlighted Codex updates as the top thing to experiment with, specifically three new features: Annotations (for editing specific parts of documents or websites), an expanded plugin ecosystem with function-specific packs (e.g., for salespeople), and 'Sites,' which allows users to convert any Codex project into a website or web app with a single click. The host believes Sites could make websites a fundamental unit of knowledge work.

The policy landscape is also shifting rapidly. Bernie Sanders published an op-ed in the New York Times calling for the government to own 50% of major AI labs. Separately, the Trump White House is reportedly considering taking equity stakes in leading AI companies, suggesting the Overton window on government-industry collaboration in AI is moving quickly. Both Anthropic and OpenAI released papers this week indicating they are observing early signs of recursive self-improvement in current AI systems, which the host suggests will intensify the policy debate significantly in the near future.

The host closed with takeaways: enterprises need to think architecturally about token efficiency (model routing, context management) and invest in agent-centric training programs. Solo practitioners should begin building personal systems now—context management, skill integration—before cost pressures increase further. The SpaceX IPO was flagged as the major event to watch the following week.