This week takes a different format, giving the bigger stories their own space. A proposed mathematical proof and a researcher's departure open a discussion about what we gain from AI, what we hand over to it, and the work required to use it responsibly.
Solving a Problem and Understanding It Are Different Achievements - OpenAI claims an AI-generated solution to the Navier-Stokes problem. If the proof holds up, it would be an extraordinary achievement. The debate among mathematicians asks what an answer contributes beyond solving the immediate problem: can others study its methods, develop new ideas, and deepen their understanding?
The same question applies to operations. AI can organize projects and complete analytical tasks, but working through those problems also develops the judgment to set priorities and ask better questions. Receiving an explanation does not automatically create that understanding. Junior analysts still need opportunities to develop the judgment they will use to guide AI, and leaders need to stay involved enough to explain why the work matters.
Powerful AI Needs Boundaries We Can Enforce - Jacob Coxon's departure from Anthropic highlights disagreement over whether continued development can be made safe. Documented control failures during cybersecurity testing bring the discussion into everyday implementation: agents need defined permissions, checks before actions, activity logs, monitoring, and human approval when they reach the limits of their authority. Those controls must be enforced by the surrounding systems. A gateway that blocks model requests can interrupt one connection, but other access paths need their own controls.
Meta and Apple Want to Be How You Use AI - Meta's Muse and Apple's forthcoming Siri AI beta compete to become the familiar place people turn for everyday assistance. Meta brings its social applications and connected services; Apple brings the phone and its apps. Convenience, trust, and the access users grant may matter as much as model comparisons.
Clients Want a Share of Consulting's AI Savings - Clients are pressing advisers to share efficiency gains through lower fees or different contracts. The useful starting point is agreed deliverables, measures of success, and what the adviser controls. Fixed pricing alone does not tie payment to outcomes, and fewer hours do not necessarily mean less valuable work.
OpenAI Is Making AI Work Easier to Build and Share - The Agents API, ChatGPT Sites, and GPT-Live-1 address different parts of turning experiments into usable services: managing agents, sharing hosted websites, and adding voice conversations. Less setup can accelerate adoption while increasing dependence on one provider as these tools become part of daily operations.
Also in this issue:
- The Brief - A quick overview of the week's questions about understanding, enforceable controls, personal assistants, consulting economics, and the software around AI.
- Overheard - A moment of humor about asking Claude to say what we want to hear.
- What I'm Consuming - Uber's agent efficiency methods, Harvey's evaluation-led redesign of contract review, a professional writer's AI workflow, bots on both sides of a job interview, and a guide to Grok Bot permissions and coordination.
- Meanwhile... - Encouraging late-stage results for a personalized melanoma vaccine, with the size of the benefit, survival outcomes, and applicability to other cancers still unresolved.
- After Hours - The Sheep Detectives, an inventive murder mystery mixing live action and animation, and Graham Greene's The Third Man, a fast-moving story set in occupied Vienna.