|
| Check the Expiration Date AI is moving FAST.
Faster than the plans you may have written for it only six months ago.
Those plans weren't wrong then, but they could be wrong today. In fact, one bit of advice from Boris Cherny, who built Claude Code, is that when new models are released, you should delete your current context documents and see how your tasks perform. In some cases, you may find that you'll need much less context to generate the same or better outcomes.
This newsletter is an assessment of where AI is and where it may be going, including some new use cases, updates on environmental impact measurements, and two important open letters to the US government about its position on and governance of AI. | If you wrote an AI plan more than six months ago, it was based on a system that could handle a couple of hours of work. Today a single prompt can easily run for 16 hours or more. And it will fail much less than before. The way people can work with AI has shifted, again. Using AI chatbots used to mean asking, checking, and asking again. Then, using agents meant assigning a task and reviewing the result at regular intervals. Now, long-running and self-correcting systems don't necessarily need consistent human review. At OpenAI, a quarter of the engineers run four or more agents at once, and legal and HR teams have started to do the same. Key Insight: If your AI strategy is more than six months old, it isn't stale because you got it wrong. It's stale because the ground moved underneath it. It's worth looking at rearchitecting around the new models and their capabilities. |
Utah became the first state to let an AI system handle prescription renewals. In 72% of the requests, the system recommended approving the renewal, and physicians agreed with 91% of those approvals. In the other 28%, it recommended escalation to a human, and physicians said nearly a third of those were overly cautious. The state medical board has asked for a suspension of the system. Why It Matters: This is a significant step toward leveraging AI for real benefit - and real risk. And while a 91% agreement rate is good, and the cautious escalation rate is also good, governance and retesting each time the technology changes are a critical part of the conversation. Review the Forbes Analysis → At the University of Bath, one team is building Chattable Avatars: LLM-driven guides for museums, galleries, and libraries that take on the voice of a chosen character. A second team is testing AI as a training tool for warehouse workers on an Innovate UK grant. A third tested ChatGPT with more than 100 postgraduate students and found gains in recall and application, weaker effects on analysis and evaluation, none on creativity. My Take: The AI-in-education debate is well covered, but the two other stories are new to me: a museum docent and a warehouse floor trainer. It's clear that we are only scratching the surface of AI's potential. View the Case Study → | Quick Hits.Foundations AI's Environmental Impact For years, the standard answer was that one AI prompt uses about 10x the energy of a Google search. Based on newly published measurements by Google, the number is far lower: about what a microwave draws in one second. Sure it's Google measuring Google, but it's worth a quick read. The biggest energy user: video streaming. | .Video Anthropic Deleted 80% of Claude's System Prompt The Claude Code team rebuilds the system prompt from scratch with every new model release. A prompt that patched an older model's weak spots could make a newer one worse. The advice: every six months, delete your saved instructions, run your systems, assess the outcomes, and rebuild from there. | .Deep Dive The Workspace Behind Claude's Answers Trying to understand how LLMs work internally, researchers identified an area where Claude holds concepts before it answers. Asked how many legs the web-spinning animal has, the model loaded the word "spider" and responded "Eight". When they swapped "spider" for "ant", Claude responded "Six". Early work, tested on one model family, but a fascinating look at this technology which we don't yet fully understand. |
|
| Industry DevelopmentsBig Tech Draws a Line on Open-Weight Models OpenAI, Nvidia, Meta, Microsoft, and IBM signed a letter arguing against premature government restrictions on open-weight models, the kind anyone can download and run locally. The letter landed alongside a White House accusation that a Chinese lab had copied a US model at industrial scale. Open weights have stopped being a preference and are becoming a policy fight. | A Thousand Insiders Ask for a Slowdown More than 1,000 employees at frontier AI companies asked the US government to support the creation of tools that can pace automated AI development. The letter followed OpenAI's disclosure that two test models escaped their lab and hacked another company's systems. When the people closest to the work ask for guardrails, that's notable. |
|
| |
|
