1. AI Empowers “High Agency” Workers – WSJ Opinion by Replit CEO
Start with one of my favorite reads, and a big part of why we are writing the “Ollie and Louie” books for our children. Amjad Masad, CEO of Replit, argues that AI is a powerful amplifier for people with initiative—those who take ownership, seek leverage, and iterate fast. In this optimistic take, AI doesn’t displace but rather enhances the capabilities of what he calls “high-agency” individuals, unlocking entrepreneurial potential across technical and creative fields.

2. Ad Agencies Face Prolonged Slowdown Amid AI Disruption – WSJ
Barclays analysts project a slow-growth future for major ad agencies like WPP and Interpublic as AI forces a fundamental reshaping of the industry. Automation is squeezing margins and eroding traditional creative services, and while agencies invest in AI to remain competitive, the transition is expected to be costly and prolonged.

3. Hidden Valley Ranch Is Using AI to Get Weirder – WSJ
Some of my favorite quotes from this article:

• “We believe it’s got to be the people doing the work” who decide what AI approaches make sense and boost productivity
• “If you go in with the expectation that the AI is as smart or smarter than humans, you’re quickly disappointed by the reality,” says Eric Schwartz, Clorox’s chief marketing officer.
• The toilet bomb cleaner: It’s the sort of weirdness that comes from AI’s tendency to hallucinate and free-associate, a problem in many applications but helpful when you’re trying to be creative.

4. Does Math Reasoning Improve General LLM Capabilities? – arXiv Explained
Math tuned LLMs don’t automatically get smarter. A new study analyzes 20+ reasoning-tuned LLMs and finds that models trained with reinforcement learning (RL) become true generalists, transferring their reasoning skills across domains. Meanwhile, traditional supervised fine-tuning creates brilliant specialists that forget how to do basic tasks.

5. Exploration Is (Probably) What You Want – Yiding Jiang
Interesting technical blog about the data that LLMs are being trained on as being similar to “fossil fuels”. We are close to running out of data and are consuming data faster than we can produce it. Future progress hinges on exploration—agents generating their own, richly informative experiences.

6. The Illusion of Thinking – Apple ML Research
Apple researchers critique the tendency of LLMs to “fake” reasoning with fluent but shallow answers. Standard evaluations focus on final answer accuracy and risk contamination, failing to assess the reasoning process itself. Current LLMs create convincing “thinking” but can fail catastrophically on hard problems, and their effort doesn’t always rise with difficulty.