Relative to your interests, Tuesday

I have been on vacation for a little while, in Finland for the third time. Finland feels different than the other parts of Europe I have been to. Perhaps it’s because I spent my time there in rural Finland. But it also feels like most of Finland is rural. In that way, it feels a lot like the spaces between the big cities in Texas. There are also massive, 24 hour hypermarkets, which makes living life much easier than the hyper-local, closed on Sunday retail culture of continental Europe. Plus, lots of parking, and you don’t have to pay to pee in public spaces.

And now, the links:

• European Marketers Say AI Won’t Replace Employees, But The Reality Is More Complicated – “Nearly half (46%) of European marketers say their organizations have already reduced headcount or replaced marketing employees with AI.”
• weeknotes 5 July 2026, Neil Williams – One mindset to make presentations useful tools: “McKinsey team’s output is all slides (which they call ‘pages’ – and ‘voice over’ rather than present), backed by analysis in spreadsheets. No wordy docs.”
• Liberty Mutual preps legacy systems for AI scale
• The Right Amount of Spec for Agentic Development – Using short feedback loops to experiment and direct software development is always a good idea, especially with AI.
• LLM cliché highlighter – find common AI sentence structures.
• Herdr: one terminal for the whole herd – iterating on tmux.
• White-collar workers are quietly rebelling against AI as 80% outright refuse adoption mandates – possible future: only tech people like AI, normies are more like “this is yet another way for executives to fuck me over.”
• How people are using Claude Cowork – “largest category of use is for ‘business process and operations’—things like pulling scattered updates into a single report, building onboarding checklists, and reconciling spreadsheets—at 33.4%. This makes sense, because business operations tasks span many different roles: people in finance, HR, and administrative jobs, for example, are all likely to make use of Claude Cowork for these types of tasks.” Also, see an overview of multiple surveys on what normies use AI for.
• How to manage AI investments in the agentic era – Expecting AI to improve dysfunctional parts of your business probably won’t go well. Optimizing the already optimized is optimal: “The strongest candidates are workflows that repeat at meaningful scale, have clear ownership, and can be measured for quality, risk, and business value.”
• The FDE Arms Race: Why Every AI Company Is Spending Billions to Embed Engineers Inside Customers – when things are not optimal, you need optimize them before you can apply a revolutionary new technology.
• 🤖 The PRD Isn’t Dead, It Just Needs to Evolve – Traditional PRDs are too slow. Modern teams adapt by using lightweight briefs, decision logs, RFCs, and prototypes to keep alignment without creating bottlenecks.
• USSFCU Secures Financial Data on a Sovereign Private Cloud – “Developing the secure data layers needed for large language model (LLM) use cases. Because its AI compute and data storage operate within the same strict perimeter, the credit union can confidently leverage highly sensitive member data. This makes Retrieval-Augmented Generation (RAG) easy to implement while keeping member information completely isolated from public APIs.”
• The CEO of AWS on why Amazon is hiring 11,000 interns and junior employees – Yes: “The company has cut roughly 30,000 corporate jobs since last October.” But: “Amazon, he noted, is hiring 11,000 interns and new grads this year.” // That’s one way to cut pay and claw back RSUs. Yay capitalism.
• Do Coding Agents Really Understand Enterprise Java? Or Are We Measuring the Wrong Things? – “Those are precisely the kinds of repository-wide reasoning tasks that remain difficult for current AI agents.”
• Viability of local models for coding – 🤖: “local models are viable for autocomplete and chat-style work, agentic coding is still messy and hit-or-miss, and nothing here is plug-and-play for developers unwilling to tinker. If you do tinker, a mid-size MoE model at 4-bit quantization (Qwen3.6 35B) is currently the sweet spot.”