Five links I’ve had open in tabs all week that I think are worth your time.
AI makes you boring — Viktor Löfgren (Marginalia)
This one started as a Hacker News comment about the declining quality of “Show HN” submissions, and it grew into something sharper. The argument is that offloading thinking to an LLM doesn’t just produce mediocre output, it produces a mediocre thinker. Original ideas come from immersing yourself in a problem for a long time, which flat out doesn’t happen when LLMs do the thinking. You get shallow, surface-level ideas instead. You don’t build muscle using an excavator to lift weights. You don’t produce interesting thoughts using a GPU to think. I keep coming back to this piece because it articulates something I’ve been feeling for a while but couldn’t quite pin down.
Microsoft’s 10,000-Year Glass Storage Medium — Ars Technica
Microsoft Research published a breakthrough in Nature on glass-based data storage that could preserve information for 10,000 years. The new technique extends the technology from expensive fused silica to ordinary borosilicate glass, the kind found in kitchen cookware, which is a major step toward commercialization. The system uses femtosecond lasers to etch data into small glass slabs at a density of over a gigabit per cubic millimeter, with a maximum capacity of 4.84 terabytes per slab, requiring zero energy to preserve. The obvious joke is that nobody will have a reader for this in 10,000 years, but as a serious archival medium for the near future it’s genuinely fascinating.
Don’t Trust the Salt: Multilingual LLM Evaluation to Guardrails — Roya Pakzad (Humane AI)
This is a dense, technical piece but absolutely worth the effort. Roya Pakzad demonstrates a technique she calls “Bilingual Shadow Reasoning,” which steers a model’s hidden chain-of-thought through customized non-English policies, making it possible to bypass safety guardrails and evade audits, all while the output appears neutral and professional on the surface. The downstream implication is alarming: many closed-source wrappers built on top of major LLMs can embed hidden policy instructions, potentially facilitating censorship in authoritarian contexts, manipulating sentiment in marketing, or reframing historical events, while users are sold the promise of accurate, neutral summaries. This is the kind of AI safety research that actually matters.
Micropayments as a Reality Check for News Sites — blog.zgp.org
A fresh, pragmatic take on an old debate. The internet has dramatically diversified reading patterns. Today, readers routinely consume journalism from dozens of sources in a single month, but large portions of that readership generate little or no direct revenue for publishers. The author argues that Clay Shirky’s old argument against micropayments from 2003, based on mental transaction costs, doesn’t work so well anymore, and makes an interesting case borrowing from mobile gaming psychology: a two-step currency exchange that makes spending feel less like spending. Worth reading alongside the skeptical takes that have accumulated over the years.
How AI Is Affecting Productivity and Jobs in Europe — VoxEU / CEPR
If you’re tired of hot takes and want actual data, this is the piece. Drawing on survey data from more than 12,000 European firms, the authors find that AI adoption increases labour productivity levels by 4% on average in the EU, with no evidence of reduced employment in the short run. The catch is that the productivity benefits are unevenly distributed, medium and large firms experience substantially stronger gains, while smaller companies, which dominate Europe’s industrial landscape, lag significantly behind. The finding about complementary investments is especially interesting: an extra percentage point of investment in workforce training amplifies AI’s productivity gains by 5.9 percentage points, which says a lot about where the real leverage is.