I disagree! Just because I think that CMSes are still and will be a thing. I was thinking of telling my agent to post for me in a website, but I've lately seen how a coding agent actually can build a wordpress theme, create or migrate content and even manage it. Is it perfect? Of course not! Do we need everything AI at the moment? Not at all! For some, a content management system is still the best thing to deliver content. But, the counterintuitive bit here is that AI may strengthen the case for a CMS rather than kill it. Reynolds gets that right because the boring machinery of publishing still matters: permissions, workflows, APIs, content models, revision history, and the dull miracle of letting a non-technical person press Save without invoking an agentic séance. WordPress is the named battlefield, but the real issue is institutional memory. A CMS is rarely just a page editor. It is where years of decisions, mistakes, forms, redirects, taxonomies and strange little business processes accumulate. Reynolds’ 24-year archive makes that unusually visible, but plenty of organisations have quieter versions of the same problem. They just call it “the website”. Cloudflare’s EmDash example is the tell. Even the “spiritual successor to WordPress” still looks like a CMS because humans still need handles on the machine. The CMS isn’t dead. The lazy CMS sales pitch is :)

The gap between the headline and the evidence is the whole story here! The Vitalists want death treated as a moral error, yet the article is admirably blunt about the awkward bit: the effective anti-ageing treatments they want do not yet exist. That’s a fairly large hole in the launch deck. Hamzelou gets the important thing right by separating this from the usual wellness slop. Vitalism is not just another supplement label with shinier typography. It is a political project. Cheng’s argument is simple enough to sound unanswerable: life is good, death is bad, therefore society should prioritise defeating death. Neat. Too neat. The hard part is not agreeing that dying is unpleasant. The hard part is deciding what proof, safety, access and opportunity cost look like before the movement starts rewriting budgets. Death may be bad. Hype with a halo is still hype.

This piece gets one big thing right: great research needs room to look inefficient. Bell Labs’ magic, as the author tells it, wasn’t just money or clever hiring; it was the willingness to give serious people serious problems and then resist the managerial itch to poke them every fortnight for a dashboard update. That’s rare. It was probably rare then. Today’s tech giants have had more money than early AT&T, yet none has built another Bell Labs. Apparently, the harder problem is taste. Who gets freedom? Which problems deserve years? How much apparent idleness can an institution tolerate before someone with a spreadsheet starts twitching? Mervin Kelly is doing a lot of work in this story. Maybe too much. Modern technology companies love research branding, but their incentives often pull towards product cycles, platform advantage and defensible moats. Basic research becomes acceptable when it can be routed through hiring, PR, patents or cloud consumption. Bell Labs, in this telling, was less a lab than an institutional contradiction: commercially funded patience.

Usually, the people selling “just scale it” have not killed mathematics; they’ve changed its job description. Kvinge’s most distinctive claim is that maths may become less like an architect’s blueprint and more like physics after the experiment: a way to explain strange empirical regularities once the machine has already surprised us. That feels right. Annoyingly right. Deep learning has repeatedly made theory look overdressed for the occasion, particularly when more data, more parameters and more compute produce capabilities that existing explanations didn’t predict. But, if mathematics becomes post-hoc explanation, who funds it when the product already ships? A team chasing latency, margin and model leaderboard position may not pause for topology unless it exposes a failure mode, cuts compute or satisfies a regulator. AlphaFold3 is the useful irritant here. If a non-equivariant architecture can succeed on data with obvious symmetries, then the old instinct to bake every known structure into the model gets harder to defend. Still, compute is not infinite. Nor is trust. In domains where errors are expensive, the case for mathematical structure returns quickly...

What a perfect way to finish this edition, on reading slowly :) “Read the obits” sounds charmingly low-tech, and the article’s most distinctive claim is genuinely useful: creativity may improve when we feed the mind distant, unrelated material rather than another personalised drip of the already-familiar. That’s a clean jab at modern information habits - even for those that receive this newsletter everyday for the past 2 years. I do think that creative work often comes from making odd connections, and - even - obituaries are unusually dense packets of career pivots, accidents, migrations, inventions and obsessions. Curiosity takes time; time is usually the first thing we pretend to value, then remove it. Anyway... obituaries won’t make anyone a genius! They might make the next idea less obedient...