Today’s thread is AI realism: a debate over whether billion-dollar startup compounding is a fair path to extreme wealth sits alongside a pushback on “everyone uses AI,” warnings that huge context windows are less useful than advertised, questions over whether Rio’s “homegrown” public LLM was mostly a merge, and a case for formal methods as AI makes proof-heavy programming more practical. Elsewhere, security and identity run through a Honda Civic infotainment exploit and a UK plan to ban under-16s from social media via age checks. Also in view: Swiss voters rejecting a population cap, an offline web archiving tool, and Python 3.14 garbage collection changes.
Paul Graham posted an essay arguing that becoming a billionaire is mathematically possible through startup compounding, and that founders can get there without cheating if they build something users love in a big market. Readers largely thought he dodged the real issue, which was not whether compounding exists, but whether anyone can fairly or morally end up personally holding that much wealth.
Gabriel Weinberg posted a rebuttal to the claim that “everyone is using AI for everything,” pulling together survey, telemetry, and public-opinion data that show usage is meaningful but far from universal. Readers largely agreed the hype outruns the numbers, while adding that AI is strongest in software and search-like tasks, weakest when bolted onto deterministic systems, and increasingly hard to avoid when vendors hide it inside existing products.
A security researcher showed that 10th-generation Honda Civic infotainment updates can be spoofed and flashed from a USB stick because Honda signed packages with Android’s widely known test key. That gives anyone with brief physical access to the front USB port a path to run custom code on the car’s head unit, and commenters split between calling it a serious security failure and a useful form of owner hackability.
A UK proposal to bar under-16s from social media set off a familiar split: many people like the goal of shielding kids from addictive feeds, but they see age checks and ID verification as the real story. The strongest reaction was that the policy is being sold as child protection while quietly normalizing broader online identity checks for everyone.
A blog post argues that giant LLM context windows are being oversold for coding work, because models often get worse long before they hit the advertised token limit. The comments mostly agree that long sessions degrade, but the useful takeaway is not "never use big context" so much as "treat context as a scarce resource and design your workflow around resets, docs, and subagents."
A GitHub issue alleges that Rio de Janeiro’s newly announced public LLM was not a homegrown post-trained model at all, but mostly a straight weight merge of Nex-N2 Pro and Qwen 3.5. The key dispute is not open-model reuse, but whether the team falsely claimed original training work and only added the merge disclosure after being caught.
Swiss voters rejected a proposal that would have forced the government to curb immigration once the country reached 10 million residents, a move widely understood as a challenge to Switzerland’s free-movement agreements with the European Union. Comments focused less on the result itself than on what is really driving anti-immigration politics: housing and infrastructure strain, wage pressure, and cultural backlash wrapped in a cleaner political package.
Jane Street published an overview of how it uses formal methods, theorem proving, and proof-aware languages to make software more reliable, and argued that AI may finally make this approach practical at broader scale. The comments mostly agreed the idea is powerful in well-defined domains, but pushed hard on where specs become the real bottleneck and where this still feels too expensive or too rigid for messy product work.
Kage is a new open source tool that crawls a website in a real browser, saves the rendered pages for offline use, and can package the result as a folder, a ZIM archive for Kiwix, or a self-contained executable. Readers mostly treated it as a modern HTTrack for JavaScript-heavy sites, with interest strongest around docs, wikis, flights, and long-term archiving.
A blog post explains why Python 3.14’s new incremental garbage collector was rolled back in 3.14.5 after it cut pause times but caused much higher memory use for some workloads. The comments mostly treat the revert as the right operational move, then widen into bigger complaints about Python’s packaging, governance, and release churn.
A 2023 paper on St. Paul’s 2021 rent control law argues that property values fell about 4.4% to 5.8% within nine months, shifting wealth away from landlords and homeowners and toward tenants. The comments mostly dispute what that result means, with many saying the short time window and Twin Cities-specific shocks make the paper a weak guide to long-run housing policy.
Phoenix released LiveView 1.2, adding colocated JavaScript and CSS for components plus other quality-of-life updates to its server-driven web UI framework. The comments mostly treated it as more evidence that the Phoenix and Elixir stack can deliver modern app UX with far less frontend sprawl than the usual React or Next.js setup.
A PlanetScale post argues that large deletes in PostgreSQL are usually the wrong primitive, because row-by-row DELETE creates a lot of extra database work while dropping a whole table or partition is nearly instant. The comments mostly agree on the mechanics but say the title overreaches: this is really about retention-style, time-partitioned workloads, not general CRUD apps.
Phoronix highlighted a ReactOS milestone: the open source operating system that aims to be Windows-compatible can now run the original Half-Life with real 3D acceleration on physical hardware, not just in a virtual machine. Readers treated it as a small but real proof that ReactOS can drive old Windows graphics stacks and legacy software on actual machines, even if it remains far from a modern Windows replacement.
A browser toy called Firewood Splitting Simulator let people chop virtual logs into neat pieces and stack them around a stump. Readers mostly treated it as a surprisingly polished bit of mindless fun, while wood splitters immediately turned it into a debate about how real log splitting actually works.
A blog post argues that many of Ruby’s most-loved features came from Lisp, then repackaged them in friendlier syntax and an object-oriented shell. Commenters mostly agreed that the comparison explains Ruby’s appeal, but pushed on where Smalltalk, Perl, macros, and pipeline syntax deserve more credit.
An Aptiv article argued that USB Power Delivery is turning USB-C into a common power standard for cars and other devices by scaling from small gadget charging up to 240 watts. The comments were less interested in the marketing pitch than in the messy reality of USB-C products that look standard but fail to charge correctly because vendors skip basic compliance.
A post from the zeroserve author says the project can now run a subset of Caddy configs and, in the author’s benchmark, serves simple HTTPS traffic at about 3 times Caddy’s throughput with much lower latency. The comments liked the engineering but mostly treated it as an impressive niche benchmark with missing features, awkward client-certificate behavior on the demo site, and open security questions around io_uring and JITed eBPF.
A developer released an experimental OrcaSlicer fork for FDM 3D printers that tries to strengthen weak layer-to-layer bonds by printing internal channels and then injecting molten plastic into them. The catch is big: the author says clean physical prints are not working yet, so readers focused as much on whether the materials process can work at all as on the slicer idea itself.