Editorial: Today's thread runs from playful to urgent. One indie project turns GPU architecture into a hands-on puzzle; three policy and infrastructure stories highlight how software and platform choices ripple into real-world access and performance. Read for a mix of curiosity, UX frustration, and system-level consequences.
In Brief
A game where you build a GPU
Why this matters now: The mVidia project makes GPU architecture accessible by letting players assemble logic gates up to GPU building blocks, offering a tactile path into hardware concepts many software engineers rely on but rarely see end-to-end.
This Show HN toy — mVidia: build a GPU — is a puzzle-sandbox that walks players from truth tables to latches, capacitors, sense amps and other GPU-y pieces. Players praised the learning curve: one commenter called it a “really cool approach to teaching GPU architecture,” and comparisons to tactile titles like Turing Complete and Silicon Zeroes keep the reference frame clear.
"Really cool approach to teaching GPU architecture." — Hacker News commenter
The project is early-stage: users reported a steep jump in difficulty for some puzzles, a capacitor/enable-gate simulation bug, and requests for better hints. The author notes LLM assistance in development, which shows how these tools are helping indie creators prototype ambitious educational interfaces. The experiment matters because GPUs underlie gaming, graphics, and AI — hands-on learning could demystify parallelism for engineers and students.
How many products does Microsoft have named 'Copilot'?
Why this matters now: The audit shows at least 75 distinct "Copilot" labels across Microsoft, which is actively creating communication and support friction as AI features spread across products.
A reader compiled an interactive map and found the Copilot name applied to apps, features, platform integrations, a keyboard key and a laptop line — and couldn't find a clean pattern. The write-up is collected at How many Microsoft Copilot are there?.
"I couldn’t… because the name ‘Copilot’ now refers to at least 75 different things." — author
HN responses split between disbelief and strategic interpretation: some say this destroys the usefulness of the term for troubleshooting and product conversations, others argue Microsoft might be aiming for a seamless AI umbrella across services. The practical cost is clear — help desks, engineers, and users lose signal when a brand word no longer points to a specific product.
LLM Wiki — an "idea file" for long-term knowledge
Why this matters now: Andrej Karpathy's LLM Wiki pattern lays out a practical habit for turning ephemeral retrieval into a persistent, interlinked knowledge asset that compounds over time.
Karpathy proposes that instead of fetching sources on demand, you have an LLM ingest and maintain a markdown wiki that sits between you and raw documents: read, summarize, cross-reference, and continuously lint it. See the example gist.
"The LLM 'incrementally builds and maintains a persistent wiki— a structured, interlinked collection of markdown files that sits between you and the raw sources.'" — Karpathy
This approach trades immediate freshness for a persistent working memory you can query, which many power users already find addictive. The downsides are model drift and a new kind of maintenance debt: if the wiki becomes self-referential and nobody audits sources, errors can compound. Still, it's a useful operational pattern for teams and researchers wrestling with large, changing corpora.
Deep Dive
German implementation of eIDAS will require an Apple/Google account to function
Why this matters now: Germany’s eIDAS 2.0 reference architecture ties national digital-ID key issuance to hardware-backed attestation and live device vulnerability checks, meaning many phones will need platform attestation services (Apple/Google) to use the wallet.
The reference architecture spells out a Mobile Device Vulnerability Management (MDVM) flow: devices are monitored for OS and hardware vulnerabilities, and if a device is flagged, keys secured by the wallet are blocked. The spec language is explicit:
"the solution provides for monitoring identified vulnerabilities for the HKS and the operating system of user devices through a mobile device vulnerability management (MDVM during operation) ... if vulnerabilities are known for a user device that could compromise the user's authentication mechanism ... the use of keys ... is prevented." — German eIDAS wallet spec
On modern phones that practically means relying on vendor attestation signals like Android Key Attestation, Google Play Integrity, and Apple’s DeviceCheck/AppAttest — services that generally require a vendor account or platform services to be present and up-to-date. Hacker News commentary frames this as an effective dependence on two U.S. companies: account bans, vendor policies, or proprietary attestation changes could lock citizens out of national ID services.
There are alternatives — SIM-based Mobile-ID, external hardware tokens (YubiKey), or attestation solutions compatible with AOSP/GrapheneOS — but they either raise UX friction or require policy changes and vendor collaboration. Implementers say they’re aware and are prioritizing Android initially, but the specification’s current wording has already sparked debate about digital sovereignty and whether a government ID should hinge on a handful of platform vendors. This is an urgent policy design trade-off: security via hardware attestation versus accessibility and independence from dominant mobile ecosystems.
AWS engineer reports PostgreSQL perf halved by Linux 7.0, fix may not be easy
Why this matters now: Near-final Linux 7.0 appears to drop PostgreSQL throughput by roughly 50% on some arm64, high-core machines, potentially affecting cloud customers and systems shipping Ubuntu 26.04 LTS.
An AWS engineer’s tests (summarized by Phoronix) show PostgreSQL spending far more time in a user-space spinlock after a kernel change that restricts preemption modes — effectively removing PREEMPT_NONE as a valid default. The regression was bisected to that change and benchmarks on a Graviton4 96-core instance showed around 0.51x throughput. Peter Zijlstra, a kernel maintainer, suggested the fix is for PostgreSQL to adopt the rseq (restartable sequences) slice extension, shifting corrective action to userspace.
"The fix here is to make PostgreSQL make use of rseq slice extension." — Peter Zijlstra (quoted in the thread)
Options on the table are a kernel revert (restore PREEMPT_NONE), PostgreSQL changing to use rseq, or distro-specific workarounds. Each path has trade-offs: kernel reverts delay new scheduler behavior, userspace changes need careful testing and rollout, and distro patches create maintenance divergence. Commenters also note the regression seems concentrated on arm64 many-core setups (Graviton4) and wasn’t reproducible on an equivalent amd64 machine.
The practical advice for ops teams: if you're running PostgreSQL on many-core arm servers, test your workloads against Linux 7.0 pre-releases and coordinate with your distro/vendor for patches. The episode is a reminder that kernel defaults matter — they can silently change performance across a large ecosystem, and the fix may force difficult choices about whether responsibility lies in kernel land or in userland libraries and runtimes.
Closing Thought
A playable GPU teaches hardware by doing; an exploding product name erodes communication; a national ID spec collides with mobile platform realities; and a kernel tweak shows how low-level defaults can break high-level services. These stories share a theme: systems are social as well as technical — design choices ripple from UX to public policy to performance. When building or governing platforms today, ask not just whether something works, but who it makes dependent and how hard it will be to fix later.