HN Engineering Weekly — Week 25, 2026

Architecture

Project Valhalla, explained: how a decade of work arrives in JDK 28

• phkahler (extends): The article understates what remains — this JEP covers value types only. Specialized generics (making ArrayList<Point> actually flat) are the part most practitioners care about for real-world collection performance.
• cpeterso (extends): The 64-bit atomic write constraint means classes with two int fields or a double may not flatten. The null flag also costs one bit. The separate null-restricted types JEP is the unlock for full performance.
• kaba0 (agrees): Valhalla changes == semantics — for value objects it now checks substitutability (same class, recursively equal fields) rather than reference identity. That's a breaking change at the language level, even in preview. Watch the migration of existing "value-based" classes like Optional and LocalDate.

There are no instances in ATProto

• pessimizer (challenges): Bluesky's relay is effectively a centralized instance. If Bluesky PBC controls the main relay and AppView, the architecture is "federated in name only" — most users won't run their own relay at any cost.
• danabramov (extends, author): Responds that a relay has been cheap to run ($20/month on Cloudflare) for over a year, and multiple community relays exist (Blacksky, Constellation). The protocol doesn't require Bluesky's relay.
• bnewbold (extends): The relay is intentionally "dumb" — it stores and forwards. Intelligence lives in the AppView, and anyone can build one. The cost argument applies to the relay specifically, not to the aggregation layer.

Zero-touch OAuth for MCP: Enterprise-Managed Authorization is stable

• hhamalai (extends): This is the missing piece for enterprise MCP adoption. The SAML/OIDC analogy is exact — before SSO standardized, every SaaS app had its own login wall. EMA is that standardization moment for AI tool access.
• jayphelps (challenges): Centralizing trust in the IdP and MCP host raises the attack surface. If the IdP is compromised, every connected MCP server is compromised. The security model needs defense-in-depth beyond SSO.
• kevincox (agrees): EMA also solves the personal-vs-corporate account mixing problem. Previously, MCP servers couldn't enforce that users connected with their work account — a real data-leakage risk for enterprise deployments.

LLMs are complicated now

• jordanb (extends): This is the "bitter lesson to feature-engineering lifecycle" playing out again. Early gains come from scale; later gains require careful architecture. Same pattern in every prior ML domain.
• charcircuit (challenges): The Llama 3 vs. Nemotron 3 Ultra comparison is unfair — GLM-5.2, released one week earlier, uses standard attention. Architectural divergence is expected between different model families, not an industry-wide trajectory.

So you want to define a well-known URI

SRE

Google Workspace is threatening to block Firefox access

• shkkmo (challenges): The remediate URL pattern suggests this is about Web Environment Integrity or device attestation, not browser capability. Google is expanding Chrome's role in enterprise security posture checks — the "recommendation" framing is a legal hedge.
• kmeisthax (extends): This follows a consistent pattern: Workspace features that "coincidentally" require Chrome — offline Docs, client-side encryption in Meet. The boundary between "not supported" and "actively blocked" has been eroding for two years.
• bogwog (agrees): If 80% of enterprise users come to feel they need Chrome for work, Firefox's market share becomes structurally untenable. A browser monoculture is itself the security risk Google claims to be preventing.

AURpocalypse now: a look at the recent AUR attacks

• ameliaquining (extends): Clarifies the attack vector precisely: this isn't anyone creating new malicious packages, it's the orphaned-package adoption mechanism. The risk is silent compromise on update of a package you already had. That's distinct from PyPI/npm's new-package injection pattern.
• jchw (agrees): The open-adoption model has long been low-hanging fruit for attackers. AUR has always carried "use at your own risk" status, but most users treat it like any other package repo and never read PKGBUILDs.
• nickjj (extends): yay v13+ now supports skipping recently-modified packages via its Lua extension system. Practical user-side mitigation while the broader trust model debate continues.

Let's Encrypt had a higher error rate for 90 minutes

• tptacek (extends): Let's Encrypt's incident patterns are notably benign compared to commercial CAs — when LE has issues, existing certs keep working and only new issuance is affected. Transparency on the extended ISP investigation is good operational practice.
• avianes (agrees): Thirty-one minutes of elevated errors is short enough that ACME clients with retry logic likely auto-recovered. Real risk is for certs expiring during the window with no retry buffer.

Satellite reveals immense scale of GPS signal tampering

• p_l (extends): GNSS timing affects far more than navigation: cellular towers, financial transaction timestamps, power grid synchronization, and scientific instruments all depend on GNSS time signals. Jamming that degrades position data often degrades timing data first.
• bobthepanda (challenges): Satellite detection may not distinguish military jamming (deliberate) from solar activity interference or equipment malfunction. The "tampering" framing may overstate intentionality in the data.
• topaz0 (extends): GNSS receivers are improving at detecting spoofed signals, but jammers are improving too. Space-based monitoring opens a new front — you can't jam the satellite measuring your jamming.

Performance

GPT-5.5 hallucinates 3× more than MIT-licensed GLM-5.2

• jackmott (challenges): AA-Omniscience specifically tests "knowing when you don't know" — a model that refused every question would score 0% hallucination. High calibration on this benchmark doesn't directly map to real-world usefulness in tasks where partial answers have value.
• hughw (extends): DeepSeek's 94% rate tracks with user experiences of it confidently fabricating API functions and historical facts. The root cause is the training philosophy: always have an answer, never admit uncertainty.
• slashdave (agrees): The "bitter lesson" — scale is all you need — is getting empirically challenged. Larger models memorize more but don't necessarily generalize or calibrate better.

Surprising economics of load-balanced systems

• bijowo1676 (challenges): The model assumes Poisson arrivals and an infinite queue. Real-world traffic has correlated bursts — retries, timeouts, thundering herd — that break the independence assumption and can cause the opposite of what the model predicts.
• mjb (agrees, Marc Brooker himself): Acknowledges the limitation directly: "exponential service time is especially wrong for real services, but the qualitative insight holds." The model is "reasonable, albeit wrong."
• lmeyerov (extends): Their GPU-backed analytics service is actively revising load-testing methodology from these insights. The non-obvious part is knowing which charts and metrics to focus on — p99.9 behavior at different server counts tells a different story than p50.

The token compression illusion: why I'm skeptical of RTK

• lackoftactics (agrees, author): Management pressure to cut LLM costs is causing teams to wrap every command in RTK without understanding the tradeoffs. "Now people are wrapping every possible command in rtk."
• compuficial (challenges, user): Defends RTK from direct experience — 3.7M tokens saved on 3.9M tokens of input is real savings. Accuracy benchmarks would be useful, but the tool works in practice.

Ask HN: Will programmers write more efficient code during the memory shortage?

Databases

Datasette Apps: host custom HTML applications inside Datasette

• simonw (extends, author): The sandbox blocks fetch() to arbitrary origins by design — only explicitly granted HTTPS origins are allowed. This is a deliberate choice to prevent data exfiltration.
• acaloiar (agrees): Favorable comparison to Observable and Streamlit — the self-contained pattern with direct SQL access is ideal for internal dashboards that should live alongside the data, not be deployed separately.

Show HN: Are You in the Weights?

• dnissley (extends): LLM training data is effectively a census of human notability, biased toward English-language internet content. Athletes and musicians score higher than scientists in the same objective tier of accomplishment.
• pmontra (challenges): The methodology conflates recognition with name frequency. High scores may just reflect how often a name appears in training data near positive-connotation text — not that the model "knows" the person.
• gs17 (extends): The tool inadvertently measures model censorship. Some controversial figures score lower not because they're less known, but because safety filters suppress detailed responses — the score is a function of both training data and RLHF policy.

This week's signal