Eight stories crossed the 200-point threshold on Hacker News today. Here’s what the community cared about and why it matters.
htmx 4.0.0 Released — 786 points
The htmx team shipped version 4.0 after eight months of development, and the headline change is an internal rewrite from XMLHttpRequest to fetch(). From a user perspective, the behavioral differences from 2.x are deliberately minimal — the team spent significant time rediscovering why the old implementation did things the way it did, then porting those decisions forward.
Three breaking changes matter: attribute inheritance is now explicit (you add :inherited to opt in, replacing the old implicit CSS-like cascade that confused everyone), event names are standardized to a htmx:phase:action pattern, and history no longer uses localStorage by default (re-fetching on back-nav instead, which sidesteps a whole class of bugs from third-party DOM mutations). New features include morph swaps via an improved idiomorph algorithm, an <hx-partial> tag for cleaner multi-target updates, and a new hx-live scripting solution that the team positions as an Alpine.js alternative with “HATEOAS-friendly reactivity.”
Notably, 4.0 is tagged as next on npm — they won’t force-upgrade anyone using unversioned CDN URLs until early 2027. The team also ships llms.txt and llms-full.txt for LLM consumption of their docs, which is either pragmatic or a sign of the times depending on your perspective. The htmax.js bundle packages htmx with popular extensions in a single file.
Source: htmx 4.0.0 announcement
Inception-Style Curved Map for Turn-by-Turn Directions — 591 points
Orbify demoed a navigation interface that warps the map into a curved, Inception-like perspective — the road ahead stretches vertically while the surrounding environment curves away. It’s a patent-pending approach (PCT/EP2026/058725) built on the PlayCanvas 3D engine.
The demo itself is a WebGL-heavy interactive page with WASD controls and mouse drag for panning/rotation. The visual effect is genuinely striking — it solves the problem of traditional top-down maps losing context about what’s actually ahead of you by creating a forced-perspective view that emphasizes the route corridor. Whether this is a real UX improvement or just a cool demo that would fatigue users on a 30-minute commute is an open question. The company is based in Norway and appears to be seeking pilots and investment.
Source: Orbify Demo
Boot a Virtual iPhone via Apple’s Virtualization.framework — 360 points
vphone-cli is an open-source tool (MIT license, 9.2k stars) that boots a virtual iPhone on Apple Silicon Macs using Apple’s Virtualization.framework and PCC research VM infrastructure. It requires macOS 15+, an Apple Silicon chip, and — critically — SIP/AMFI relaxation to allow private entitlements with unsigned binaries. That’s a significant security trade-off that limits this to research and development use.
The tool automates the full pipeline: IPSW download, boot chain patching, DFU restore, CFW installation, and first boot. You can manage VMs (create, clone, export/import, configure CPU/memory), and it cross-compiles a guest daemon (vphoned) via Xcode. The dependency list is substantial — python@3.13, aria2, wget, cmake, libusb, and more — reflecting the complexity of the iOS boot chain. This is clearly a security research tool, not something for casual app testing.
Source: vphone-cli on GitHub
I Accidentally Turned LLM Memory into Program Analysis — 259 points
Security researcher Jordy Zomer hit a wall using LLM agents for vulnerability research: after a few hours, models lose track of established facts, suggest already-ruled-out approaches, and confidently reason from invalidated observations. The standard RAG approach — embed old conversations, retrieve relevant chunks — didn’t satisfy because it relies on the LLM to re-derive conclusions from retrieved memories every time, which fails when earlier observations get corrected.
The insight: this is a program analysis problem, not a memory problem. Zomer built Lemmalog, a Datalog engine for LLMs. The LLM handles the fuzzy extraction work (converting debugger output, source code, and natural language into structured facts), while Datalog handles deterministic derivation and — crucially — automatic retraction. When a fact changes, all dependent conclusions update automatically rather than requiring the LLM to notice the inconsistency.
This is genuinely clever. The retraction problem in Datalog (you can’t just delete a fact if other derivations still support the same conclusion) maps perfectly to vulnerability research where multiple observations might independently support the same finding. The approach separates what LLMs are good at (fuzzy understanding) from what they’re bad at (maintaining consistent logical state over long sessions).
Source: I accidentally turned LLM memory into program analysis
EVE Online Moves to Python 3 — 232 points
EVE Online’s codebase has been running on Stackless Python 2.7 since 2010 — sixteen years on the same version. CCP Games (now under Fenris Creations) announced the migration to Python 3 has begun, with first changes deployed to the Singularity test server. The game launched in 2003 on Stackless Python, whose lightweight “tasklets” let a single server node handle thousands of concurrent pilots.
The migration rationale is straightforward: Python 2.7 hit end-of-life in 2020, modern libraries and tools are Python 3-only, and recent Python 3 releases bring significant performance improvements. The team’s definition of success is telling — “it should be completely unnoticeable, aside from the occasional moment when something runs more smoothly.” They’re not promising new features from the migration, just a healthier foundation. The fact that EVE stayed on Python 2.7 for 16 years is either a testament to Stackless Python’s stability or a warning about technical debt accumulation, depending on your perspective.
Source: The Move to Python 3 Begins!
Samsung’s Processing-in-Memory (PIM) — 219 points
Samsung presented their LPDDR5X-PIM at Hot Chips 2026, embedding MAC (multiply-accumulate) units directly into DRAM banks. Each of the 16 banks gets a PIM block that can access its attached DRAM without going through the external bus, collectively utilizing 614 GB/s of internal bandwidth versus 76.8 GB/s for standard external accesses. That’s an 8x improvement in raw bandwidth for compute operations.
The PIM blocks support INT8, FP8, and 4-bit formats, delivering 2.4 TOPS per chip at 4-bit precision. Eight LPDDR5X chips together would hit 9.6 INT8 TOPS — roughly matching Intel Meteor Lake’s NPU, but requiring 128 GB of memory (eight 16 GB chips). The key design choice is staying within the standard LPDDR5X protocol, so the chip functions as normal memory when PIM isn’t active. This is the pragmatic path — no new interconnect, no host-side changes, just compute that happens to live where the data already is. The question is whether the per-chip TOPS figure is compelling enough for real workloads when discrete NPUs exist.
Source: Hot Chips 2026: Samsung’s Processing-in-Memory (PIM)
The Internet Is Kind of a Predatory Cesspit Now — 209 points
Stephen Diehl’s essay argues that the internet’s transformation from a scrappy human-built commons into an “industrial system for manufacturing weakness at scale” is now complete. The grift economy isn’t a peripheral problem — it’s the organizing principle. Platforms are reinforcement-learning loops that continuously experiment on human weakness, optimizing for attention and retention. The consumer, salesman, and product have “collapsed into the same exhausted person.”
The piece is strongest when it describes the participatory nature of the grift economy: aspiring influencers with 47 followers are free labor for platforms and cheap distribution for course sellers. The affiliate marketer buys a course about affiliate marketing, then recovers costs by selling the same course. The life coach coaches new life coaches. “The pyramid is social before it is financial.” Diehl’s observation that satisfaction is churn and misery is recurring revenue cuts to the core of why these systems resist reform — they’re working exactly as designed.
Source: The Internet Is Kind of a Predatory Cesspit Now
Glacier Mice — 203 points
Glacier mice are colonies of multiple moss species found on glaciers in Alaska, Chile, Greenland, Iceland, Svalbard, Uganda, Venezuela, and several Subantarctic islands. They form free-moving globular structures that host nematode worms, springtails, and water bears (tardigrades). They move across the ice at roughly 2.5 cm per day, and — here’s the interesting part — their movement appears to be non-random, exhibiting herd-like behavior that doesn’t correlate with wind or slope direction.
The leading hypothesis for their movement: the dark moss absorbs solar energy on the sun-facing side, creating a small depression as the ice melts beneath it. The moss ball rolls into the depression, and the process repeats, creating gradual directional motion (southward in the Northern Hemisphere, northward in the Southern). Accelerometer studies confirm they actually rotate and roll rather than sliding, exposing all surfaces over time. They persist for six years or longer, creating self-contained micro-ecosystems on otherwise barren ice. First described in 1950 by Icelandic meteorologist Jón Eyþórsson.
Source: Glacier mice — Wikipedia
Today’s Throughline
Three themes dominate today’s front page. First, infrastructure that works shouldn’t be replaced just because it’s old — EVE Online ran on Python 2.7 for 16 years and htmx deliberately kept XMLHttpRequest for years after fetch() existed, both waiting until the migration justified itself. Second, the separation of concerns is having a renaissance — htmx 4 splits inheritance into explicit opt-in, the LLM/Datalog paper splits fuzzy understanding from deterministic reasoning, and Samsung’s PIM splits compute from the host processor. Third, the internet’s incentive structures are broken — Diehl’s essay and the Orbify demo (cool tech seeking a business model) both reflect a landscape where the gap between genuine innovation and extractive grift keeps widening.