This evening’s batch covers the release of Iroh 1.0’s key-based networking layer, Fox’s acquisition of Roku, Salesforce buying Fin for $3.6B, and Hetzner tripling dedicated server prices. On the research side, two separate Alzheimer’s studies — one using a copper-transport drug, the other high-dose psilocybin — report results that are promising but preliminary. Jane Street reverses 25 years of skepticism toward formal methods, OpenRouter launches a multi-model deliberation API, and a takedown analysis shows Rio de Janeiro’s “homegrown” LLM is a weighted merge of two existing models.

---

AI & Tech Policy

Launch HN: Drafted (YC P26) – Models for residential architecture

Summary: Drafted, a YC P26 startup, trains specialized models that generate complete residential architecture from structured design constraints including square footage targets, lot boundaries, footprint shapes, room placement preferences, and spatial relationship rules. The system produces floor plans with matching exterior elevations in seconds, offering 2D and 3D visualization, interior furnishing, material experimentation, and export to CAD and PDF for the downstream pre-construction workflow. Custom home design typically costs $10,000–$50,000 and takes months; Drafted targets the majority of homes currently built without any architectural involvement at all.

HN Discussion: The Launch HN thread had not yet accumulated comment discussion at the time of selection.

Apple Foundation Models

Summary: Apple’s Foundation Models framework, opening in iOS 27 and macOS 27, exposes a public LanguageModel protocol that lets third-party cloud providers like Anthropic and Google plug their models into Apple’s on-device AI infrastructure. A Swift package integrates Claude as a server-side language model, letting developers invoke any registered provider through a unified API surface. Apple also maintains on-device models running on the Neural Engine, creating a two-tier system of local and cloud model access behind a single developer abstraction.

HN Discussion: Commenters saw this as Apple commoditizing LLMs while retaining UX control and reinforcing their hardware business. Storage bloat was flagged as a concern if multiple apps each download separate on-device models, with no apparent shared-model mechanism. Some developers had hoped Claude Code features would run locally on the Neural Engine rather than through the server-side integration actually delivered.

Openrouter Fusion API

Summary: OpenRouter’s Fusion API runs a multi-model deliberation panel where several expert models analyze a prompt in parallel with web search enabled, then a judge model synthesizes responses into a structured analysis covering consensus, contradictions, partial coverage, unique insights, and blind spots. Quality and Budget presets control the panel composition, and developers can override the panel and judge models entirely via plugin configuration. Pricing is the sum of all panel-member completions plus the judge call, positioning Fusion for high-stakes research rather than routine queries.

HN Discussion: A developer who built a similar system reported that model-judges tend to evaluate based on resemblance to their own preferred answer rather than actual quality, adding rounds without genuine improvement. Benchmarking showed Fusion at roughly 7x slower and 4x more expensive than calling a single top-tier model directly. Theoretical discussion questioned whether SFT and RL shift a model’s knowledge distribution enough that cross-model aggregation provides real signal beyond what one model captures alone.

Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding?

Summary: An Ask HN thread asked whether developers have fully replaced cloud-based Claude or GPT with local models as their primary coding tool, specifically requesting setups and tokens-per-second benchmarks for production daily use rather than side experiments. The question targets the practical inflection point where local inference becomes fast enough and capable enough to serve as a daily driver for professional software development.

HN Discussion: One developer reported running DeepSeek V4 Flash on dual RTX Pro 6000 Blackwell GPUs at 160 tok/s, using an automated write-and-review pipeline rather than interactive coding. Another described MiniMax 2.7 on a Strix Halo at 30 tok/s generation as usable but slow, noting that frontier-quality models like M3 at 460B parameters cannot fit in 128GB of RAM. Discussion touched on whether per-prompt RLHF fine-tuning could eliminate the behavioral quirks that make general-purpose models frustrating for individual workflows.

Anthropic’s Safety Superpower

Summary: Ben Thompson’s Stratechery essay argues that Anthropic’s safety positioning has become a strategic moat, leveraging safety narratives to position itself as the indispensable gatekeeper for frontier AI deployment. The piece covers ITAR export controls applied to Anthropic’s Mythos model restricting access to US citizens and green-card holders only, and examines how guardrail-protected models can be jailbroken once released publicly. Thompson warns that Anthropic’s leadership effectively claims final say over AI governance broadly, which combined with assertions about AI conducting all economic activity amounts to seeking unprecedented private power.

HN Discussion: Critics argued the thesis collapses because Anthropic’s models get distilled into free Chinese alternatives within months, proving the bottleneck is compute and data rather than model weights. The ITAR restrictions were seen as proof Anthropic lacks real gatekeeping power, since resorting to blunt nationality-based export controls is what you do when you cannot technically prevent knowledge transfer. Commenters flagged the central concern that if one company claims sole authority over AI deployment while also claiming AI will drive all economic activity, that amounts to power over everything.

Rio de Janeiro’s “homegrown” LLM appears to be a merge of an existing model

Summary: The municipality of Rio de Janeiro released Rio-3.5-Open-397B as a homegrown Qwen3.5 fine-tune, but tensor-level analysis shows every weight is a consistent 0.6 × Nex-N2 Pro + 0.4 × Qwen blend across all 60 layers, to thousands of standard deviations. Nex-N2 Pro was released about a week before Rio’s model, and the weighted merge was identified through straightforward statistical comparison. The finding implies that Rio’s benchmark improvements over base Qwen may come entirely from incorporating Nex-N2 Pro’s weights rather than from novel training by the municipal team.

HN Discussion: Commenters noted the remarkable robustness of large transformer models, where a simple linear interpolation of every weight across all layers can enhance rather than degrade performance. A charitable interpretation suggested Rio may have used on-policy distillation absent from the uploaded weights, with the Qwen base disclosed but Nex Pro not. The irony was highlighted that open-source forensic analysis caught the reuse within days, since open weights inherently enable this kind of audit.

Show HN: Can Europe train a frontier AI model on the compute it owns?

Summary: A GitHub project called Euromesh provides a sourced model and short report investigating whether Europe can train a sovereign frontier AI model using only the public compute infrastructure it already owns, rather than waiting years for gigawatt datacenters to connect to the grid. The project aggregates publicly documented HPC and GPU resources across European institutions to assess whether distributed training at frontier scale is feasible today. The analysis arrives amid Europe’s broader debate over AI sovereignty, complicated by the EU AI Act and cross-border coordination challenges.

HN Discussion: Skeptics argued the real problem is organizational: Europe lacks the capital aggregation and cross-border coordination that the US uses to build large-scale models, making fragmentation the binding constraint rather than raw FLOPS. The EU AI Act, data privacy regulation, and cultural risk aversion were cited as structural barriers that would prevent frontier training even with adequate hardware. Distributed training for frontier models was acknowledged as technically possible but practically difficult, with projects like Glasswing illustrating the coordination problems.

---

Security & Privacy

Show HN: Exploiting Slack’s video embeds to achieve E2EE communication

Summary: A developer discovered that Slack’s Block Kit video block accepts any URL without runtime content verification beyond checking for a 2xx or 3xx HTTP response, effectively creating an unrestricted iframe inside the Slack client. The exploit builds an end-to-end encryption layer on top of this: a client-side app uses browser crypto APIs to generate key pairs, encrypts the private key server-side, and uses the video iframe to perform decrypt and sign operations that the Slack server never observes. This turns Slack’s video embed system into a covert E2EE messaging channel, since the server only ever handles encrypted blobs and iframe rendering.

HN Discussion: A commenter drew parallels to the WeChat and Alipay mini-app ecosystems, where fully-featured applications run inside messaging clients, suggesting the Slack video block is a similar but unintended app platform.

---

Business & Industry

Fox to buy Roku

Summary: Fox has agreed to acquire Roku, placing a major media company in direct control of the streaming OS running on tens of millions of American household televisions. The deal gives Fox ownership of Roku’s advertising platform, content channels, and the service-agnostic streaming interface that has been a key hardware differentiator. The central question is whether Roku’s platform neutrality toward third-party streaming services can survive ownership by a company that also owns competing content.

HN Discussion: Long-time Roku customers expressed pessimism that Fox ownership fundamentally compromises the platform’s service-agnostic architecture. Antitrust concerns were raised about a single media company controlling the TV hardware interface for an estimated 30–50% of American households. Some users threatened to permanently disconnect their Roku TVs, while others noted confusion about which corporate Fox entity is the buyer following the Disney-era corporate split.

Salesforce to Acquire Fin (formerly Intercom) for $3.6B

Summary: Salesforce signed a definitive agreement to acquire Fin, formerly Intercom, for approximately $3.6 billion, adding Fin’s AI customer agent platform to its enterprise cloud portfolio. Fin’s proprietary Apex AI model is purpose-built for customer support and claims resolution rates that outperform frontier commercial models across live chat, email, WhatsApp, SMS, phone, and Slack channels. The acquisition positions Salesforce to compete with Sierra, the $15.8B AI support company founded by Benioff’s former co-CEO Bret Taylor, alongside Decagon at $4.5B.

HN Discussion: Commenters mapped the competitive dynamics: Marc Benioff acquiring Fin to compete directly with Sierra, started by his ex-co-CEO Bret Taylor. Some questioned the long-term viability of helpdesk SaaS as businesses train their own local support agents using open models like Gemma. Positive AI customer-service experiences were reported — one Starlink caller rated it above 95% of human interactions — though execution quality was seen as the deciding factor rather than the underlying technology.

---

Tech Tools & Projects

Google Flight Simulator

Summary: Google released an experimental flight simulator mode within Google Earth that lets users fly around the world using the platform’s 3D satellite imagery and terrain data. The feature runs in the browser through the Maps Platform rendering infrastructure and is documented in the Google Earth developer docs. Google Earth also supports generated 3D buildings, though that feature reportedly originated from Sidewalk Labs and may have been deprioritized after years of development.

HN Discussion: Pilots and flight-sim enthusiasts criticized the flight model, saying the controls demonstrate a fundamental misunderstanding of how airplanes actually work. Nostalgic commenters recalled using the old Google Earth plugin for browser-based 3D flight data visualization decades ago. While not competing with Microsoft Flight Simulator, users found the Google Earth VR experience and flight mode genuinely charming as a lightweight exploration tool.

Asciline – real-time ASCII video rendering engine

Summary: Asciline is a real-time ASCII video rendering engine that streams binary-encoded frames via WebSockets for ultra-low-latency playback at 30 FPS using HTML5 Canvas and requestAnimationFrame. The engine processes video into character-based output on a server and pushes rendered frames to a browser client, minimizing client-side computation. The project is built without heavy client dependencies, relying on the WebSocket protocol and canvas rendering for the display layer.

HN Discussion: Commenters noted the project drew skepticism for suspiciously rapid GitHub star growth, with questions about whether genuine interest or coordinated promotion was responsible. Critics pointed out the README contains AI-generated padding and that the system renders using fixed characters with only color changes rather than true ASCII brightness mapping, making it closer to aalib/libcaca but vibe-coded. The practical use case was questioned, with one commenter arguing that visual coolness should be the priority for an ASCII video tool.

Show HN: I wrote a C++ ray tracer from scratch without AI

Summary: A developer built a C++ path tracer called Luz entirely from scratch with zero third-party libraries, starting from a ray tracer assignment at the 42 coding bootcamp at age 17. The project was developed consistently for over a year, then sporadically revisited, implementing physically-based rendering without relying on existing graphics frameworks. The author later used AI tools for code cleanup and additional features, though the initial architecture and core implementation were entirely hand-written.

HN Discussion: Commenters noted the slight contradiction between the title claiming “without AI” and the README admitting AI was later used for cleanup and feature additions. Building a ray tracer was described as a rite of passage in graphics programming, with many practitioners citing the popular raytracing.github.io tutorial series or Sebastian Lague’s video series as their entry point. The enduring appeal was attributed to ray tracing’s conceptual simplicity combined with deep implementation challenges around materials, sampling, and performance.

Ported my C game to WASM, here’s everybug that I hit

Summary: A game developer documented every non-obvious bug encountered while porting a C game engine — using bgfx, SDL2, miniaudio, and cimgui — to WebAssembly via Emscripten. The root cause of most bugs was WASM’s 32-bit address space: raw pointers in serialized asset structs changed from 8 bytes to 4 bytes, silently corrupting pak file deserialization. Other issues included needing a 32-bit native build for debugging, audio API differences, and Emscripten-specific build pipeline quirks.

HN Discussion: Commenters strongly recommended proper serialization and deserialization instead of dumping structs directly to files, noting that compiler padding differences cause the same portability problems on native platforms too. Another developer described porting a 20-year-old DirectX 9 codebase to WASM via bgfx, with the fixed-function-to-shader conversion being the bulk of the work. Frustration was expressed about WASM shipping with 32-bit pointers despite being a modern technology, with questions about why the design repeated historical mistakes.

---

Web & Infrastructure

Iroh 1.0

Summary: Iroh 1.0 is released as a stable networking library built around addressing devices by cryptographic keys rather than IP addresses, enabling connections that persist across network changes and NAT boundaries. The public relay infrastructure has handled over 200 million endpoints in the past 30 days alone, with production uses spanning distributed AI training, video streaming, real-time sync, IoT, and payments. The Rust-based system supports IPv4, IPv6, and relay transports with QUIC connections, providing Tailscale-style hole punching as an embeddable crate rather than a separate VPN service.

HN Discussion: A production user described Iroh as “Tailscale-style hole punching as a Rust crate” with additional P2P capabilities layered on top. Developers clarified the transport abstraction design, explaining they avoid feature-flag bloat by making non-standard transports like WebRTC, BLE, and LoRa pluggable. Skeptics questioned the fundamental value proposition, arguing IP with DNS and IPv6 already solves connectivity adequately and that a pricing page for a protocol seems contradictory.

Hetzner increased dedicated server prices 3-4x

Summary: Hetzner has dramatically increased bare metal dedicated server pricing for the second time in months, with the AX102 jumping from €124 to €454 and the AX162 (256GB) from €244 to €844 per month. This follows a prior approximately 30% increase earlier in the year, making the cumulative increase 3–4x for popular configurations. The hikes end Hetzner’s long-standing position as the budget dedicated hosting leader and significantly close the cost gap with hyperscaler cloud providers.

HN Discussion: Commenters attributed the increases to real hardware cost inflation, particularly for RAM and storage, and questioned how hyperscalers like AWS and GCP manage supply-chain leverage to keep cloud pricing stable. The sheer magnitude was considered extraordinary, with multiple commenters calling a 3x jump wild compared to expected 25–50% adjustments. One take was that after countless “I saved 10x by moving to Hetzner” blog posts, the company was simply capturing value it had been leaving on the table.

Show HN: Nxui – Copy-paste animated UI components for Vue

Summary: Nxui is a collection of over 180 copy-paste animated Vue 3 components built with Tailwind CSS v4 and motion-v, the Vue port of Framer Motion. Components include text animations, visual effects, animated backgrounds, buttons, and hero sections, all with dark mode support and TypeScript via script setup. Components install through the shadcn-vue CLI or by copying source directly, following the copy-paste distribution model rather than adding an npm package dependency.

HN Discussion: The Show HN thread had not yet accumulated comment discussion at the time of selection.

---

History & Science

Making glass-to-metal seals for homemade vacuum tubes

Summary: An electronics hobbyist documents the process of creating glass-to-metal seals needed to pass electrodes through the glass envelope of homemade vacuum tubes. The technique exploits copper’s red oxide bonding to borosilicate glass, where the bond proves stronger than bulk glass itself, and demonstrates vacuum retention using a tesla coil to check residual gas ionization. The writeup covers practical details including matching thermal expansion coefficients between metal and glass, sourcing premade tube stock, and using a rotary vane pump to achieve working vacuum levels.

HN Discussion: Commenters debated whether the achieved vacuum level is actually sufficient for a functional triode, with some questioning whether industrial tube manufacturing used significantly harder vacuums. Inductive coupling through glass was suggested as an alternative to physical feedthroughs for high-voltage applications, avoiding the sealing challenge entirely. The loss of historical manufacturing knowledge from the vacuum tube era was noted, with questions about how mass-production techniques from the mid-twentieth century could translate to modern hobbyist scale.

Teenagers Stayed Overnight at Their School and Found Hidden Ancient Roman Ruins

Summary: During a January 2021 school occupation protest against extended remote learning, students at a Roman high school across from the Colosseum discovered a well-preserved ancient Roman villa sealed in a basement beneath the building. The villa dates to the mid-second century CE and contains extensive graffiti spanning more than a century scrawled across its walls. Archaeologists plan to eventually open the sprawling subterranean space to the public.

HN Discussion: Expatriates described the common Italian experience of construction projects halted for years when Roman ruins surface, with one school expansion uncovering a catacomb that nuns reportedly kept quiet. A Terry Pratchett passage about Ankh-Morpork’s layered cellars drew parallels to how Rome is literally built on repeated layers of earlier construction. American commenters expressed awe at the density of ancient history visible throughout the city, contrasting it with the comparatively young historical record in North America.

There Is(Ǝ) – Such That (∋)

Summary: A creative coding project defines a small domain-specific language for building clock-like generative art compositions using mathematical set notation and bird-themed glyphs. The DSL operates on a circular canvas where vectors of time — hours, minutes, seconds — are represented as ravens, crows, and mag-pies, with operations for scalar extraction, vector modification, and visual glyph rendering. Built during a residency at the Recurse Center, the project includes composition patterns called Habitats that arrange glyphs and vectors into functional visual pieces.

HN Discussion: Mathematically trained readers were confused by the use of the set-membership symbol ∋ to mean “such that,” noting the unconventional choice is never explicitly justified. Commenters loved the concept of a bird-and-math clock DSL but found the gap between the poetic framing and the mechanical specification hard to bridge without more connective explanation. The abstract writing style led some to wonder whether it was AI-generated, though the project itself was recognized as a genuinely creative visual graph layout system.

---

Academic & Research

Copper transport drug restores memory and clears toxic Alzheimer’s proteins

Summary: Monash University researchers report a copper-transporting drug compound that restored memory and cleared amyloid-beta plaque buildup in mouse models of Alzheimer’s disease. The compound works by restoring intracellular copper levels, which disrupts the aggregation of misfolded proteins associated with neurodegeneration. The drug has already passed safety evaluations for other human conditions, potentially enabling a faster transition into clinical trials than would be possible for a novel compound.

HN Discussion: Commenters cautioned that amyloid-directed therapies have repeatedly failed in human trials despite strong mouse evidence, citing Derek Lowe’s decades of documented skepticism. Multiple commenters with family members affected by Alzheimer’s noted the disease has distinct subtypes and that current plaque-removing drugs show questionable efficacy in actually slowing cognitive decline. The mouse-to-human leap was the central concern, with one commenter emphasizing how little is still understood about the disease’s heterogeneity.

Improvement in advanced Alzheimer’s disease following high-dose psilocybin

Summary: A case report published in Frontiers in Neuroscience documents transient multidomain functional improvement in a patient with advanced Alzheimer’s disease following administration of high-dose psilocybin-containing mushrooms. This is a single-patient case report rather than a clinical trial, meaning the observed improvement could reflect spontaneous remission or other confounders rather than a treatment effect. The improvement was short-lived, and the biological mechanism by which psilocybin might affect neurodegeneration remains unclear.

HN Discussion: Commenters criticized the submitted title as heavily editorialized, stressing that a single case report is far from evidence of a successful treatment. The phenomenon of terminal lucidity — where Alzheimer’s patients spontaneously improve briefly before death — was raised as a possible confounder regardless of any intervention. Discussion referenced ongoing drug development targeting the same serotonin receptors as psilocybin, aiming to replicate potential cognitive effects without inducing a psychedelic experience.

How to build a virtual cell and biology scaling laws

Summary: Markov Biosciences, a San Francisco startup, is applying the “bitter lesson” from AI to biology: that large unbiased datasets and the right training objective outcompete models with hard-coded biological priors. Founder Adam Green argues that single-cell RNA-seq data should be treated as a ranking problem rather than raw counts — a concept traceable to a 1927 psychophysics paper — and that observational data alone enables clean scaling laws for virtual cell models. Their virtual cells pre-trained on plain observational data show monotonically improving performance on predicting unseen perturbations as models grow, beating state-of-the-art models built specifically for perturbation prediction.

HN Discussion: The HN thread had not yet accumulated comment discussion at the time of selection.

Formal methods and the future of programming

Summary: Jane Street’s Yaron Minsky reverses 25 years of institutional skepticism and announces the firm is now investing heavily in formal methods, driven by the thesis that AI-generated code shifts the bottleneck from code creation to verification. The post acknowledges that formal-methods costs remain very high — citing seL4 as a multi-year effort — but argues the cost-benefit equation fundamentally changes when AI floods development with code requiring verification. Jane Street’s approach leverages their existing OCaml type system as a bridge to heavier formal verification, particularly for financial systems where correctness is paramount.

HN Discussion: A proof-of-correctness veteran recounted decades of progress from early SAT solvers through the Boyer-Moore prover, emphasizing that lemma suggestion remains the key human bottleneck. Developers reported using expressive Scala 3 type systems to create compile-time proofs that prevent AI agents from producing low-quality code through what they termed “noun accretion.” Frontier models like ChatGPT-5.5 were noted as surprisingly effective at completing manual proofs in Rocq/Coq, sometimes proving in minutes what takes a human hours, making AI simultaneously the verification problem and part of its solution.

---

System Administration

My Homelab AI Dev Platform

Summary: A homelab operator built an AI-assisted development platform using OpenCode Web UI with Git integration, where AI pushes changes as pull requests that are manually approved before GitOps deployment through Arcane. The setup manages roughly a dozen Docker Compose stacks and uses AI to summarize release notes for container updates, reducing hours of manual review to minutes. The author migrated away from Claude Code due to token-limit cost pressures and now runs OpenCode as a server with persistent coding sessions synced across devices.

HN Discussion: Commenters shared comparable setups using n8n, Git, ArgoCD, and k3s with local models like Qwen and Gemma for automated workflows. Some extended the pattern with agent-driven PR creation and Helm-based deployment tracking through tools like ReARM. Questions focused on inference hardware and model selection, with interest in running OpenCode inside Proxmox LXC containers with Discord integration via Kimaki.

Exploring building a tiny FUSE filesystem

Summary: A detailed walkthrough of building “magicfs,” a small but functional FUSE filesystem that mounts at /magic with a real backing store separating metadata from file contents. The filesystem stores names and inode numbers in a metadata.json file while file contents live as plain blob files, illustrating the request-loop nature of FUSE operations and deliberate omissions around write ordering and kernel caching behavior. The post covers the Unix everything-is-a-file philosophy and how FUSE extends it to userspace implementations of lookup, read, write, and stat operations.

HN Discussion: A glibc maintainer noted that the glibc test suite already uses tiny FUSE filesystems for edge-case testing, including one that fabricates files to exercise O_CREAT|O_EXCL failure paths. Commenters appreciated the educational approach of building a minimal version to understand a complex system. The kernel FUSE documentation was reported as a dead link, reflecting broader issues with FUSE documentation discoverability.

The only scalable delete in Postgres is DROP TABLE

Summary: PlanetScale argues that large batch DELETE operations in Postgres are inherently unscalable because they don’t immediately reclaim disk space, add write amplification through WAL logs, and create bloat through dead tuples requiring VACUUM cleanup. The recommended approach is schema designs that allow bulk data removal as DROP TABLE or TRUNCATE operations, particularly using partitioning so old data can be dropped by detaching and dropping entire partitions. The post traces the mechanics through Postgres MVCC internals: every DELETE creates a dead tuple, autovacuum must clean them, and during that window table bloat degrades both query performance and index efficiency.

HN Discussion: Critics called the headline clickbait, arguing that DELETE with well-tuned autovacuum works fine at terabyte scale and that a manual VACUUM after large deletes reclaims space promptly. The point that deletion is as expensive as insertion was called obvious rather than counterintuitive, since both operations must write logs, update indices, and replicate. Experienced DBAs warned that DROP TABLE requires an ACCESS EXCLUSIVE LOCK, which can block all queries if the lock queue stalls, making it risky for production systems without careful planning.

---

Other

TinyWind: A pixel pirate sailing game with real wind physics (380k+ kms sailed)

Summary: TinyWind is a browser-based pixel-art pirate sailing game that simulates real wind physics, with players having collectively sailed over 380,000 kilometers. The game features two free-to-play modes playable directly in the browser and is currently in playtest with around 245 active captains. The developer is actively seeking playtest feedback from the community during this early stage.

HN Discussion: The developer engaged directly in the thread, inviting players to try the two free game modes and requesting feedback during the playtest phase.

CrankGPT

Summary: CrankGPT is a satirical product concept for a human-powered AI device operated by a hand crank or pedals, positioned as a privacy-preserving and climate-friendly alternative to cloud AI. The site presents tiered power levels from a 20W hand-crank for basic queries to 2000W+ pedal-powered setups for agent swarms and model fine-tuning, with genuine technical documentation underlying the parody. The marketing skewers tech companies for abandoning climate pledges to build gas-burning power plants for AI data centers, offering caloric computation as the absurdist alternative.

HN Discussion: Commenters appreciated the satire but found the actual technical documentation more interesting, noting it references real models that run at low wattage. Fitness enthusiasts connected the concept to rowing machines that display wattage output and calculated how much rowing would be needed to power inference on an M5 Max MacBook. The protest imagery was critiqued for conflating gas-burning smokestacks with nuclear cooling towers, which one commenter called representative of muddled anti-AI environmental arguments.

Firewood Splitting Simulator

Summary: A browser-based toy from the screen.toys collection that simulates firewood splitting with click-to-split mechanics and drag-to-rotate interaction for positioning logs. The simulator is designed as a mindless diversion rather than a realistic physics model, in the spirit of Goat Simulator and other casual browser toys from the same collection.

HN Discussion: Experienced firewood splitters critiqued the physics, noting that real splits push halves outward, the challenge is execution rather than target selection, and you must avoid knots and axe-handle collisions. Matthias Wandel pointed out multiple inaccuracies: the game lets you slice across the grain like bread, ignores the difficulty of standing up crooked logs, and allows cuts that would damage an axe handle. Commenters appreciated the thread as a refreshing break from AI-dominated discussion, enjoying nerdy gripes about simulator accuracy applied to a physical skill.