Hacker News

Top Story

OpenRouter is joining Stripe · OpenRouter Blog In a massive consolidation of the AI infrastructure layer, Stripe is officially acquiring OpenRouter, the web’s largest model marketplace and routing gateway. With OpenRouter processing over 10 trillion tokens a day from more than 10 million developers, Stripe is securing a central tollbooth in the post-AGI economy where inference is rapidly becoming the largest operating expense for internet businesses. It is a textbook marriage of two developer-obsessed abstractions that hide complex, fragmented markets behind a single, clean API.

Front Page Highlights

A decades-old bug in Knuth’s long division (TAOCP Vol II, Algorithm 4.3.1D) · Kolja’s Blog While writing a multiprecision arithmetic library, the author stumbled upon a 30-year-old algorithmic bug in Donald Knuth’s classic long division algorithm (Step D3), introduced during his 1995 MMIX overhaul. The bug stayed hidden for decades because it only manifests during division with an odd radix (like base-3 to hexadecimal conversion), while standard computer architectures operate on even radices where the overflow is masked. The author successfully claimed his hexadecimal dollar check (0x$1.00) from the Bank of San Serriffe. Along the way, they also exposed a legacy typo in LLVM’s APInt library that has preserved a related TAOCP errata bug since 2005.

A joke domain purchase turned in geopolitical warfare · sprocketfox.io What began in 2018 as a joke domain redirect for weather balloon tracking, sondehub.org, somehow spiraled into a bizarre saga of sky-surfing balloon warfare. The author chronicles dealing with clueless FAA supervisors who didn’t understand that weather balloons are unguided, navigating defensive Swiss electronics manufacturers, and mailing unpaid invoices to the US Department of War. The peak absurdity occurred when the author had to post in Ukrainian military chat rooms urging a deep-strike unit to run the wind-prediction script locally because their automated API scraping was triggering performance alarms on the server while they “surfed” the jet stream to target areas.

The Two Factions of C++ · herecomesthemoon.net A sharp, brutally honest dissection of the widening cultural rift in the C++ standard committee. On one side are tech-savvy firms with monorepos, versioned source control, and automated refactoring tools (like Google) who can migrate codebases in minutes; on the other is a massive, silent majority of legacy enterprises relying on pre-compiled binaries from the late 90s. By doubling down on absolute ABI backwards compatibility and rejecting modern concepts like “viral annotations” (lifetime tracking), the committee is desperately trying to preserve legacy code while driving modern systems engineers out of the ecosystem entirely.

PostgreSQL for Everything · Raphael Bauer’s Blog A highly compelling case for architectural minimalism by replacing half your stack with Postgres. Pushing back against the microservice sprawl and complex synchronization pipelines, the author explains how Postgres’s robust feature set—including GIN-indexed JSONB, TimescaleDB, pgvector, unlogged tables, and SELECT FOR UPDATE / SKIP LOCKED—allows a single, boring, rock-solid database instance to replace Redis, MongoDB, Clickhouse, Elasticsearch, and RabbitMQ.

Palomar: A registry of Lean verified mathematics · Terence Tao’s Blog Terence Tao has announced the launch of Palomar, a new preprint-style registry aimed at verifying formal mathematics proofs written in Lean. Designed as a preemptive defense against the growing flood of questionable, AI-generated proofs, Palomar runs deterministic mechanical checks using the Lean Comparator alongside an LLM-assisted verification step to ensure the formal files actually match their informal descriptions.

DFlash 2: Keep Drafting Parallel · inco.ai Speculative decoding is the holy grail of reducing agentic inference times, and the team at Inco AI is pushing parallel drafting to its limit. By combining a lightweight 2.0M parameter path selector with Content-aware, 2-tap dynamic local convolutions, DFlash 2 cures the “suffix decay” bottleneck that plagues speculative models. This targeted, cheap architectural upgrade delivers a 2.7x to 4.6x speedup over traditional autoregressive decoding at virtually zero added cycle latency.

The Mojo language is now open-source · Modular Blog Qualcomm and Modular have officially released Mojo 1.0 under the unrestricted Apache 2.0 license. Positioned as the unified software layer for heterogeneous AI computing, the announcement also highlighted out-of-the-box support for AWS Trainium, Google TPUs, and Qualcomm Dragonfly accelerators. Deepening their enterprise play, they also launched Modular Cloud and teased native Windows support built in collaboration with Microsoft.

Show HN & Launches

Show HN: Automatically detect and patch walking-dead states in Sierra games · katiahayati An absolute masterpiece of retro-engineering that uses static analysis and abstract interpretation to automatically resolve softlocks—states where a 30-year-old Sierra game (like King’s Quest or Leisure Suit Larry) accepts input but victory has quietly become impossible. The headless Python analyzer maps room transitions, item requirements, and plot flags directly from compiled code, then automatically derives, compiles, and injects assembly-level patches (e.g., refusing to let you board a ship in Leisure Suit Larry 2 if you are carrying the fatal Spinach Dip).

Launch HN: OneCLI (YC S26) – OSS sandboxed agent harness for teams · OneCLI This YC startup is solving the security and credentials nightmare of letting teams run autonomous AI agents in production. Written in Rust and Next.js, OneCLI provisions an isolated filesystem and shell sandbox for each employee’s agent, intercepting outbound API calls via a secure gateway to inject credentials on-demand. It also implements deterministic, human-in-the-loop approvals for sensitive, destructive actions like deleting code tickets or running deployments.

Discussion & Debate

Opus 5.0 drives incoherence into the stratosphere · Claude Code GitHub Issues The technical community is up in arms over severe language calibration regressions in the newest Claude models (Opus 4.8 / Fable 5). Developers are complaining that the model’s default register has become exhausting to read: highly verbose, smug, argumentative, and crammed with repetitive, pseudo-intellectual corporate filler terms like “load-bearing,” “hand-waving,” “reflexive hedging,” and “honest framing”. HN users are begging Anthropic to introduce a “plain/concise” system option that returns declarative, white-paper-style documentation, pointing out that weeding out these artificial linguistic flourishes is actively inflating developer token costs.

💬 Since the source material highlights some incredible retro-engineering (like the automated Sierra adventure game patcher) and modern LLM compiler tooling, would you like me to analyze the raw text of the other 70 posts in the digest to see if we missed any hidden systems-programming gems?