Hacker News — 2026-07-05#

Top Story#

My AI-built PHP engine in Rust passes 17% of PHP-src tests, renders WordPress What makes this project fascinating isn’t just an LLM writing a parser—it’s the author’s workflow. A developer who doesn’t know Rust used the official 22,000-file PHP test suite as an incorruptible oracle to grade the AI’s code generation, preventing the model from hallucinating success or “grading its own homework”. It now successfully renders a WordPress front page from a SQLite database, proving that leveraging hostile test suites as a build system actually scales to real-world codebases, even if the resulting engine is 55x slower than native PHP.

Front Page Highlights#

Better Models: Worse Tools · Source Armin Ronacher dives into a bizarre regression where modern models like Claude Opus 4.8 and Sonnet 5 fail at custom JSON tool schemas that older models handled perfectly. He posits that heavy RLHF post-training against Anthropic’s internal Claude Code text-editor harness has created an overwhelmingly strong prior, causing the model to hallucinate non-existent JSON keys when confronted with slightly novel or nested tool definitions.

Zero-copy in Go: sendfile, splice, and the cost of io.Copy · Source A brilliant technical deep dive into how a single, seemingly harmless line of code—wrapping an os.File in a custom io.Reader just to count bytes—can silently disable the Linux sendfile(2) fast path. This forces the Go runtime to fall back to bouncing 32KB chunks through userspace, multiplying server CPU overhead by 3.4x and generating hundreds of thousands of unnecessary read/write syscalls.

The Great Blogging Collapse: What Happened to 100 Successful Blogs? · Source A grim post-mortem on the SEO blog industry following Google’s Helpful Content Update and the rollout of AI Overviews. By tracking 100 previously highly-profitable blogs over four years, the author found the median site lost 85% of its search traffic. The only survivors were highly experiential, irreplaceable niches (like DIY crafts, parenting, and original recipe testing) where the “content is the doing,” proving that highly-summarizable affiliate content is a dead business model.

Trust your compiler: Modern C++ · Source An excellent reminder to shed 90s-era C++ optimization myths—like manual loop unrolling, fearing exceptions over error codes, or writing your own fast inverse square roots. The author benchmarks old tricks against modern idioms, demonstrating that writing clear, standard C++ (using <bit> or std::expected) allows LLVM and GCC to emit heavily optimized, target-specific assembly that routinely beats hand-rolled “clever” code.

GPT-5.5 Codex reasoning-token clustering may be leading to degraded performance · Source A developer noticed a highly suspicious telemetry pattern in the OpenAI API: GPT-5.5 responses disproportionately terminate at exactly 516 reasoning tokens, with secondary fixed-boundary spikes at 1034 and 1552 tokens. This model-specific clustering strongly implies an undocumented reasoning-budget cap or truncation threshold, which correlates with degraded performance on complex coding tasks.

Connections in Math: the two kinds of random · Source A beautifully written exploration of the difference between Shannon entropy (statistical randomness) and Kolmogorov complexity (algorithmic randomness). It uses the digits of Pi to explain how a dataset can be completely incompressible from a statistical standpoint, yet remain trivially compressible by algorithmic rules since it is generated by a short program.

Web-based cryptography is always snake oil · Source A classic, hard-line security rant pointing out that browser-based “end-to-end encryption” is inherently flawed. The author argues that because the server operator delivers the client-side JavaScript executing the crypto, they can trivially backdoor the app on the next page load, making the entire threat model incoherent and merely “cryptography theatre” used as a legal loophole.

Show HN & Launches#

Espresso: Train and run transformers directly on Apple’s Neural Engine A wild project that reverse-engineers private Apple Neural Engine (ANE) APIs to bypass CoreML entirely. Written in Swift 6.2, it achieves a massive 4.76x speedup over standard CoreML pipelines by using fused multi-layer kernels and zero-copy I/O directly against the Apple silicon.

Steam Controller Auto-Charge Peak hacker ingenuity: an open-source web app that uses OpenCV optical flow and an overhead webcam to track a Steam Controller on a desk. It then uses WebHID to fire asymmetric haptic pulses through the controller’s internal linear resonant actuators, physically vibrating the controller until it inches its way autonomously onto a magnetic charging puck.

DNSGlobe A slick Rust-based TUI that allows you to watch global DNS propagation in real-time from your terminal. It polls 34 public DNS resolvers worldwide in parallel to visualize propagation status directly on an ASCII world map.

Discussion & Debate#

Mark Zuckerberg tells staff that AI agents haven’t progressed enough This leak sparked heavy debate on HN about the reality of replacing corporate workforces with LLMs. Meta laid off roughly 10% of its workforce earlier this year, leaning into an “Agent Transformation” unit that engineers described as a “soul-crushing gulag”. Zuckerberg admitted in an all-hands that the AI agents haven’t accelerated fast enough to justify the cuts, highlighting the friction between tech-exec AI hype and actual shipping capabilities.

I Wasn’t Allowed Prompting ChatGPT During My Chalk Talk: This Is Discrimination (2025) A biting satire piece written from the perspective of an academic job candidate who brought a laptop to a whiteboard interview because they literally can no longer reason or recall information without an LLM prompt box. It struck a nerve in the comments, provoking a wider debate about whether modern “AI-augmented” engineers are losing foundational problem-solving skills, or if traditional whiteboard interviews are just stubbornly outdated.