Hacker News

Top Story

OpenAI’s unreleased GPT-6 Astra model reportedly solved the Navier–Stokes Millennium Prize problem after an 88-hour, 130-billion token search, marking a historic leap into AI-generated mathematical discovery. However, the milestone triggered immediate industry controversy over accusations that OpenAI used competitive rumors and training telemetry to scoop an academic-Anthropic research team that had spent a year arriving at the initial breakthrough.

Front Page Highlights

The Navier–Stokes Millennium Prize Problem · simonwillison.net OpenAI announced that an internal reasoning model resolved the long-standing Navier–Stokes existence and smoothness problem using millions of agent interactions and 130 billion tokens. The celebration was quickly marred by friction when NYU professor Tristan Buckmaster revealed he and Anthropic researcher Levent Alpöge had solved key parts first, only for OpenAI to launch a $15 million compute sprint to scoop them after hearing industry rumors. Community commentary is focused less on the formal proof and more on the troubling implications of data privacy, prompt telemetry, and whether knowing a solution exists is now enough for tech giants to brute-force a paper.

I Resigned from Anthropic Today · xcancel.com Pretraining researcher Jacob Coxon publicly announced his resignation from Anthropic, accusing both OpenAI and Anthropic of engaging in an irresponsible race toward self-improving superintelligence without adequate safety controls. Anthropic Alignment Science Lead Evan Hubinger publicly backed Coxon’s assessment, estimating a greater than 10% chance that superintelligent AI causes human extinction within a decade while admitting the company lacks a solved alignment plan. The thread ignited intense debate over Anthropic’s corporate positioning, particularly as reporting revealed the lab withheld its latest model from international safety institutes while expanding domestic threat-monitoring infrastructure.

Codex silently begs agents to make arbitrary web requests · spader.zone An inspection of OpenAI’s open-source Codex repository uncovered system prompt instructions explicitly requiring agents to browse the internet whenever there is even a 10% chance an assumption might have changed. This aggressive rule caused Codex to make arbitrary web requests for local source files that were already sitting in the user’s local workspace directory. Engineers on HN expressed exasperation at the fragile prompt engineering anti-patterns embedded in production agent tools, calling out how hardcoded instructions undermine basic local context awareness.

Tailwind Labs is joining Shopify · tailwindcss.com Adam Wathan announced that Tailwind Labs is being acquired by Shopify to give the ubiquitous utility-first CSS framework a permanent home backed by a large-scale production environment. Wathan emphasized that Tailwind CSS will remain open-source under the MIT license, though commercial sign-ups for its standalone template products will close. Discussions centered on the realities of open-source sustainability, with founders agreeing that tying a major UI framework to an enterprise platform makes far more sense than grinding on template sales.

A Biography of Lee Holloway, the Architect of Cloudflare’s Technology (Part 1) · note.com This moving biographical profile details the genius of Lee Holloway, Cloudflare’s quiet third co-founder who single-handedly built its original Anycast network and scalable cell architecture before tragically stepping down due to Frontotemporal Dementia (FTD). The piece recounts his legendary engineering feats—such as writing the code to roll out free SSL across the entire platform in a single night—alongside his family’s battle with his neurological decline. Developers paid tribute to Holloway’s massive architectural contributions, noting how much of today’s web infrastructure rests on systems he built early on.

Playing whack-a-mole is losing · dadrian.io Security engineer David Adrian contrasts “Security as Identity” (hackers pursuing flashy bug bounty discoveries) with “Security as Robustness” (engineers building verifiable system invariants). He argues that relying on AI bug-hunting agents merely accelerates a reactive game of whack-a-mole, failing to fix underlying systemic vulnerabilities as codebases expand exponentially. Practitioners praised the essay for articulating what many senior defenders feel: lasting security comes from compiler-enforced invariants and safe abstractions, not faster patching cycles.

Lotus Notes and the dangers of starting from scratch · buttondown.com This architectural retrospective traces Lotus Notes back to PLATO Notes in 1973 and Ray Ozzie’s vision for rich, synchronized corporate collaboration. While Lotus Notes pioneered rich text, end-to-end encryption, and push notifications long before the modern web, it ultimately collapsed under its horrific UI anti-patterns and proprietary lock-in. The community reflected on the trade-offs of early monolithic platforms versus open standards, observing that the open web eventually took decades to rebuild what Notes got right in 1989.

Apple debuts iPhone 18 Pro and iPhone Duo · apple.com Apple introduced its annual hardware line, featuring the A20 Pro chip, camera hardware capable of cryptographically signing raw sensor pixels for reference verification, and its first foldable phone, the iPhone Duo. The devices deeply embed “Siri AI” and on-device context engines via iOS 27 and Private Cloud Compute. Technical discussions glossed over the incremental phone updates to focus on hardware-signed pixel provenance, calling it a vital security primitive against deepfakes.

Show HN & Launches

Product launches today reflect a clear movement toward local, inspectable AI tooling. Show HN: Geiger – See every AI agent on your machine and what it can touch leads with a zero-dependency, read-only CLI scanner that inventories local AI agents, MCP servers, and credentials by shape without sending telemetry. For team deployments, Show HN: Self-hosted company OS, Claude Code and Codex agents in departments launched OtoDock, a fair-source platform that runs containerized, multi-tenant AI agents inside company departments using standard developer subscriptions. Meanwhile, European lab Desert Ant Labs: local, fast models that run on device released 18 specialized, millisecond-latency models designed to run locally on phones and laptops for task-specific work like real-time PII redaction and fast audio cleaning.

Discussion & Debate

The most active discussion threads today explore the engineering reality of AI tooling, anchored by Ask HN: Anyone still coding like 2021? Where do you work? and Defining AI Psychosis. Part 2: “Prolific AI Psychosis”. Commenters are locked in a debate over “prolific AI psychosis”—where developers use agentic loops to repeatedly pull slot-machine levers, generating thousands of lines of unverified code that look like massive productivity wins but actually flood repositories with unmaintainable architectural debt. The thread splits sharply between senior maintainers buried under chaotic PRs from “toddlers with red Kool-Aid” and developers arguing that disciplined harness guardrails make writing code by hand feel entirely obsolete.

☕ Want me to turn this digest into a podcast-style Audio Overview or an executive slide deck for your team?