HackerNews Digest

OpenAI announced its first custom inference processor, Jalapeño, developed with Broadcom and built on the company’s AI models. The chip targets the inference workload—running pre‑trained models for user requests—and is claimed to deliver markedly higher performance‑per‑watt than existing state‑of‑the‑art solutions, though it remains in testing. Early results suggest reduced operating costs for real‑time coding models, while pre‑training is expected to continue relying on Nvidia GPUs. The partnership, publicly confirmed in October, reflects OpenAI’s effort to lessen dependence on Nvidia’s accelerators, mirroring similar moves by Google and Amazon. OpenAI’s leadership emphasized a stack‑wide optimization strategy, encompassing chip architecture, kernels, memory, networking, scheduling, and deployment, to make models faster, more reliable, and cheaper for users. The announcement positions the company to control both model development and the underlying hardware infrastructure.

None

The GitHub issue serves as an umbrella for proposing and documenting syntax extensions for LuaJIT 3.0. The first comment will host evolving documentation, and contributors are invited to discuss design, semantics, and improvements while keeping feedback constructive and focused on functionality rather than cosmetic symbol choices. Accepted extensions must enhance developer quality‑of‑life, be proven in other languages or Lua dialects, avoid syntactic ambiguities, preserve backward compatibility, and remain tool‑friendly (e.g., for formatters and language servers). The maintainer explicitly rejects adding the syntactic complexity found in languages such as Perl, Ruby, C++, or Rust. Prior related issues (#63 and #1379) have been closed in favor of this consolidated discussion. An accompanying image references the user “@MikePall”.

Cloudflare has expanded its OAuth offering to all customers, allowing them to create and manage OAuth applications for delegated access to the Cloudflare API. The move follows an upgrade of the underlying Hydra OAuth engine: first to the latest 1.x release (with concurrent index creation and custom SELECT‑only queries to avoid exclusive table locks) and then to a 2.x version using a blue‑green deployment. To preserve writes during migration, token expiry was extended and a revocation‑replay queue was added, ensuring revocations are not lost. Post‑upgrade, stricter refresh‑token invalidation caused errors, which were mitigated by adding refresh‑token coalescing in a routing worker and by leveraging Hydra’s configurable refresh‑token grace period. Additional steps included data cleanup for new schema constraints, coordinated cutovers of related services, and monitoring for session corruption that led to 403 responses. The upgrade resulted in improved performance, stability, and a clearer consent/revocation experience, enabling broader SaaS integrations and internal developer platforms on the Cloudflare platform.

John Gruber criticizes intrusive pop‑ups and “accept cookies” overlays, arguing that a webpage should simply display its content without obstructive prompts—comparable to an email that only links to a web page. The author reflects on this meta commentary, noting that blogging often feels like stating the obvious, which can seem unimportant or redundant. Despite the perceived lack of novelty, the author argues that pointing out widely ignored annoyances (e.g., intrusive UI elements) can produce valuable posts, especially when supported by examples (“receipts”). The key takeaway is that effective blogging may hinge on a willingness to articulate observations that seem evident to the writer but are rarely voiced, or to amplify similar statements from others with clear endorsement.

LinkedRecords is a NoSQL Backend‑as‑a‑Service that can be accessed directly from single‑page applications without custom backend code. It provides built‑in, per‑record authorization: the creator specifies read permissions, eliminating separate backend rule sets. The API is deliberately minimal—few endpoints that can be composed to support varied use cases (RBAC, etc.)—and operates independently of the consuming SPA, delivering real‑time updates. A continuous load test simulates document management: each `createDocument` builds a blueprint of eight attributes (key‑value, long‑text, collaborators, comments, references, configuration). The test runs three simulated users (continuous creator, test user creating one per ten documents up to 300, occasional creator) and measures response times for `createDocument`, `fetchDocuments`, and `fetchDocument` as total document count grows, with results plotted automatically. Deployment is configured via optional environment variables. In confidential‑client mode a session token is stored in an HttpOnly cookie; when the frontend and backend are on different domains, the token must be stored in the browser (public‑client mode). The SDK is initialized with OIDC parameters. Blob attributes can be stored in S3 (recommended) or fallback to PostgreSQL.

The author recounts early attempts with *War and Peace* and later reading *Crime and Punishment*, noting that despite long length and unfamiliar Russian names, the prose of Russian classics is clear, direct, and often humorous. Translations—particularly older ones like Constance Garnet’s—are described as readable, with modern translators (Pevear & Volokhonsky) recognized as higher quality but not essential for comprehension. The piece argues that perceived difficulty of great literature stems more from academic commentary than from the texts themselves, citing G.K. Chesterton’s observation that “Great Books” are inherently accessible. By contrast, works such as *Ulysses*, *Gödel, Escher, Bach*, and *The Glass Bead Game* are highlighted as genuinely challenging. The author concludes that Dostoevsky’s lasting appeal lies in the simplicity and vividness of his storytelling, not in any intrinsic complexity, urging a reread of the final chapter of *The Brothers Karamazov* to confirm this view.

None

The paper argues that conventional programming languages restrict domain‑specific geometric expression to linear text, limiting developers who could benefit from visual syntactic constructs. It proposes a hybrid language approach where visual and textual syntax are treated equally and can be composed without disrupting standard workflows. The authors introduce **Hybrid ClojureScript**, the first language to support visual interactive syntax as a composable extension: programmers can embed visual elements directly within source code, and an enhanced IDE renders these embeddings as miniature GUIs for interaction, while other IDEs display a textual representation. The design preserves static reasoning and aims to be adoptable across domains. Evaluation shows that developers can integrate visual syntax without major learning overhead, and the paper discusses requirements for adapting the mechanism to other languages. The work was submitted on 16 Mar 2026, revised on 27 Mar 2026, authored by Leif Andersen and four co‑authors.

None