Tech News

Revolut has moved roughly 2,000 engineers onto self-hosted GLM for coding, making Claude Code and Codex exceptions rather than defaults. According to AI ops lead David Abitbol, the setup uses 56 Nvidia B200s at a blended 30 cents per million tokens—about a tenth of what Revolut previously paid Anthropic and OpenAI.

A reported issue alleges that Postman silently auto-installs AI plugins and attempts to share all API data, raising privacy and security concerns. Users described unexpected changes to their coding agents after opening Postman on macOS. The report has been filed in the issue tracker, where duplicate submissions are expected to be closed.

A new paper argues that LLMs can natively read and write neural network weights, enabling “weight-space programming.” The news piece reports that Qwen 3.8 27B running on a laptop can also interpret model weights, suggesting the capability may be accessible on local hardware.

A recent post claiming Astra can read and write weights to improve Qwen is misleading. The experiments involving reading raw weights and those that fix Qwen rely on completely different techniques, so they do not support the implied connection.

Fujifilm plans to increase its magnetic tape production fivefold to meet growing archiving demand driven by AI. Magnetic tape remains attractive for archival storage because it is low-cost and offers offline security.

A car owner jailbroke their vehicle’s navigation system just to change the clock, highlighting how locked-down infotainment software can block basic settings. They note that the workaround shouldn’t be necessary but found the process enjoyable.

A fun experiment back-ports 30 years of chess engine improvements to run on a 1996 Pentium Pro, testing whether modern techniques on period hardware could have defeated Kasparov in 1996. It explores what might have been possible if researchers had today’s engine knowledge at the time.

OpenAI has built Pando, a Git-compatible backend designed to handle the development scale made possible by its Codex coding tools. The system aims to overcome Git’s scaling limits as AI-assisted coding increases repository activity and development throughput.

Stringman is an $800 open-source room-scale gripper robot whose concept intrigued commenters, though some said they’d wait for a full-body humanoid version. The discussion quickly turned darkly humorous, with readers joking about the robot strangling or stabbing users and comparing it to Futurama and Douglas Adams-style absurdity. No technical details were available because the article itself couldn’t be fetched.

Google is giving schools more ways to restrict student access to its A.I. tools. However, schools that block A.I. entirely will lose student access to Google’s search engine.

GitBook’s CTO confirmed the company has been dealing with a security attack since yesterday, taking a large portion of services offline to protect users and progressively restoring them while monitoring the situation. The company is keeping users updated via email, Slack, and its status page, and discussion praised its outreach and decision to rotate tokens itself. Some readers also questioned GitBook’s claim that the attack involved AI agents, asking what evidence distinguishes it from human activity.

Users are discussing claims that OpenAI has silently re-enabled model training on Codex chats after users opted out, though the original article could not be fetched and no proof has been provided. Commenters debate how hard such changes are to verify, recommend self-hosted or gated enterprise inference for sensitive data, and cite a similar Mistral opt-in/opt-out incident as evidence of broader trust concerns.

Running Macs in a datacenter requires custom MacBook racks, virtualizing macOS using Apple’s Virtualization framework, and managing large VM images across the fleet. The piece details the infrastructure and operational work needed to scale Mac capacity at datacenter scale.

Quili AI is a playful human-powered alternative to AI chatbots that has sparked debate over relative water use, with commenters estimating a person answering an average 11-minute chat may consume roughly four times the water of an AI interaction. Others caution the comparison depends on electricity use and how water re-enters the cycle, and some criticize the project’s claims as overly propagandistic while pointing to Chile’s broader water crisis.

Intel’s stock slide deepened Monday, with weakness spreading to memory and storage chip stocks as a broader $3.5 trillion semiconductor rout intensified. Most of Wall Street otherwise moved past last week’s tech sell-off.

Kalshi and its co-founders face felony charges for allegedly violating Utah’s gambling laws amid the state’s total ban. The charges come as the Trump administration reportedly tries to protect the prediction-market platform from Utah’s crackdown.

Researchers have revealed writing on PHerc. 306B, a carbonized Herculaneum papyrus, for the first time in roughly 2,000 years. The breakthrough highlights how advanced imaging and computational techniques are making it possible to read fragile ancient scrolls without physically damaging them.

A new effort is working to shrink NixOS’s RAM footprint and reduce its memory requirements to as low as 512 MB. The goal is to make the OS viable on lower-resource systems by minimizing resource usage.

JIVX lets users paste a Japanese YouTube video and turns its lines into sentences they can type or say, with each line graded. Users can also practice with videos that others have already added.

A discussion of “Resisting Cognitive Surrender” warns that relying on LLM-generated code and content can erode human understanding, turn review into rubber-stamping, and degrade codebases over time. Commenters support scaffolding and review workflows but argue that unchecked adoption risks persistent dependence and lost cognitive competence, making continued investment in human capability and control essential.

WorkOS says a security incident was isolated to its internal go-to-market systems and that no customer platform data was accessed. The CEO said the attack involved ARTEX, an autonomous AI agent framework used for offensive security operations, and that the incident has been contained with independent experts investigating. Customers seeking more details can contact WorkOS’s security team.

Muvon has released Octomind, an open-source AI coding agent and runtime built as a single static Rust binary. It is model-agnostic and MCP-native, aims to save tokens, and can run in the terminal, CI, or as a daemon.

A new study estimates that 75% of NFL players who died over a 15-year span showed signs of chronic traumatic encephalopathy, or CTE. It is the highest estimate to date of the degenerative brain disease among professional football players.

Cloudflare is currently experiencing multiple active incidents, according to its real-time status page. The page provides live updates and incident history across Cloudflare services, network locations, and scheduled maintenance.

Antimony, a code-based CAD tool rooted in Lisp-machine ideas rather than drafting-table workflows, has seen maintenance taper since around 2016, though a more actively maintained fork called Stibium is now available. Discussion around the piece highlights renewed momentum in code CAD and AI-driven CAD projects, citing OpenSCAD’s long-awaited release candidate, PythonSCAD, CADquery, Build123D, and Halfspace. Commenters debated the practicality of code-based parametric CAD while noting desired features like SolveSpace-style 2D sketching and fillets/bevels.

ByteDance has developed USDB, a new storage system aimed at scaling the infrastructure needed for next-generation recommender systems. The arXiv paper outlines the design and motivation for handling the growing storage demands of advanced recommendation workloads at the company.

An Ask HN post seeking a contact with a Cleveland-area ham radio club drew helpful replies from local amateur radio operators, including a suggestion to reach out to Case Western Reserve University’s W8EDU club. Commenters also offered to connect the poster with nearby radio enthusiasts and briefly discussed how the post reached HN’s front page.

A year after their first writeup, the author revisits how they use AI and finds their workflow and tools have changed entirely. The piece documents the stark contrast between their earlier practices and their current approach.

The article uses the phrase “No brother of mine will play the flute” to frame a discussion of how two wrongs can sometimes produce a right. It argues that rigid rules and retaliatory choices can occasionally lead to surprisingly workable outcomes, while cautioning that such cases are the exception rather than the rule.

The article examines how infinite loops in GPU shaders can hang the GPU, even though shaders are meant to be fast, small, and finite. It notes that SPIR-V does not prevent describing a shader with an infinite loop, so such loops should be enforced at shader compilation time. Reader discussion welcomes the piece and stresses the need for safeguards against pathological shader code.