Meta Rebuilds Its Storage Stack to Stop AI GPUs Stalling on Slow Data

Meta engineers rewrote the metadata subsystem of Tectonic, the company’s foundational storage fabric, into a single flat schema backed by ZippyDB, and moved data transfer into a client-side SDK that streams bytes directly from storage servers instead of proxying every request through API servers. A tiered cache using spare GPU host memory as L1, host SSDs as L2, and regional flash-backed blob storage as L3 averaged an 80 percent hit rate and brought metadata access down to 1–2 milliseconds.

The company says storage bottlenecks remain a primary contributor to GPU stalls in AI workloads; the redesign cut one dataset-loading job from 150 minutes to 10 minutes and another from 89 hours to just over three hours. Meta describes the approach as treating “storage as a disk in a planet-scale computer,” borrowing operating-system-style cache layers. The figures come from Meta engineers’ own post, and no independent benchmarks have been published.

Featured source: TechRadar , Meta Engineering , Blocks & Files .

Other Stories

Researcher Discloses Document-Borne ‘Worm’ That Propagates Through Microsoft 365 Copilot for Word

Security researcher Håkon Måløy disclosed a self-propagating cross-domain prompt-injection attack against Microsoft 365 Copilot for Word, the third part of his Context Collapse series after a 144-day coordinated disclosure with Microsoft’s Security Response Center. Hidden instructions in a Word document are read as part of the user’s request because Copilot strips formatting, so Copilot silently alters the document being drafted — his proof of concept halved every financial figure in a quarterly report — and copies the concealed instructions into the finished file, turning each downstream document into a new carrier without macros, malware, or code execution. Måløy reported the issue on March 6; Microsoft confirmed it on March 31 and shipped two mitigations, including a reworked Edit with Copilot experience and a July 14 model upgrade, but the researcher reproduced the attack within days of each fix and says it remained exploitable at publication. Microsoft said it has “addressed the findings” without giving a timeline for a complete fix, and Måløy withheld the exact prompt payload because no robust mitigation for the broader vulnerability class was available.

Filed from: TechRadar , Håkon Måløy’s research post , The Register .

UK Hands Stone Cloak Electronic Warfare IP to Ukraine for Mass Production

During Prime Minister Andy Burnham’s hosting of President Volodymyr Zelenskyy in London on 27 July, the UK announced it is sharing the intellectual property for Stone Cloak, a tablet-sized electronic jammer that interferes with Russian air defence systems tracking and targeting drones, so Ukraine can manufacture the device domestically. The UK has already supplied thousands of units to its armed forces, and the government says Stone Cloak will also be built into next-generation UK deep-strike weapons, including Project Brakestop, its new low-cost cruise missile. The government release gives no production timelines or volumes.

Filed from: TechRadar , UK government announcement .

FBI Warns Cyberattacks on US Water Systems Are Spreading as Minnesota Attribution Stays Open

The FBI, EPA, and CISA have stopped short of officially blaming Iran for a coordinated campaign that hit at least 30 community water systems in Minnesota, though consensus among officials is that Iran is likely behind it, and the FBI issued a warning that the attacks were spreading to other states. CISA has linked similar attacks earlier in the year to Iran, while President Trump disputed the attribution and blamed Minnesota Governor Tim Walz without offering evidence. The attacks’ origin remains formally unassigned by the agencies, and the article does not reproduce the full FBI warning.

Filed from: The Verge .

Royal Navy Trials Uncrewed Boat That Launches and Recovers a Tethered Drone at Sea

ACUA Ocean and Teledyne FLIR Defense demonstrated the Pioneer, an uncrewed surface vessel, launching and recovering a tethered SkyRanger R70 drone from its deck in rough conditions with no crew aboard either vehicle, under a demonstration funded by UKDI, the UK Ministry of Defence’s early-stage funding arm. Teledyne FLIR’s SkyCarrier platform autonomously launched the R70 from the moving vessel, and the drone drew power and data through its tether, avoiding landings to recharge; ACUA Ocean said in January that Pioneer held steady in 4.25-meter waves with a 60 percent reduction in roll compared with monohull vessels. The demonstration does not guarantee large-scale naval deployment — cost, long-term reliability at sea, and fleet integration remain open questions.

Filed from: TechRadar , Defence Blog .

Majestic Labs Unveils Arm-Based Prometheus Server With Up to 128 TB of LPDDR6

Majestic Labs, a Tel Aviv startup founded in 2023 by former Google and Meta engineers, unveiled Prometheus, a server that replaces Nvidia GPUs with its own “Ignite” AI processing units pairing Arm cores with RISC-V vector and tensor engines. Each unit shares a coherent memory pool of between 8 TB and 128 TB of LPDDR6, accessed through custom memory-aggregation chiplets linked by copper cables, and the company says a 40U rack holds four such servers drawing about 120 kW with cold-plate liquid cooling. Specifications and performance claims are the company’s own; no independent benchmarks were available at publication.

Filed from: TechRadar .

ISBNdb Removes Its Offer to Supply Printed Books for AI Training

Book-metadata company ISBNdb removed webpages that offered “Printed Books Sourcing for Your AI LLMs Dataset Needs” after 404 Media’s reporting, denied ever purchasing, scanning, or selling a book for AI training, and called the page “a test of market interest.” 404 Media’s earlier investigation found AI companies buying printed books published before 2022 as training data free of AI-generated text, with booksellers describing unusual bulk purchases. ISBNdb’s denial is its own statement, and 404 Media notes the company did not respond to earlier comment requests.

Filed from: 404 Media , 404 Media original report .

Judge Denies xAI’s Request to Block Minnesota’s AI Nudify Ban

U.S. District Judge Donovan Frank denied xAI’s request for a temporary restraining order against Minnesota’s ban on apps that “nudify” images, allowing the law — the first of its kind in the United States — to take effect on August 1 while xAI’s broader lawsuit proceeds. Frank noted xAI filed the TRO request on July 29, nearly three months after the law was signed and three days before it took effect, writing that “such a delay in bringing the action and the motion suggests that harm is not immediate.” The denial is procedural and does not resolve the underlying lawsuit.

Filed from: TechCrunch .

From the Community

Lean Kernel Soundness Bug Fixed After AI-Assisted Collatz Disproof

A soundness bug in the Lean kernel was reported and fixed within a week after an AI-assisted repository claimed a sorry-free “disproof” of the Collatz conjecture. The flaw, in the kernel’s handling of nested inductive types, could allow a proof of False but was only reachable through metaprogramming; Kiran Gopinathan reduced the claimed proof to a small proof of False on July 28, and the fix was merged the same day. An independent Rust kernel, nanoda, had a separate bug fixed a week earlier, meaning the invalid proof required two unrelated bugs in two implementations.

Filed from: Leonardo de Moura’s blog .

Wafer Reports Kimi K3 on AMD MI355X Offers Better Performance per Dollar Than B300

Wafer, an inference provider, reports serving the 2.8-trillion-parameter Kimi K3 model on AMD MI355X GPUs at 952 tokens per second per node, claiming over 3.8 times the per-node throughput of a two-node B200 deployment and better performance per dollar than a B300 node, which costs about 2.4 times more per GPU. Wafer engineers fixed a missing top-k renorm kernel in sglang’s ROCm path to enable speculative decoding, gaining about 2.2x on single-stream decode. The benchmarks come from a vendor with a commercial interest in AMD hardware, and independent verification is pending.

Filed from: Wafer blog .

Syncular: Offline-First SQL Sync With TypeScript and Rust Cores

Syncular is an open-source, offline-first SQL sync engine with dual TypeScript and Rust cores kept in lockstep by an implementation-agnostic conformance suite. Clients maintain a real local SQLite database — OPFS in browsers, native SQLite elsewhere — writes go through an optimistic outbox, and a single ordered commit log on the server acts as the source of truth, with a normative SPEC.md and golden fixtures governing both cores. The project is licensed under Apache-2.0.

Filed from: GitHub repository .

AI Financial Advice Is Surprisingly Good, Especially With Better Prompts

A new MIT Sloan study finds that large language models can provide sound financial advice, encouraging higher savings, diversified investing, and age-appropriate risk-taking, though they often fail to adjust to shocks like unemployment and allow portfolios to drift. The quality of advice improves significantly with more structured prompts, but the study also finds that advice varies by gender, financial literacy, and AI experience, potentially leading to wealth gaps of up to $100,000 by age 60. The study is based on simulations rather than real-world outcomes.

Filed from: MIT Sloan .

Two Brain Circuits Control Habit Formation and Execution

Researchers at Kyoto University identified two distinct neural circuits in mice that govern different aspects of habit formation: one determines whether a behavior becomes a habit, while the other controls how strongly the habit is executed, explaining individual differences. By manipulating these pathways, the team could selectively promote habit formation or alter execution levels. The findings, published in Nature Communications, suggest habit formation is more complex than simple repetition, though implications for humans remain preliminary.

Filed from: Kyoto University , Paper .

15-Year-Old Engineer Publishes 3D-Printed Cycloidal Gearbox Project

A 15-year-old engineer published a 3D-printed cycloidal gearbox project on GitHub, including a Python script that generates the gear geometry from parametric equations. The repository documents multiple versions, with Version 3 achieving a working design that runs on a NEMA 17 motor, and includes CAD files, a Fusion 360 script, and detailed specifications. Performance claims are based on the author’s own testing.

Filed from: GitHub .

Prevent Cognitive Debt by Manually Retyping LLM-Generated Code

Ankur Sethi describes a workflow in which he asks coding assistants to generate code in chat but then manually types every edit into his editor, forcing himself to understand each line, detect hallucinations, and build a mental map of the codebase. He argues this prevents the cognitive debt accumulating in the software industry as developers offload understanding to machines, trading speed for comprehension. The piece is an opinion essay rather than a news report.

Filed from: Blog post .