AI Weekly Malaysia

This is a link to a PDF by Tristan Buckmaster related to Navier-Stokes equations, but the article text is unreadable raw PDF binary data with no extractable content.

Mistral raised €3B in a Series D round at a €21B+ post-money valuation, led by Samsung with co-leads EQT's Scaleup Europe Fund and PSG Equity. The funding will expand Mistral's compute capacity for training, infrastructure, and international footprint across its current 20 countries of operation.

A blog post argues that GLM 5.3-flash, an open-weight model from Z.ai (formerly Zhipu AI), combined with DeAlignAI's 'abliterated' version that scores 0% on Harmbench-320 safety benchmarks, puts capable, unrestricted AI hacking tools in anyone's hands for ~5-15k USD in hardware. The author claims we have roughly a year to use frontier LLMs to find and fix vulnerabilities industry-wide before threat actors exploit this, noting the M5 Mac Studio with 256GB unified memory (releasing September 22) will run the model at ~30 tokens/second locally.

Singtel's 26.9% stake in Bharti Airtel is worth S$41.7 billion, accounting for roughly 56.6% of Singtel's S$73.5 billion market cap. Temasek, which owns Singtel, calls India its best-performing market over 10 years, with recent exits like Schneider Electric India delivering a ~10x return on S$860 million invested in 2020.

Huawei launched the Kirin 950 Pro, a processor it claims is entirely free of US technology, featuring its own LinxiCore CPU, GPU, and NPU designs. It uses a 'Tau Scaling' architecture to improve performance without relying on TSMC or Intel's advanced chipmaking processes, powering the $2,979 three-screen Mate XT2 folding phone. Analysts view it as an interesting development but not a major breakthrough, and Huawei hasn't detailed plans to adapt the chip for data center or AI roles.

Malaysian IC design firm SkyeChip, operational since 2020, reported Q1 FY2027 revenue of US$12.25M with silicon IP contributing 52.9% and custom ASIC 46.9%, selling primarily into the US (54.6%) and China (42.1%). Kenanga analyst Tan Woon Pin sees product and royalty revenue as a 'natural extension' of its current NRE and licensing business but has not yet factored it into earnings forecasts due to timing uncertainty. The company employs 365 engineers out of 391 staff, has filed 113 patents (36 granted), and spent RM59.7M on R&D in FY2025—roughly half of revenue.

Eric Wu, former Opendoor CEO, launched a new company out of stealth in May 2026 building AI copilots for construction workers — described as a hands-free expert coach for field laborers. The construction industry faces a shortfall of ~349,000 workers this year, worsened by aging workers, stricter U.S. immigration enforcement, and surging data center projects that now require 4,000–5,000 workers per campus versus the previous ~750.

Arm announced its next-generation semi-custom Neoverse CSS N4 compute subsystem platform (codenamed 'Falcon' in the headline, 'Ranger' in the URL), packing up to 128 cores per die on TSMC's N3P process. The article body was not accessible — only the title and site boilerplate were provided.

A March 2024 Google core update appears to suppress all pages except the main page of brand-new root domains from search results, lasting up to a year or longer. Weird Gloop reports ~90% of independent wikis launched on new domains since then have been affected, despite content quality or originality. Their workaround is launching new wikis (Overwatch, Fortnite) on subdomains of weirdgloop.org instead of new root domains.

Jellyfin 12.0 is a major release that changes the database schema and actively rewrites data on first boot, making a full manual backup the only rollback path. Upgraders must be on 10.10.7 or 10.11.x first, usernames are now case-insensitive (migration fails if duplicates exist), a full library scan is required post-upgrade, and all third-party plugins built for 10.11 will not load until authors release 12.0-compatible builds.

Matthew McPherrin factored two 512-bit RSA root CA keys from the defunct Canadian CA E-Certify, shipped in Netscape 4.51 in March 1999. He used Claude Code to extract old root certificates from browser archives on archive.org, then ran CADO-NFS on a Ryzen 9 5950X, taking 32 and 29 hours respectively to recover the private keys.

Rockstar Games faces an employment tribunal in Glasgow over the dismissal of 31 workers accused of sharing confidential information on Discord ahead of GTA VI's release. The Independent Workers' Union of Great Britain claims the firings targeted trade union activities, while Rockstar cites gross misconduct. The tribunal is ongoing.

Anthropic's Lance Martin outlines three concrete ways to cut Claude API costs without sacrificing performance: maximize prompt cache hit rate, remove prompt anti-patterns when upgrading to frontier models, and calibrate effort to the task. The article details prompt caching mechanics—cache is model-pinned, requires byte-exact prefix matches, has a TTL—and warns against volatile values in system prompts, reordering tool definitions, and changing effort mid-conversation (except on Claude Opus 5 and Fable 5.1).

OpenAI publishes a case study claiming 1Password achieved 20.9% measured productivity improvement among engineers using Codex, a 10.9% reduction in median PR cycle time, and ~90% reduction in investigation time on a complex production issue, with an estimated $0.8M annual engineering capacity value.

OpenAI is launching a journalism support initiative starting with the 2026–2027 academic year, providing 400+ ChatGPT Edu licenses to journalism students at CUNY's Newmark J-School and Northwestern's Medill School, in collaboration with the Tow-Knight Center and Knight Lab. The program covers tools, training, and partnerships across classrooms and newsrooms.

TALA (Terrastruct's AutoLayout Algorithm) is now open-source under MPL-2.0 and bundled in D2 v0.9.0, usable via --layout=tala. It's an orthogonal layout engine optimized for software architecture diagrams, blending multiple graph-drawing research techniques to optimize for symmetry, flow, and clustering. A standout feature is that node positions and sizes can be locked to specific coordinates while TALA handles edge routing — useful for agentic workflows where LLMs place nodes but still struggle with routing.

Konstantin Ryabitsev reports that git.kernel.org now spends more CPU cycles rendering commits for abusive scrapers than on all legitimate access combined, with 14 cores across 5 geo-distributed nodes dedicated solely to serving crawler traffic. Simon Willison highlights this as a growing concern for any project serving large numbers of crawlable web pages, including his own Datasette.

OpenAI Chief Scientist Jakub Pachocki argues that the strongest case for rapidly training smarter models is the need for defensive AI systems that can secure infrastructure, counter rogue agents in real time, and invent new protective measures. He simultaneously cautions that the seriousness of these stakes should not become an excuse for reckless racing forward.

An Amazon Prime Air cargo jet (Boeing 767-300 operated by 21 Air as Flight 7598) overran the runway at Miami International Airport by about 1,300 feet on Sept. 6, 2026, striking vehicles inside and outside the airport boundary. At least five people were killed and five injured; NTSB chair Jennifer Homendy described the scene as 'utter devastation' but offered no details on the cause.

Latent Space built a Frontier AEO (Answer Engine Optimization) Tracker running 6 prompt variations across 7 frontier models in 161 product categories, from coding agents to payroll software. They found clear self-bias (Claude models recommend Claude Code, Astra recommends Codex, Grok recommends Cursor), 28 categories with a universally dominant primary choice across all models, and consequential recommendation flips between model generations (Opus→Fable, Sol→Astra) that signal shifts in training data and RL priorities.

Broadcom has quietly removed public download pages for VMware's Virtual Disk Development Kit (VDDK), with support reportedly telling customers it is no longer available for general use. The VDDK is a core library used by migration tools including Azure Migrate, Nutanix Move, Red Hat Migration Toolkit, virtv2v, and nbdkit, making it significantly harder for customers to move workloads off vSphere to competing platforms.

TechCrunch published a living AI glossary covering terms from AGI to AI agents, aimed at people who encounter jargon in meetings and pitches. The notable new entry is 'opaque recurrence,' described as the reasoning technique in OpenAI's new Astra model that has AI safety researchers concerned.

An interactive 3D visualization renders every surviving building in Los Angeles as a box, appearing in the year it was built from 1880 to 2026. It uses LARIAC 2020 building outlines and LA County Assessor roll data for heights, with heights exaggerated when zoomed out.

Simon Willison used Claude Fable 5.1 in Claude Code for web to build a browser-based video compressor tool powered by the WebAssembly build of FFmpeg, so he could optimize a phone-recorded demo video before publishing it on his blog.

SOCRadar disclosed PEEP, a post-exploitation toolkit that injects a fake 'Smart Bookmarks' extension (ID: ejkndncpkdcjcikfhiamcdehdoegilbj) directly into Chrome/Edge profiles by forging Chromium's Secure Preferences integrity values, bypassing Web Store checks entirely. A native messaging host binary (nm_host.exe) escalates it from credential theft to full host-level command execution, polling a C2 server every 30 seconds over plaintext HTTP. PEEP is built on the open-source RedExt red-teaming framework and contains Chinese-language source artifacts, though attribution remains unconfirmed.