AI Weekly Malaysia

UK AISI is publishing evaluation results through EvalEval's Evaluation Cards platform using the Every Eval Ever (EEE) schema, covering five benchmarks (HealthBench, FrontierMath, Humanity's Last Exam, SWE-Bench Pro, Terminal-Bench 2.0) across six frontier models including Claude Opus 4/4.5/4.6 and GPT-5/5.2/5.4. The release includes verified results, configuration details, and transcript-level transparency to make evaluations reproducible.

OpenAI published a position statement outlining its priorities and principles for third-party safety assessments of its frontier AI models, covering areas like validating safety claims, testing safeguards under realistic conditions, and enabling independent assessors with deep access including chain-of-thought visibility and internal deployment data. The post proposes four priority areas for deeper assessment and emphasizes independence, scientific rigor, and security practices.

Hugging Face's transformers library now supports loading GGUF (llama.cpp quantized) models directly via from_pretrained, reusing llama.cpp's ggml kernels for performance. Initial support targets Apple Silicon and the Qwen3.5 architecture, with recommended starting quantization Q4_K_M (2.74 GB for Qwen3.5-4B vs 8.42 GB unquantized).

Jun Kim, creator and maintainer of oMLX (an MLX-based library for local AI on Apple Silicon), has joined Hugging Face to work on MLX full-time. oMLX remains Apache 2.0 and Jun continues leading it, but the project moves from a side project to a funded effort. A key focus will be streamlining the conversion path from transformers model definitions to reference MLX implementations that different inference engines can consume.

Ron Johnson, who built Apple's retail store network starting in 2000, argues that AI agents won't fundamentally change how people shop, especially for high-value personal purchases like laptops. He concedes AI may narrow choices but insists consumers will still want physical experiences before spending, even as Google pushes its Universal Commerce Protocol and OpenAI turns ChatGPT into a shopping destination.

TypeSafe AI unveiled Jev, a new model category they call 'System One' (or 'decision models') that accepts text input but returns floating-point numbers instead of text—confidence scores for yes/no questions, probability distributions across choices, or numeric scores along a defined range. Input is priced at $0.042/million tokens with output free, undercutting GPT-5 Nano ($0.05/million), and questions are evaluated in parallel so you can batch many against a single document at similar latency to one.

Meta's Muse AI personal agent app, powered by Meta's Muse Spark model family, overtook ChatGPT as the top free iOS app in the U.S. on Sep 19, accumulating 730,000 downloads in roughly five days following its Sep 8 launch. Sensor Tower data shows Muse's post-launch download numbers exceeded ChatGPT, Claude, and Grok over the same window, marking Zuckerberg's major push into the consumer AI agent market.

The article coins 'spymark' to distinguish hidden tracking signals from traditional watermarks, calling out Google's SynthID as a prime example. SynthID-Image can embed a 136-bit payload (64-bit database ID + 72 bits error correction) in a 512x512 image, imperceptible to humans, potentially linking output to user identity records including name, IP, address, and more. OpenAI and others are developing similar systems at scale.

Texas Gov. Greg Abbott ordered the state's environmental regulator to halt all data center-related permits until ERCOT completes an audit of the existing grid connection waitlist, expanding an August moratorium that already paused new grid approvals. Texas was projected to surpass Virginia as the world's largest data center market by decade's end per a JLL report, but this directive may slow that trajectory. Data center buildout has become a major issue in the 2026 midterms, with an NBC News poll showing 64% of Americans leaning against it.

Colin Breck argues that AI-generated professional writing—design docs, PR summaries, tickets, meeting notes, even personal messages—has become pervasive but unreadable, lacking context, perspective, and human voice. He distinguishes between using AI to help himself write better versus being forced to consume machine-generated summaries that skip the thinking and consensus-building that writing is supposed to drive.

Cloudflare's Python Workers are now generally available after a two-year preview, making Python a first-class language on their edge platform. Python code runs compiled to WebAssembly via Pyodide inside the V8-based workerd runtime, which means multiprocessing and threading are non-functional. Local development uses the pywrangler tool (published as workers-py on PyPI) which runs a 123MB workerd binary simulating the full stack locally.

Trump and Xi are expected to meet in Washington to shore up a fragile trade truce, with AI policy, tariffs, export controls, and rare earths among the contested issues. The article is a summit preview with no concrete outcomes yet, citing Treasury Secretary Scott Bessent on 'deliverables' and China expert Claire Reade noting 'inconsistent viewpoints' within the negotiations.

Ars Technica reports that Muse, Meta's AI assistant described as having extraordinary privileges, has a serious 0-day vulnerability. The article body was not properly extracted—only cookie consent boilerplate was captured—so specific technical details about the vulnerability, its impact, and any proof-of-concept are unavailable from the provided text.

Diogo Almeida, CEO of TypeSafe AI and InstructGPT co-author, explains why Jev—a model whose launch video drew ~40M views—targets 'System One' production use cases rather than chat-tuned LLM behavior. He argues API frontier models took a wrong turn on alignment, refusals, and reliability, and advocates RLCD over three flavors of RLHF. The podcast covers concrete Jev patterns: coding agents with an official guide, linting, compacting tool calls, voice + browser/computer control, a Doom demo, analytics replay, entity resolution, and 'Jev as judge,' plus his note on the 'Tyranny of the KV Cache.'

The article URL points to an Ars Technica story about a Ukrainian naval drone sinking a Russian kamikaze drone boat, but the captured page content contains only cookie consent boilerplate—no article text was retrieved.

The headline announces Saudi Arabia's launch of Ceer, a new EV brand with two planned vehicles, as part of its effort to build a domestic car industry. However, the article body provided is entirely cookie consent boilerplate with no substantive content.

OpenAI published a blog post proposing global safety standards for frontier AI development, focusing on alignment research and recursive self-improvement (RSI), a computing technique where AI systems iteratively improve themselves. The company called for international cooperation building on existing AI safety institutes, with standards targeting frontier models, developers, and benefit-risk management for automated AI researchers. Anthropic released its own safety proposals the prior week amid escalating warnings about AI risks.

OpenAI announced an independent math advisory group hosted at the Institute for Advanced Study in Princeton, after an internal model reportedly resolved over 100 open math problems including the Navier-Stokes Millennium Prize problem. 25 Fields Medal winners signed an open letter criticizing AI labs for rushing famous math solutions. The advisory group can assess and coordinate release of results but explicitly cannot slow or redirect OpenAI's internal math research.

Xiaomi released MiMo v2.6, details unavailable as the article page could not be retrieved. The HN post drew unusually high engagement with 1118 points and 476 comments, suggesting significant developer interest.

AMD crossed a $1 trillion market cap for the first time, with shares up ~24% over a five-day rally and an intraday high of $613.92. The company's Q2 revenue was $11.54 billion (up 50% YoY), driven largely by its Data Center unit at $6.7 billion, up 107% YoY on AI chip demand.

Transformer Explainer is an interactive web visualization that breaks down the GPT-2 (small, 124M parameter) model architecture, showing how embeddings, self-attention, MLP layers, and output probabilities work step by step. It lets users tweak temperature, top-k, and top-p sampling in real time to see how next-token prediction responds.

TechCrunch Disrupt 2026 runs October 13-15 at Moscone West in San Francisco, featuring 200 Startup Battlefield companies and speakers from OpenAI, Anthropic, and Replit. Early-bird pricing offers up to $200 off plus 50% off a second ticket, ending September 25 at 11:59 p.m. PT.

Linear's engineering team reworked their CI pipeline because AI coding agents accelerated code production faster than validation could keep up, making CI a bottleneck. Despite test suites nearly quadrupling since January, they cut PR wait time from 6+ minutes to ~5 minutes and halved runner time per test by moving off GitHub Actions to faster third-party runners (34% faster jobs), switching to tsgo native TypeScript compiler (73% faster tsc), rewriting lint rules to drop TypeScript type dependency (68% faster API lint), and adopting Oxlint.

Appfigures estimates Meta's Muse AI app reached 1.8M iOS downloads in the U.S. and Canada within its first 12 days, versus ChatGPT's 1.3M at the same point. Muse also hit 642K U.S. daily active users compared to ChatGPT's 231K, though launch strategies differed (Muse on both iOS and Android but U.S./Canada only; ChatGPT iOS-only but global).

The article link points to a Cloudflare CAPTCHA verification page; no actual article content was retrieved. The headline claims NASA's Mars Sample Return mission is dead, but the body text provides zero detail to confirm or elaborate.