Gemma 4 QAT models: Optimizing compression for mobile and laptop efficiency

Google has released Gemma 4 checkpoints optimized via Quantization-Aware Training (QAT) to enhance inference efficiency on edge devices and consumer GPUs. Unlike standard PTQ, QAT minimizes performance degradation by integrating quantization into the training process, enabling the Gemma 4 E2B model to run with a sub-1GB memory footprint. The release includes Q4_0 and mobile-specialized formats featuring static activations, channel-wise quantization, and targeted 2-bit quantization, with immediate support across ecosystems including llama.cpp, vLLM, and Ollama.

Read more

Meta enables ADB on deprecated Portal devices [video]

Meta CTO Andrew Bosworth announced that developer tools recently released for Quest are now compatible with Portal devices, enabling cross-platform hardware experimentation. Bosworth demonstrated a "vibe coded" home hub project, highlighting the shift toward AI-assisted rapid prototyping. The update coincides with broader AI integrations across Meta’s hardware ecosystem, including Ray-Ban Meta glasses and advanced computer graphics workflows.

Read more

Fine-tuning an LLM to write docs like it's 1995

The author utilized QLoRA to fine-tune Llama 3.1 8B and Qwen 2.5 7B on a 37-million-word corpus of 1990s Microsoft technical manuals to evaluate style transfer capabilities. The experiment demonstrated that fine-tuning is more effective than RAG for capturing specific structural and linguistic registers, with Qwen models maintaining the target persona better than Llama. Results also highlighted the interplay between LoRA rank and epochs, noting that lower-rank adapters often showed stronger commitment to the training style by reducing the model's ability to deviate from the corpus.

Read more

S&P 500 rejects SpaceX, also blocking entry for OpenAI and Anthropic

S&P Dow Jones Indices refused to waive eligibility requirements for SpaceX’s IPO, effectively blocking accelerated S&P 500 entry for OpenAI and Anthropic. The decision upholds standard rules regarding profitability, a 12-month seasoning period, and a 10% minimum public float, preventing billions in passive investment from flowing into these currently unprofitable AI firms. While the Nasdaq and FTSE Russell have eased entry paths, the S&P 500 maintains its financial viability screens despite the massive market caps and AI infrastructure spending of these entities.

Read more

The Pentagon is running an AI propaganda mill targeting Latin America

The Pentagon is utilizing an AI-driven content mill called La Tilde to conduct influence operations across Latin America. Operated by SOCSOUTH, the platform leverages LLMs to generate bilingual articles that blend generic news with pro-U.S. military propaganda. While current outputs exhibit low-fidelity AI artifacts and detectable machine-written text, the strategy highlights a shift toward using generative AI to rapidly scale and localize state-sponsored messaging.

Read more

AI Agents Enable Adaptive Computer Worms

Researchers have demonstrated a self-propagating AI worm that leverages open-weight LLMs on compromised hosts to generate target-specific attack strategies in real-time. By utilizing stolen compute across Linux, Windows, and IoT environments, the worm bypasses centralized safety controls and creates a zero-marginal-cost threat model. This shift from fixed exploit code to autonomous, reasoning-based malware marks the emergence of generative adversaries capable of adaptive synthesis and propagation.

ArXiv

Dense Contexts Are Hard: Lexical Density Limits LLM Context Windows

This research identifies lexical density—the rate of distinct information introduction—as a key, overlooked factor systematically reducing LLM effective context windows, beyond input length and needle position. Benchmarking open-weight LLMs (9B-685B) with "find-the-needle" tasks, performance sharply collapsed from near-perfect to below 60% retrieval in higher-density contexts of identical length. Reducing density restored performance, indicating that effective context capacity is a function of lexical density, with implications for LLM systems handling information-rich inputs.

ArXiv

Rethinking the Value of Generated Tests for LLM Software Engineering Agents

Research on SWE-bench Verified indicates that agent-written tests do not significantly improve LLM performance in resolving repository-level issues. Analysis shows that agents primarily use tests for observational feedback via print statements rather than formal assertions, with success rates remaining stable regardless of test-writing frequency. Prompt-intervention studies confirm that current testing practices increase interaction costs and alter workflows without meaningfully impacting final task outcomes.

ArXiv

Unlocking Non-Uniform KV Cache for Efficient Multi-Turn LLM Serving

Tangram is an LLM serving system designed to optimize non-uniform KV cache compression by addressing memory fragmentation and scheduling overhead. It utilizes deterministic budget allocation for static memory footprints, head group paging for efficient memory reclamation, and AOT load balancing to maximize GPU utilization. These optimizations enable up to 2.6x throughput improvements while fully preserving model accuracy.

ArXiv

Tracing a powerful GNSS interference source over Europe

Researchers developed a detection framework and identification techniques using received-power and TDOA measurements to analyze wide-area GNSS interference events affecting Europe and North America. By processing data from terrestrial reference stations, the study identifies the source of these transient events as a constellation of Russian early warning satellites operating in Molniya orbits.

ArXiv

Open Code Review – An AI-powered code review CLI tool

Open Code Review is an open-source CLI tool from Alibaba that automates code reviews using a hybrid architecture of deterministic engineering and LLM agents. It mitigates common agent issues like position drift and incomplete coverage through deterministic file bundling, rule matching, and reflection modules. The tool supports Git diffs, custom review rules, and integrates with CI/CD pipelines and coding agents like Claude Code.

Repo

Lowfat – pluggable CLI filter that saved 91.8% of my LLM tokens

Lowfat is a lightweight CLI tool designed to minimize LLM token costs by filtering unnecessary command-line output before it is processed by an agent. It features a composable, local-first architecture with native support for Claude Code, OpenCode, and standard shell integrations. Users can optimize context window efficiency through configurable compression levels, custom plugins via a dedicated DSL, and detailed token-saving analytics.

Repo

Jo – AI-native language to catch prompt injection at compile-time

Jo is a statically typed language compiling to Ruby and Python that uses compiler-enforced capabilities to secure AI-generated code. By tracking resource access at the type level, it allows developers to confine AI agents within strict boundaries, preventing unauthorized network or filesystem access through a "two-world" architecture. This approach provides compile-time proofs that confined code only utilizes explicitly granted capabilities.

Repo

I created a RAW to HDRI stacker in (mostly) Common Lisp

rawtohdri is a high-performance utility for stacking bracketed RAW images into OpenEXR HDRIs, recently rewritten in SBCL Common Lisp. It features a multi-threaded architecture for parallel demosaicing and EXR encoding, utilizing AVX2 SIMD instructions to accelerate floating-point stacking. The engine optimizes memory overhead by processing 16-bit integer buffers on-the-fly into a single target float buffer, achieving near-IO-bound speeds for high-resolution data pipelines.

Repo

A Simplistic UI for Rich Hickey's Design in Practice

A single-page UI renders Rich Hickey-style design decision matrices from structured text, specifically DIP-XML or DIP-YAML. This approach overcomes the limitations of spreadsheets for design artifacts, making them easier for both humans and LLMs to write, edit, generate, review, and version, while still providing a visual matrix for comparison.

Repo