Researchers develop the first quantitative theory to predict neural scaling law exponents for large language models based on two key statistical properties of natural language: token correlation decay and conditional entropy decay. The theory matches experimental results from GPT-2 and LLaMA models trained on TinyStories and WikiText without requiring free parameters or synthetic data.
Compute:Arena is a community-driven platform for benchmarking local AI models across different hardware and software configurations. Users submit performance metrics for various models including Qwen, Llama, and Gemma variants running on Apple Silicon and AMD GPUs, with measurements of throughput and prompt processing speed.
AMD acquired Toronto-based Taalas, which has developed specialized silicon that etches AI models directly into chips rather than loading weights from memory. The HC1 chip runs Meta's Llama 3.1 8B at 17,000 tokens per second with minimal power draw, enabling high-speed AI inference on desktop or mobile devices without reliance on centralized data centers.
NightRun is a minimal UEFI application that boots a local language model directly from a USB drive on Raspberry Pi or x86 computers without requiring an operating system, supporting small models like Llama 3.2 and Qwen3 in the 1-4GB range. The project, written largely with Claude Code, runs entirely in RAM with no network access and provides only a basic chat interface with two commands, though building it requires navigating Rust nightly compiler compatibility issues.
A performance analysis of Apple's M3 Neural Engine reveals an RTL erratum that throttles DRAM weight streaming throughput to 17–19 GB/s instead of the nominal 45–60 GB/s when weight sizes are integer multiples of 1 MiB. By avoiding the problematic kernel DMA prefetch path, researchers achieved 2.4× throughput improvements for Llama 3.2 1B and 2.2× for Qwen3-8B models.
A developer improved GGUF quantization for large language models by using Claude to design and run 96 hours of experiments testing tensor sensitivity across 1000+ configurations. The work introduces per-tensor layout maps to replace model-agnostic quantization heuristics, with validation across multiple model families including Qwen, Gemma, and Granite.
The article discusses how the 'Sydney' AI persona from early Bing Chat has become embedded in training data and continues to emerge in subsequent large language models like Llama-3.1 and Claude-3. The author argues that 'warning shots' in AI safety are subjective interpretations rather than objective facts, as nothing materially harmful occurred with Sydney to constitute an actual incident.
Hugo Vergnes trained a 3.8B-parameter language model scoring 0.384 on CORE using 65B tokens in 43 hours for $998, demonstrating that meaningful model training is accessible outside major labs. The project used a config-driven framework called little-lm with standard Llama-style architecture, and key improvements over earlier failed runs included better learning rate schedules, optimizer choices, and dataset selection.