Morning Edition · Wednesday, July 15, 2026
Tech
PrismML's 1-Bit Bonsai 27B Compresses a 27-Billion-Parameter Model to 3.9 Gigabytes
The company reports the 1-bit build keeps 89.5 percent of full-precision quality and the ternary build 94.6 percent, using quantization-aware training rather than post-training compression.
Geopolitics
Anthropic Publishes a Cross-Industry Jailbreak Severity Scale After a 19-Day Export-Control Suspension
The Cyber Jailbreak Severity framework, built with Amazon, Microsoft and Google, rates attacks from CJS-0 to CJS-4 on exponential bands modeled on the CVSS software-vulnerability standard.
Tech
NVIDIA Forms a Nemotron Coalition of Open-Model Labs Across Twelve Countries
The group pools GPU allocation and data to train permissively licensed frontier models, and NVIDIA says Harvey matched leading closed models on legal tasks at roughly a tenth of the cost per run using a post-trained Nemotron.
Tech
Meta's Brain2Qwerty v2 Decodes Typed Sentences From Non-Invasive Brain Signals at 61 Percent Word Accuracy
The pipeline reads magnetoencephalography while a person types and reconstructs the text in real time, but it still requires a 306-sensor scanner in a shielded room and has only been tested on healthy participants during actual typing.
Tech
Samsung Readies a Standalone 4-Nanometer AI Accelerator for PCs, Codenamed Gaia
HP and Lenovo are validating the neural processing unit, which pairs Samsung's 4-nanometer node with processing-in-memory and is aimed at mass production as early as 2027.
Markets
NVIDIA Argues Performance Per Watt, Not Raw FLOPS, Now Governs AI Data-Center Economics
With power the fixed constraint, the company frames tokens generated per watt as the metric that sets an AI facility's revenue and margin.
Tech
MonkeyOCRv2 Pretrains a Document Vision Model on 113 Million Images Across 17 Languages
A dual objective that jointly learns image-to-text generation and pixel-level document reconstruction yields a reported 2.8-point absolute gain on the MDPBench multilingual parsing benchmark over the prior best open model.
Tech
ChatGPT Returns to WhatsApp Across Europe After the EU Ordered Meta to Restore Rival Access
OpenAI re-enabled the assistant across all 27 EU states plus Iceland, Liechtenstein and Norway on July 13, six months after Meta barred general-purpose AI bots from the WhatsApp Business API.
Tech
A New Linear-Attention Variant Adds Phase-Controlled Delta Memory to Fix Long-Context State Tracking
Semidirect Fourier Delta Attention keeps a fixed recurrent state instead of a growing key-value cache, using constructive chunk-WY kernels to preserve exact state tracking that linear attention normally loses.
Tech
OpenAI Temporarily Lifts the Five-Hour Usage Cap on GPT-5.6 Sol and Codex After a Demand Surge
The change follows the model's July 9 launch, keeps weekly caps intact, and comes with inference optimizations the company says free roughly 10 percent more usage from the same allowance.
Markets
Researchers Scale Point-in-Time Language Models to Strip Lookahead Bias From Financial Backtests
Models trained on unrestricted internet corpora embed information from the future, and the paper studies how to build language models whose knowledge is bounded to a chosen historical date.
Tech
Study Finds Prompt Formatting Alone Can Flip LLM Leaderboard Rankings
Under a token-controlled protocol, wrappers that differ only in formatting shift model scores enough to reorder benchmarks, and the authors propose a Format Sensitivity Index to measure it.