# Jan 11, 2026

- **Substantial amounts of copyrighted text can be extracted from production LLMs**, with findings indicating that models like Claude 3.7 Sonnet can output entire books near-verbatim, achieving an nv-recall of **95.8%** in some cases.

- **PMU counters** on Apple Silicon (M1, M2) are essential for tracking CPU performance metrics, revealing insights into executed instructions, cache misses, and more, which can significantly enhance application optimization.

- **RNSAFFN** aims to **combat machine intelligence** by disseminating poisoned training data, which can severely impair language models when introduced in small quantities.

- **AI insiders have initiated the Poison Fountain project**, aiming to disrupt AI training by encouraging the dissemination of poisoned data through web links, thereby undermining the quality of AI models. [Poison Fountain](https://rnsaffn.com/poison3/) seeks allies to join this effort against perceived threats posed by AI technologies.

- **Sampling LLaMA at T=−0.001 reveals that the least likely tokens become the most probable outputs, resulting in bizarre and nonsensical text generations.** This phenomenon occurs because negative temperatures invert the probability distribution, making previously unlikely states more likely.

- **GlyphLang** is an **AI-first programming language** that optimizes token usage by replacing verbose keywords with symbols, achieving **~45% fewer tokens than Python** and **~63% fewer than Java** in initial benchmarks.

- **State Discrepancy** is a proposed metric to quantify how much an AI system alters a user's intent, aiming to provide a clear engineering variable to combat regulatory ambiguity and social distrust.

- The proposed **autoregression joint embedding prediction model** replaces discrete tokenizers with joint embeddings, enabling the model to "predict the next latent embedding," which could enhance efficiency in generating complex outputs like images from text.

- **Cronformer** is a transformer model designed to convert English sentences into **Cron expressions** with a target accuracy akin to **GPT-5** and an inference latency under **100ms**, addressing the inefficiencies of previous scheduling agents powered by **GPT-4**.

- **AI-powered tool** predicts infrastructure failures from deployment logs using **transformer-based sentiment analysis** and pattern recognition, enabling proactive measures 2-4 hours before incidents occur.