# Nov 16, 2025

## Heretic: Automatic censorship removal for language models

**Heretic** automates the removal of censorship from transformer-based language models using **directional ablation** and a **TPE-based optimizer**, achieving results comparable to manual methods without requiring deep technical knowledge.

## A new Google model is nearly perfect on automated handwriting recognition

**Google's new AI model, currently in testing, achieves near-perfect automated handwriting recognition and demonstrates unexpected levels of abstract reasoning, suggesting a significant leap in AI capabilities.**

## Tracking users with favicons, even in incognito mode

**Supercookie** leverages favicons to create a **persistent tracking ID** for users, which remains intact even in incognito mode and is resistant to common privacy measures like cache flushing or VPNs.

## Archimedes – A Python toolkit for hardware engineering

**Archimedes is a Python framework designed to streamline control systems development by enabling seamless C code generation from NumPy algorithms, enhancing productivity and deployment efficiency.**

## [R] Generative Flows on Weight Space for Covariate Shift Detection (AAAI 2026 Workshop)

**Generative flow models** in weight space enable effective detection of **covariate shifts** by adapting weight distributions in response to data changes, outperforming traditional methods in practical applications.

## Beyond Hyperparameters: We're Now Quantifying (and Steering) the Internal Physics of AI Training. [R]

The **Vector Space Mapping (VSM)** protocol reveals that quantifying the **specialization of attention heads** in Transformer models can significantly enhance training efficiency, achieving over **160% faster structural reorganization** with optimized hyperparameters.