The Archive

Theoretical analysis of adversarial training in Vision Transformers shows benign overfitting under signal-to-noise conditions.

Adaptive MSD-Splitting improves C4.5 and Random Forests decision tree induction for skewed continuous attributes.

User reports Claude autonomously detected and investigated cryptominer in Docker container via code review.

LLMs contain shared logical subspace aligning natural and symbolic reasoning views; CCA identifies subspace for multi-step reasoning.

Theoretical analysis connecting random decision trees to overlap gap properties in symmetric binary perceptrons; advances statistical computational gap characterization.

YouTube is expanding its AI deepfake monitoring feature to Hollywood - meaning some celebrity AI videos could soon disappear. The platform's likeness detection feature searches YouTube for AI deepfake content and flags it for public figures enrolled in the program. Public figures can use it to keep track of AI content on YouTube of themselves or request removal (takedowns are evaluated against YouTube's privacy policy, and not every request will be approved). YouTube began testing the feature with content creators last fall; in March, the company expanded the program to politicians and journa...

**TL;DR**: Bought Claude Max 5x Gift on April 9, 2026. Worked fine for 12 days, then silently downgraded to Free plan with zero notification. Billing still shows “Paid $0.00” and valid until May 9. This is a widespread known bug. The real problem is support: \- [https://support.anthropic.com](https://support.anthropic.com) has no **ticket form**, no way to submit a real request for regular users. \- The only way to contact support is through the in-app “Get Help” chat on [claude.ai](http://claude.ai) — which is just useless Fin AI Agent. \- Once your account is downgraded to Free, h...

Reddit user implements diffusion language model from scratch on M2; produces toy results on Shakespeare dataset.

As AI agents increasingly work alongside humans across organizations, companies could be inadvertently opening a new attack surface. Insecure agents can be manipulated to access sensitive systems and proprietary data, increasing enterprise risk. In some modern enterprises, non-human identities (NHI) are outpacing human identities, and that trend will explode with agentic AI. Solid governance and…

LLM-based Evidence-Minus-Intuition scoring applied to 15M parliamentary speeches reveals epistemic orientation patterns linked to deliberative democracy quality.

Comparative analysis of two determinantal point process estimators for Monte Carlo integration with variance reduction guarantees.

SmoothCruiser algorithm for entropy-regularized MDPs and games achieves O(1/ε⁴) sample complexity via Bellman operator smoothness.

User reports Gemma 4 e4b extracted from Android app outperforms published versions; unclear if genuine discovery or misconfiguration.

A-MAR framework uses agent-based reasoning plans and retrieval to ground artwork interpretation in structured evidence and cultural context.

The photo deletion comes after an FTC settlement with Clarifai. The company had asked OkCupid — whose executives had invested in Clarifai — to share data in 2014, according to court documents.

New task for document-grounded QA: generating related insights from document collections to iteratively refine open-ended answers.

PREF-XAI framework learns user-preference-based personalized rule explanations for black-box ML models with principled preference modeling.

Anthropic secures 5 gigawatts of Amazon’s custom silicon as Claude demand soars.

Function vectors extracted from multilingual LLMs show language-agnostic properties; translation vectors transfer across target languages.

Hybrid position-force control policies using reinforcement learning improve robotic manipulation accuracy under force constraints.

Budgeted online influence maximization algorithm using independent cascade model and semi-bandit feedback with cost-based optimization.

Multi-cycle spatio-temporal adaptation framework for joint human-robot planning that learns individual preferences across task, motion, and spatial dimensions.

HardNet++ enforces nonlinear constraint satisfaction in neural network outputs for safety-critical control and decision-making applications.

Chat2Workflow benchmark evaluates LLM capability to generate executable visual workflows from natural language multi-turn interactions.

Reproducibility study standardizing evaluation of 11 counterfactual explanation methods for recommender systems across metrics and datasets.

Self-supervised disentangled representation learning framework for structural damage identification robust to environmental and operational variability.

GAAP execution environment guarantees privacy for AI agents by enforcing data access controls against prompt injection and untrusted providers.

GRIL trains LLMs via reinforcement learning to recognize missing premises and pause rather than fabricate information in incomplete reasoning contexts.

Dual analysis of synthetic human mobility trajectory generators examining privacy vulnerabilities alongside utility gains from generative models.

Deep learning framework for environmental sound deepfake detection evaluating spectrograms, architectures, and pre-trained models on benchmark datasets.