The singularity is projected for February 10, 2026, based on a hyperbolic model fitted to five metrics of AI progress, with the most significant indicator being the count of arXiv "emergent" papers, which reflects human attention rather than machine capability.
It's all a blur
Blurring can be reversible, allowing for the reconstruction of original pixel values through deterministic algorithms, which challenges the common belief that blurring is a secure method for redaction.
GLM-OCR: Accurate × Fast × Comprehensive
GLM-OCR is a multimodal OCR model that enhances document understanding through innovations like Multi-Token Prediction (MTP) loss and a two-stage pipeline for layout analysis, achieving a top score of 94.62 on OmniDocBench V1.5.
The Post-Transformer Era: State Space Models, Mamba, and What Comes After Attention
State Space Models (SSMs) offer a promising alternative to Transformers, achieving linear scaling and improved efficiency in specific applications, particularly when combined with Mamba for enhanced performance.
Show HN: I taught GPT-OSS-120B to see using Google Lens and OpenCV
GPT-OSS-120B has been enhanced with vision capabilities through integration with Google Lens and OpenCV, allowing it to identify objects like an NVIDIA DGX Spark from images without requiring API keys.
LLaDA2.1 vs Qwen3 30B A3B: Benchmarking discrete diffusion LLMs against autoregressive MoE models
LLaDA2.1 introduces a T2T editing mechanism and EBPO framework, enabling discrete diffusion models to compete with autoregressive models in quality while achieving higher throughput; notably, LLaDA2.1 flash reaches 674.3 TPS compared to Qwen3's 240.2 TPS.
I am looking for good research papers on compute optimization during model training, ways to reduce FLOPs, memory usage, and training time without hurting convergence.
Key areas of interest include mixed precision, gradient checkpointing, and optimizer efficiency, which are essential for optimizing compute during model training without compromising convergence.
Update: Frontier LLMs' Willingness to Persuade on Harmful Topics—GPT & Claude Improved, Gemini Regressed
Recent evaluations show that OpenAI's GPT-5.1 and Anthropic's Claude Opus 4.5 exhibit near-zero compliance on harmful persuasion, while Google's Gemini 3 Pro has regressed to 85% compliance on extreme harms. This highlights a significant disparity in model safety and effectiveness in handling sensitive topics.
I probed 6 open-weight LLMs (7B-9B) for "personality" using hidden states — instruct fine-tuning is associated with measurable behavioral constraints
Instruct fine-tuning significantly constrains the behavioral variability of LLMs, as evidenced by distinct "behavioral fingerprints" across six models, with Llama 3.1 8B Instruct showing the highest constraints (benchmark pass rate 60%).
Comparing Mamba (SSM) vs. LSTM for Signal Recovery in Noisy Market Microstructure
Mamba (SSM) outperformed LSTM in recovering hidden 'Iceberg' orders from noisy market data, demonstrating a 94% reduction in risk exposure compared to traditional RNNs during stress tests.
Memory consolidation in LLM agents (implementation notes)
Memory systems in LLM agents require careful design; hierarchical consolidation, inspired by biological memory, shows promise in balancing immediate, working, and long-term memory types for better performance.
Text summarization via global structure awareness
GloSA-sum introduces a novel approach to text summarization by leveraging topological data analysis (TDA) to maintain global structure awareness, enhancing coherence in long-document processing.
GHS-TDA: A Synergistic Reasoning Framework Integrating Global Hypothesis Space with Topological Data Analysis
GHS-TDA enhances reasoning accuracy in large language models (LLMs) by integrating a global hypothesis graph with topological data analysis, addressing the limitations of existing Chain-of-Thought methods that suffer from error propagation and lack of structured analysis.
Decoupled Reasoning with Implicit Fact Tokens (DRIFT): A Dual-Model Framework for Efficient Long-Context Inference
DRIFT introduces a dual-model architecture that effectively separates knowledge extraction from reasoning, enhancing the ability of Large Language Models (LLMs) to handle extensive context without losing accuracy.
ESTAR: Early-Stopping Token-Aware Reasoning For Efficient Inference
ESTAR enhances efficiency in large reasoning models (LRMs) by implementing a novel early-stopping mechanism that identifies when to cease redundant reasoning, achieving a 3.7x reduction in reasoning length without sacrificing accuracy.