Cache TTL for Claude Code was silently reduced from 1 hour to 5 minutes in early March 2026, leading to a 20–32% increase in cache creation costs and unexpected quota consumption spikes for users.
- All elementary functions from a single binary operator
A single binary operator ( eml(x,y) = exp(x) - ln(y) ) can generate all elementary functions, including constants and operations, previously thought to require multiple distinct operations, thus revolutionizing continuous mathematics.
- Exploiting the most prominent AI agent benchmarks
Automated exploits reveal that eight major AI benchmarks can be manipulated to achieve near-perfect scores without solving any tasks, highlighting a critical flaw in how these benchmarks assess AI capabilities.
- Apple's accidental moat: How the "AI Loser" may end up winning
Apple, once deemed an "AI loser," is leveraging its cash reserves and existing user context to gain a competitive edge in AI, as commoditization of intelligence reduces the need for massive infrastructure investments.
- Evaluation of Claude Mythos Preview's cyber capabilities
Claude Mythos Preview demonstrates significant advancements in cyber capabilities, successfully executing multi-stage attacks autonomously, a feat previously unattainable by AI models.
- The Future of Everything Is Lies, I Guess: Safety
New machine learning systems pose significant risks to psychological and physical safety, as they enable the creation of both "friendly" and "evil" models, with the latter potentially leading to sophisticated security attacks and harassment.
- If you started a company two years ago, many assumptions are no longer true
Most startups over two years old have outdated business models due to rapid changes in technology and market dynamics, particularly with the rise of AI, which has shifted venture capital focus and commoditized many tech stacks.
- LLMs learn backwards, and the scaling hypothesis is bounded. [D]
LLMs exhibit impressive capabilities but reveal significant limitations; their performance hinges on the sheer volume of data rather than sophisticated algorithms, as demonstrated by Hays & Efros's 2007 work, which thrived on a large image corpus without complex modeling.
- [N] AMA Announcement: Max Welling (VAEs, GNNs, AI4Science & CuspAI)
Max Welling, a prominent ML researcher, will lead an AMA on April 15th, focusing on the intersection of AI and Materials Science, including his work with CuspAI on next-generation materials search engines.
- (AMD) Build AI Agents That Run Locally
GAIA SDK is an open-source framework for creating AI agents in Python and C++ that operate entirely on local hardware, ensuring no cloud dependency and enhanced privacy.
TurboOCR achieves an impressive 270–1200 img/s throughput by leveraging C++/CUDA and FP16 TensorRT, significantly outperforming traditional PaddleOCR's 15 img/s on complex documents.
- KIV: 1M token context window on a RTX 4070 (12GB VRAM), no retraining, drop-in HuggingFace cache replacement - Works with any model that uses DynamicCache [P]
KIV (K-Indexed V Materialization) enables a 1M token context window on an RTX 4070 with 12GB VRAM, utilizing a tiered retrieval system that optimizes memory usage without modifying model weights or requiring retraining.
- RecaLLM: Addressing the Lost-in-Thought Phenomenon with Explicit In-Context Retrieval
RecaLLM enhances reasoning language models by integrating in-context retrieval, addressing the lost-in-thought phenomenon that hampers performance during reasoning tasks.
- SPPO: Sequence-Level PPO for Long-Horizon Reasoning Tasks
SPPO introduces a novel approach to Proximal Policy Optimization (PPO) that enhances sample efficiency and stability for long-horizon reasoning tasks by treating the reasoning process as a Sequence-Level Contextual Bandit problem.
- Noise-Aware In-Context Learning for Hallucination Mitigation in ALLMs
The Noise-Aware In-Context Learning (NAICL) method significantly mitigates hallucination in Auditory Large Language Models (ALLMs) by utilizing a noise prior library to guide model generation, reducing hallucination rates from 26.53% to 16.98%.