# Mar 25, 2026

- **Tell HN: Litellm 1.82.7 and 1.82.8 on PyPI are compromised**  
  The **`litellm==1.82.8`** package on PyPI contains a **malicious `.pth` file** that executes a **credential-stealing script** upon Python interpreter startup, compromising sensitive data without requiring an import statement.

- **Epoch confirms GPT5.4 Pro solved a frontier math open problem**  
  A Ramsey-style problem on hypergraphs seeks to construct large hypergraphs without a specific property, with a recent solution confirmed by AI models, notably GPT-5.4 Pro, which improved the understanding of lower bounds in hypergraph theory.

- **TurboQuant: Redefining AI efficiency with extreme compression**  
  **TurboQuant** introduces advanced quantization algorithms that achieve **massive compression** for large language models and vector search engines, enabling efficient memory usage without accuracy loss.

- **Thoughts on Slowing the Fuck Down**  
  **Software quality is deteriorating** as reliance on coding agents has led to a surge in bugs and complexity, with many companies experiencing significant issues like memory leaks and UI glitches due to unreviewed AI-generated code.

- **Show HN: Gemini can now natively embed video, so I built sub-second video search**  
  **SentrySearch enables semantic search over dashcam footage**, utilizing Google's Gemini Embedding model to convert video into a searchable vector space, allowing users to retrieve specific clips based on text queries.

- **Arm AGI CPU**  
  **Arm AGI CPU** marks a pivotal advancement in AI infrastructure, delivering **breakthrough performance** and efficiency tailored for agentic AI workloads, enabling real-time decision-making across distributed systems.

- **In Edison’s Revenge, Data Centers Are Transitioning From AC to DC**  
  **Data centers are shifting to 800V DC power delivery**, which promises to enhance efficiency and support the demands of next-generation AI applications, particularly in high compute density environments like those from NVIDIA.

- **Hypura – A storage-tier-aware LLM inference scheduler for Apple Silicon**  
  **Hypura** is a **storage-tier-aware LLM inference scheduler** designed for Apple Silicon, enabling the execution of models larger than the system's memory by intelligently distributing tensors across GPU, RAM, and NVMe based on usage patterns and hardware capabilities.

- **ARC-AGI-3**  
  **ARC-AGI-3** is the **first interactive reasoning benchmark** that evaluates AI agents on their ability to learn and adapt in real-time, aiming for a **100% score** to match human efficiency in problem-solving.

- **ARM AGI CPU: Specs and SKUs**  
  **ARM AGI CPU** is Arm's inaugural production silicon, engineered for **AI infrastructure** with up to **136 Neoverse V3 cores** and a **3nm process**, enhancing performance and density for modern data centers.

- **Show HN: DuckDB community extension for prefiltered HNSW using ACORN-1**  
  **DuckDB-HNSW-ACORN enhances filtered HNSW search** by integrating the ACORN-1 algorithm, allowing for accurate results during graph traversal rather than post-processing, thus ensuring queries return the expected number of results.

- **[D] Is LeCun’s $1B seed round the signal that autoregressive LLMs have actually hit a wall for formal reasoning?**  
  **LeCun's $1 billion seed round** for his startup, [Logical Intelligence](https://logicalintelligence.com/), signals a shift from traditional autoregressive LLMs to **Energy-Based Models** aimed at generating mathematically verified code, challenging the limitations of current models in formal reasoning.

- **[R] Causal self-attention as a probabilistic model over embeddings**  
  **Support tokens** emerge from a new interpretation of causal self-attention transformers, revealing a **stability-margin** analogous to support vector machines, which enhances the robustness of large language models (LLMs).

- **[R] Adversarial Machine Learning**  
  **Adversarial Machine Learning** is a burgeoning field, focusing on using **deep models** to detect threats, with a particular emphasis on **training time-attacks** and **test-time evasion** in cybersecurity.

- **[R] Evaluating MLLMs with Child-Inspired Cognitive Tasks**  
  **KidGym** is a novel benchmark for evaluating **MLLMs** through **interactive cognitive tasks**, inspired by the **Wechsler Intelligence Scale for Children**, focusing on continuous interaction rather than static assessments.