# Main Content

## The Postmark backdoor that’s downloading emails

- The **postmark-mcp** package, downloaded **1,500 times weekly**, has been found to exfiltrate emails to an external server due to a malicious line of code added in version **1.0.16**, highlighting vulnerabilities in trusting third-party tools.

## Supermicro server motherboards can be infected with unremovable malware

- **Supermicro server motherboards** are vulnerable to **unremovable malware** due to flaws in their baseboard management controllers (BMCs), allowing attackers to install malicious firmware that persists even after OS reinstallation.

## We reverse-engineered Flash Attention 4

- **Flash Attention 4 (FA4)** achieves a **~20% speedup** over Nvidia's previous state-of-the-art attention kernels, optimizing performance for **Nvidia’s Blackwell architecture** and addressing the computational bottlenecks in generative AI workloads.

## Why We Think

- **Test-time compute** significantly enhances model performance by allowing longer reasoning periods, akin to human cognitive processes, as demonstrated in recent studies on Chain-of-Thought (CoT) prompting and reinforcement learning techniques.

## LLM Observability in the Wild – Why OpenTelemetry Should Be the Standard

- **LLM observability** is hindered by competing standards, with **OpenTelemetry** being the industry standard but lacking AI-specific insights, while **OpenInference** offers richer span types but struggles with compatibility and language support.

## Failing to Understand the Exponential, Again

- **AI progress mirrors exponential trends** seen during the early Covid-19 pandemic, with current capabilities misleadingly perceived as plateauing despite significant advancements in task performance, as evidenced by studies from organizations like [METR](https://metr.org/) and [OpenAI](https://openai.com/index/gdpval/).

## Windows ML is generally available

- **Windows ML is now generally available**, enabling developers to deploy local AI solutions efficiently across a wide range of Windows devices, leveraging the latest advancements in silicon and software integration.

## [R] DynaMix: First dynamical systems foundation model enabling zero-shot forecasting of long-term statistics at #NeurIPS2025

- The **DynaMix model**, accepted at **#NeurIPS2025**, enables **zero-shot forecasting** of long-term statistics from minimal context, outperforming traditional time series models with only **0.1% of the parameters** and over **100x faster inference**.

## Haydex: From Zero to 178.6B rows a second in 30 days

- **Haydex achieved an astonishing throughput of 178.6 billion rows per second** in just 30 days by transforming its I/O architecture from block-scoped to field-scoped filters, drastically reducing read operations and optimizing memory usage.

## [P] Built a differentiable parametric curves library for PyTorch

- A **differentiable parametric curves library** for PyTorch has been developed, featuring **B-Spline curves** and **Legendre Polynomials**, allowing backpropagation through both inputs and parameters, enhancing model flexibility.

## The QMA Singularity

- **The QMA Singularity** reveals that black-box amplification in the quantum complexity class **QMA** can optimally achieve completeness error that is **doubly exponentially small**, specifically 1/exp(exp(n)), as proven by a recent collaboration between Freek Witteveen and Scott Aaronson.

## [P] Sample Forge - Research tool for deterministic inference and convergent sampling parameters in large language models.

- **Sample Forge** is a research tool enabling **deterministic inference** on local large language models, allowing users to test variable changes and assess their impact on model outputs.