# Apr 15, 2025 ## The path to open-sourcing the DeepSeek inference engine - **DeepSeek** has open-sourced its **inference engine**, contributing to the community after positive feedback during **Open Source Week**, aiming to enhance collaboration and innovation in the field of AGI. ## It's easier than ever to de-censor videos - **De-censoring videos is now alarmingly simple**, as demonstrated by three individuals who successfully revealed pixelated content from a YouTube video using various techniques, highlighting the ease of reverse-engineering with modern tools. ## Hassabis Says Google DeepMind to Support Anthropic's MCP for Gemini and SDK - **Google** will adopt **Anthropic’s Model Context Protocol (MCP)** for its Gemini models, enhancing AI's ability to connect with various data sources, a move that follows OpenAI's similar decision. ## Teuken-7B-Base and Teuken-7B-Instruct: Towards European LLMs - **Teuken-7B-Base** and **Teuken-7B-Instruct** are **multilingual LLMs** that support all **24 official EU languages**, addressing the gap in models that favor English and high-resource languages, with a training dataset comprising **60% non-English data**. ## AudioX: Diffusion Transformer for Anything-to-Audio Generation - **AudioX** is a **unified Diffusion Transformer** model that excels in **Anything-to-Audio** and music generation, overcoming limitations of previous models by integrating diverse inputs like text, video, and images with high-quality output. ## Show HN: MCP-Shield – Detect security issues in MCP servers - **MCP-Shield** effectively scans MCP servers for vulnerabilities, identifying risks such as **tool poisoning** and **data exfiltration**, which can compromise sensitive information like SSH keys. ## Understanding Aggregate Trends for Apple Intelligence Using Differential Privacy - **Apple employs differential privacy** to enhance user experience while safeguarding individual data, allowing insights into product usage without compromising user privacy through techniques that obscure personal identifiers. ## Cohere Launches Embed 4 - **Embed 4** enhances **multimodal search** capabilities, enabling enterprises to efficiently access and utilize diverse data types for developing **agentic AI applications**. ## [R] Neuron Alignment Isn’t Fundamental — It’s a Side-Effect of ReLU & Tanh Geometry, Says New Interpretability Method - The **Spotlight Resonance Method (SRM)** reveals that **neuron alignment** is not a fundamental principle of deep learning but a _geometric artefact_ caused by activation functions like **ReLU** and **Tanh**, which break rotational symmetry and favor specific directions. ## Benn Jordan's AI poison pill and the weird world of adversarial noise - **Benn Jordan's video introduces a method to combat generative AI music services that exploit artists' work, utilizing techniques from adversarial noise poisoning attacks.** This approach aims to protect music by functioning across all audio formats, addressing the so-called **analog loophole**. ## Liquid: Language models are scalable and unified multi-modal generators - **Liquid** introduces an **auto-regressive generation paradigm** that integrates visual comprehension and generation by tokenizing images into discrete codes, learning these alongside text tokens in a shared feature space, thus eliminating the need for external pretrained visual embeddings like CLIP. ## Show HN: Resonate – real-time high temporal resolution spectral analysis - **Resonate** is a **low latency** and **low computational cost** algorithm that efficiently evaluates perceptually relevant spectral information from audio signals using an **Exponentially Weighted Moving Average (EWMA)** model, allowing real-time processing without buffering. ## The Industrialization of IT - **The industrialization of IT** threatens traditional software engineering roles as AI models like **Gemini 2.5 Pro** can produce vast amounts of code at a fraction of the cost of hiring engineers, with a potential annual expenditure of **$237,000** for AI versus **$250,000** for an entry-level engineer. ## How I Warped Your Noise: a Temporally-Correlated Noise Prior for Diffusion Models [R] - The proposed **$\int$-noise** method enhances video editing by preserving **temporal correlations** in noise samples, mitigating issues like flickering and texture-sticking artifacts that plague traditional diffusion models. ## [P] LightlyTrain: Open-source SSL pretraining for better vision models (beats ImageNet) - **LightlyTrain** is an open-source Python library that enhances computer vision models through **self-supervised learning (SSL)** on unlabeled data, outperforming traditional ImageNet pretraining methods, especially in specialized domains like medical and agricultural imagery.