NerdNews

Nvidia's stock plummeted 16.9% after DeepSeek, a Chinese AI startup, released its R1 model, which achieves impressive results with fewer resources and less compute power, potentially disrupting the AI hardware market. This development comes amidst a changing regulatory landscape, with the US recently reversing an executive order restricting the export of advanced AI chips to countries like China.

Alibaba's Qwen team has released a new family of AI models, Qwen2.5-VL, that can perform various tasks such as text and image analysis, controlling PCs and phones, and interacting with software. The models have been benchmarked to beat other models like OpenAI's GPT-4o and Google's Gemini 2.0 Flash in certain evaluations.

Meta AI can now use Facebook and Instagram data to personalize its responses. The chatbot can 'remember' details from conversations and provide tailored recommendations based on user preferences and information shared across Meta's apps.

DeepSeek, a Chinese AI company, has developed an AI model that outperforms some of OpenAI's leading models, raising concerns about US tech supremacy. The model, called R1, was trained at a cost of $6 million, significantly less than the tens of billions of dollars invested by OpenAI and its patron Microsoft. DeepSeek's success has sparked a price war in the AI industry, with the company charging 95% less than OpenAI for access to its most powerful models.

Chinese startup DeepSeek has released an open large language model (LLM) that rivals those from Meta and OpenAI, but at a fraction of the cost. This has led to a significant drop in US tech stocks, with Nvidia down 14%, Microsoft down 7%, and Meta down 5%. The move has raised questions about the viability of spending billions of dollars on AI training and the sustainability of the current path of AI development.

Andrew Pavlo's annual database retrospective highlights significant license changes, growth of DuckDB, and new releases in 2024, including Microsoft Garnet and Valkey, with a focus on open-source databases and their impact on the industry.

MIT researchers created an optical AI chip that can process photons directly, skipping the digitization step and achieving a latency of 410 picoseconds. The chip can perform linear and non-linear operations using photons, making it potentially faster and more energy-efficient than traditional computers. The team implemented a deep neural network on the chip, which recognized spoken vowels with 92% accuracy.

SonicWall has disclosed a zero-day vulnerability in its SMA1000 remote access appliance, which is being exploited by hackers to breach customer networks. The bug allows attackers to plant malware on affected devices without needing a login. A security hotfix has been released to patch the vulnerability.

Microsoft has launched a document database platform constructed on a relational PostgreSQL back end. The company is also suggesting users can implement an open-source FerretDB solution on the platform to get started. The new platform is fully open source, with no commercial licensing fees or usage restrictions, and is designed to meet the community's NoSQL database needs.

AWS offers various generative AI tools like Amazon Bedrock, Code Review Assistant, and Agentic Code Generation to streamline development processes, improve code quality, and increase productivity. These tools can automate code reviews, support case investigations, and test generation, allowing engineers to focus on problem-solving and delivering value to customers.

DeepSeek, a Chinese generative AI platform, is collecting and sending US user data to China, raising concerns about data privacy and security. The app's privacy policy states that it stores user information in secure servers located in China, and it may use this data to improve its services and comply with legal obligations. Experts warn that this could lead to data misuse and surveillance, and that users should be cautious when using the app.

DeepSeek, a viral AI company, has released a new set of multimodal AI models called Janus-Pro, which can outperform OpenAI's DALL-E 3. The models are available for download and can analyze and create new images.

DeepSeek's open-source R1 LLM family boasts impressive benchmark scores, but its erratic responses raise questions about its training and censorship. The model has been found to misidentify itself as OpenAI's GPT-4 and has issues with loss of context and hallucinations. Despite this, experts believe that the release of R1 will improve AI models worldwide.

Nvidia has announced the deprecation of CUDA support for its Maxwell, Pascal, and Volta architectures, which will impact datacenter operators and scientific institutions still using these older models. Although the decision may cause compatibility issues, graphics drivers for these architectures will still be available and supported for the time being.

The NVIDIA GeForce RTX 5090 FE is a powerful graphics card with improved efficiency and AI features, including neural rendering and DLSS 4, making it suitable for 3D artists, game developers, and video editors with high demands.

Former Intel CEO Pat Gelsinger has chosen to use DeepSeek's open-source AI model, R1, over OpenAI for his startup Gloo. Gelsinger believes DeepSeek's model is a game-changer, offering high-performance at a lower cost. He plans to rebuild Gloo's AI service, Kallm, using DeepSeek's open-source model, citing the benefits of open ecosystems and the potential for widespread adoption of AI.

OpenAI-backed 1X has acquired Kind Humanoid, a robotics startup that focuses on creating humanoids. The acquisition marks a key consolidation in the humanoid robotics industry. Kind Humanoid's expertise complements 1X's mission to create safe and intelligent humanoids. The deal's details have not been disclosed, but it's expected to help 1X expand its Bay Area operations.

A ransomware attack on Change Healthcare in February 2024 resulted in the largest data breach of US medical records, affecting approximately 190 million people. The attack, carried out by the ALPHV/BlackCat gang, involved the theft of sensitive medical data, including health insurance information and medical records. The breach was caused by a single set password on a user account not protected with multi-factor authentication.

SaaS breaches increased by 300% from September 2023, with 85% starting with compromised identities. Traditional security measures are failing to prevent these attacks, which are often used for financial gain, espionage, and disruption. The healthcare sector was the most affected, with 14% of breaches. To mitigate these attacks, organizations should gain a comprehensive view of SaaS applications, implement least privilege access controls, and establish ongoing monitoring.

DeepSeek's Janus Pro 1B and 7B models are designed to handle image generation and vision processing tasks, challenging OpenAI's DALL-E 3. The models improve upon the original Janus model by decoupling visual encoding and maintaining a single transformer architecture. Janus Pro 7B outperforms DALL-E 3 in some benchmarks, but has limitations such as limited input resolution and image detail.