NerdNewsMarch 12, 2025 |
News & Trends
Meta is reportedly testing its first in-house AI training chip
Meta is testing its first in-house AI training chip, aiming to reduce dependence on NVIDIA and lower infrastructure costs. The chip is part of the Meta Training and Inference Accelerator series, designed for generative AI, recommendation systems, and advanced research.
OpenAI pushes AI agent capabilities with new developer API
OpenAI has unveiled a new Responses API to help developers create AI agents that can perform tasks independently. The API provides access to models that can scan company files, navigate websites, and browse the web to answer questions. The company also released an open-source Agents SDK to integrate models with internal systems and monitor agent activities.
Meta Is Ready to Rock Nvidia’s Boat With Its In-House AI Chip
Meta has started testing its in-house AI chip, aiming to reduce its reliance on Nvidia's pricey AI chips. The company plans to use the chip for training models by 2026, which could help improve content recommendations and ad targeting, and potentially save the company billions of dollars in costs.
Google has given Anthropic more funding than previously known
Google has invested over $3 billion in Anthropic, a San Francisco AI startup, with a 14% stake and a $750 million convertible debt deal. This raises questions about Anthropic's independence and the impact of Big Tech funding on AI startups.
OpenAI Launches New Tools to Help Businesses Build AI Agents
OpenAI has launched a new Responses API to help developers build custom AI agents. The API includes tools for web search, file search, and automating computer tasks. OpenAI also released an open-source Agents SDK for integrating models with internal systems and monitoring AI agent activities. |
Options & Tutorials
EU AI Act: Latest draft Code for AI model makers
The European Union has released a third draft of the Code of Practice for general-purpose AI model makers, which provides guidance on complying with the EU AI Act. The updated code includes streamlined structure and refined commitments, with a focus on transparency, copyright, and risk mitigation. The code is expected to be finalized by May, and companies can provide feedback until March 30.
.NET 10 Preview 1: Updates in Runtime, SDK, Frameworks and More
.NET 10 Preview 1 introduces enhancements to the .NET Runtime, SDK, libraries, C#, ASP.NET Core, Blazor, .NET MAUI, and more. Key updates include OpenAPI 3.1 support, simplified OpenAPI version configuration, and improved integration testing. Additionally, C# 14 adds features like field-backed properties and implicit conversions for Span<T> and ReadOnlySpan<T>.
Microsoft March 2025 Patch Tuesday fixes 7 zero-days, 57 flaws
Microsoft's March 2025 Patch Tuesday updates address 57 vulnerabilities, including 7 zero-day flaws and 6 critical remote code execution vulnerabilities. The updates fix issues in various Windows components, including NTFS, Windows Kernel, and Microsoft Office. Users are advised to apply the updates as soon as possible to protect against potential attacks.
MINJA Attack Poisons AI Model Memory
Researchers have discovered a new attack called MINJA that can poison AI model memory, allowing malicious users to manipulate the model's behavior and potentially leading to harmful consequences. The attack works by sending a series of prompts to the model that include extra details intended to manipulate its memory. The technique has been tested on several AI agents powered by OpenAI's GPT-4 and has achieved a high success rate.
Cerebras to Deploy AI Accelerators in North America and France
Cerebras is deploying over 1,000 AI accelerators across North America and France, bringing online new datacenters and expanding its inference services. The company's CS-3 accelerators offer high performance and memory bandwidth, making them suitable for large language models and other AI applications. |
Launches & Tools
Dexterity Lands $95M Funding for AI-Powered Robots
Dexterity, a startup specializing in industrial robots with human-like finesse, has raised $95 million at a post-money valuation of $1.65 billion. The company's robots use specialized AI models to perform repetitive and sometimes dangerous tasks in warehouses and factories.
Hugging Face Expands LeRobot Platform with Training Data for Self-Driving Machines
Hugging Face has partnered with Yaak to expand its LeRobot platform with a new training set called Learning to Drive (L2D), which contains over a petabyte of data from sensors installed on cars in German driving schools. The dataset is designed to support end-to-end learning for self-driving models and will be used for real-world testing this summer.
Waymo Expands Robotaxi Service Across Silicon Valley
Waymo is expanding its robotaxi service to Mountain View, Los Altos, Palo Alto, and parts of Sunnyvale, increasing its coverage in the San Francisco Bay Area to 82 square miles. The company plans to test in 10 new US cities this year and has partnered with Uber to offer rides in Austin and Atlanta.
iRobot launches eight new Roombas with lidar room mapping
iRobot has launched eight new Roomba models with lidar navigation and mapping, a technology used by most competitors. The new lineup includes the 105, 205, 405, and 505 series, with prices starting at $299. The models feature improved suction power, dual spinning mop pads, and a new design. The company has also launched a new iRobot app with real-time cleaning view and support for Matter smart home standard.
Google Enhances Data Privacy with Confidential Federated Analytics
Google introduces Confidential Federated Analytics (CFA), a technique that increases transparency in data processing while maintaining privacy. CFA leverages confidential computing to ensure that only predefined and inspectable computations are performed on user data. This approach addresses the limitation of traditional federated analytics by using Trusted Execution Environments (TEEs) to restrict computations and prevent unauthorized access to raw data. |
Quick Links
Why extracting data from PDFs is still a nightmare for data experts
The article discusses the difficulties of extracting data from PDFs, a problem that affects many industries. Traditional OCR methods are limited, but new AI-based approaches, such as large language models, are being developed to improve data extraction. However, these models also introduce new problems, such as hallucinations and instruction following. The article highlights the current state of PDF data extraction and the potential solutions being explored.
Flower Labs Launches AI Service
Flower Labs, a Y Combinator-backed startup, has launched a preview of its distributed cloud platform for serving AI models. The platform, called Flower Intelligence, can drive on-device AI mobile, PC, and web apps that automatically hand off to a private cloud when needed. It uses open models, including those from Meta's Llama family, and employs end-to-end encryption to protect sensitive user data.
Sola emerges from stealth with $30M to build the 'Stripe for security'
Sola, an Israeli startup, has emerged from stealth with $30M in seed funding to build a low/no-code platform for cybersecurity, aiming to simplify security services for organizations and make them more affordable. The platform allows users to design their own cybersecurity apps tailored to their specific needs, and it can be used to query data within existing security apps or replace certain functionality. Sola's goal is to 'democratize' how security can be approached and handled, similar to how Stripe simplified payments and Canva simplified design.
The State of Secrets Sprawl 2025
GitGuardian's report reveals 23.8 million secrets leaked on public GitHub repositories in 2024, a 25% year-over-year increase. 70% of secrets leaked in 2022 remain active today, expanding the attack surface for threat actors. The report highlights the need for proactive remediation and better security hygiene across all industries.
Microsoft Adds Voice Chat to Copilot
Microsoft has added a new feature to its Copilot AI assistant, allowing users to interact with it using voice commands. The feature is triggered by holding down the Alt + Spacebar keys for two seconds, and is available to Windows Insiders. This new functionality provides an alternative to typing out requests, and can be used to ask questions, get information, and more. |
Share NerdNewsShare your affiliate link to get commission!
Thanks for reading,
If you dont want to receive future editions of NerdNews, |