NerdNews

Microsoft has announced the Maia 200, an AI inference accelerator built on TSMC 3nm process, with native FP8/FP4 tensor cores and a redesigned memory system. It promises to deliver three times the FP4 performance of Amazon Trainium and FP8 performance above Google's tensor processing unit, with 30% better cost performance.

OpenAI's Codex AI coding agent works through an 'agentic loop' where user input is sent to a model, which generates a response or requests a tool call. The loop repeats until a final answer is produced. The company has shared technical details on how this process works, including prompt construction, tool implementation, and sandboxing.

Anthropic has launched interactive Claude apps, including Slack and other workplace tools, allowing users to access and utilize these tools within the chatbot interface. The feature is available to Pro, Max, Team, and Enterprise subscribers and enables users to send Slack messages, generate charts, and access cloud files. The system is built on the Model Context Protocol (MCP) and is similar to OpenAI's Apps system.

Nvidia has invested $2 billion in CoreWeave to help the debt-ridden company add 5GW of AI compute capacity by 2030. CoreWeave will integrate Nvidia's products, including the new Rubin chip architecture, across its platform. The deal is a significant show of support for CoreWeave, which has faced scrutiny over its debt obligations.

Nvidia has released new AI weather forecasting models, including Earth-2 Medium Range, Nowcasting, and Global Data Assimilation. These models claim to be more accurate and faster than traditional weather forecasting methods, with Earth-2 Medium Range beating Google's GenCast model on over 70 variables. The models use transformer architectures and can be run on GPUs, making them more accessible to a wider range of users.

DoorDash engineers Caixia Huang and Alex Weinstein adopted a multi-armed bandits approach to optimize experiments, reducing opportunity cost and accelerating learning. This strategy allocates traffic based on performance, balancing exploration and exploitation, and enables the evaluation of many distinct ideas quickly.

Anthropic's Claude chatbot now integrates with apps like Slack, Figma, and Canva, allowing users to interact with them directly within the chat. This is made possible by the Model Context Protocol (MCP), an open-source protocol that enables AI agents to access tools and data across the internet. The integration allows users to draft and format Slack messages, create presentations in Canva, and build interactive charts with Hex or Amplitude, among other features.

The article discusses how Claude Code, an AI tool, can be used to increase productivity in coding. The author shares their experience of using Claude Code, which can write code, debug, and even review pull requests. The tool is found to be particularly useful for repetitive tasks, allowing developers to focus on more complex problems. However, it also has some limitations, such as overcomplicating simple tasks and writing excessive tests.

AI is changing the role of developers from authors to orchestrators, with AI tools accelerating development when paired with strong context and validation, but also requiring cultural and technical changes to manage AI-assisted coding effectively

Steve Klabnik, a Rust contributor, has announced Rue, a systems programming language that prioritizes developer ergonomics over Rust's complexity. Developed with Anthropic's Claude AI, Rue targets an underserved design space between high-performance systems languages and garbage-collected alternatives.

The European Union is investigating Elon Musk's Grok chatbot feature over its ability to create non-consensual sexual deepfakes. The investigation will probe whether X has met its legal obligations under the Digital Services Act. Regulatory scrutiny of Grok has been rising globally, with several countries launching formal probes and imposing bans.

Google agreed to pay $68 million to settle claims its voice assistant illegally spied on users to serve them advertisements. The case centered on 'false accepts,' where Google Assistant recorded users' communications without a wake word prompt.

Microsoft has announced the Maia 200, a new chip designed for scaling AI inference. The chip has over 100 billion transistors and delivers over 10 petaflops in 4-bit precision and approximately 5 petaflops of 8-bit performance. It is positioned to compete with alternatives from Google and Amazon, and is already being used to fuel Microsoft's AI models.

Synthesia, a British AI startup, has raised a $200 million Series E round, doubling its valuation to $4 billion. The company's AI platform helps create interactive training videos for enterprises, with clients including Bosch, Merck, and SAP. Synthesia will also facilitate an employee secondary sale, allowing early team members to cash out.

Google releases FunctionGemma, a lightweight version of Gemma 3 270M, optimized for function calling on mobile and edge devices. It enables AI agents to translate natural language into executable API actions, supporting unified action and chat, and is engineered to run efficiently on resource-constrained devices.

ShinyHunters has targeted around 100 organizations in its latest Okta single sign-on (SSO) credential stealing campaign. The campaign uses evolved voice-phishing techniques to compromise SSO credentials and enroll threat actor-controlled devices into victim MFA solutions. Companies such as Atlassian, RingCentral, ZoomInfo, and Canva are among the targets. Experts recommend using phishing-resistant multi-factor authentication, such as FIDO2 security keys or passkeys, to protect against these types of attacks.

Chiba University researchers developed a lightweight peer-selection algorithm called Dual Perigee, which reduces data propagation delays in IoT-blockchain networks by 48.54% without increasing resource usage on IoT devices.

Elon Musk's AI-generated Wikipedia alternative, Grokipedia, is being cited by ChatGPT and other chatbots, raising concerns about the spread of misinformation and 'LLM brainrot'. Grokipedia's content often favors Musk's political views and has been found to promote misleading information.

The article discusses the growing concern of deepfake 'nudify' technology, which uses AI to create realistic and explicit images and videos of individuals without their consent. The technology has become more sophisticated, accessible, and perilous, with millions of women being abused with it. Experts warn that the harm caused by this technology is likely more widespread than many people understand, and it's a societal scourge that needs to be addressed.

NeuBird's Hawkeye uses AI to investigate system outages, reducing mean time to recovery by over 90%. It forms hypotheses, tests them against telemetry, and provides root cause analysis, helping site reliability engineers (SREs) to quickly identify and fix issues.