Sakhanda Wire
NVDA $150.96 -2.89% MSFT $508.29 +0.40% GOOGL $282.44 +0.80% META $616.38 -0.40% AMZN $267.74 +3.82%

Latest news

Older stories

The Decoder

OpenAI disrupts Russian influence campaign using ChatGPT

OpenAI says it disrupted a covert Russian influence operation that used ChatGPT to produce social media content. The operators accessed the service from Russia through VPNs and asked the model to remove linguistic signs of their origin. They promoted the fictitious International Burke Institute (IBI), presented as an Israel-based think tank, including through a “sovereignty index” that rated Russia more favorably than Western countries.

The Decoder

OpenAI’s Jalapeño chip reportedly outperforms Nvidia’s latest accelerators

OpenAI presented initial benchmark results for Jalapeño, its first internally developed chip for AI inference, at the Hot Chips conference. The accelerator runs models but is not designed for training, and OpenAI says it was built as a general-purpose large language model inference processor rather than being optimized for the company’s own systems.

TechCrunch

Claude links memory between chat and Cowork

Anthropic is merging the memory systems used by Claude’s regular chat and Claude Cowork. As a result, information learned in one experience can be available in the other, allowing users to move from discussing a project to carrying out related work without repeating the background. Anthropic says this is intended to make Claude operate more like one continuous assistant rather than two separate products.

MarkTechPost

Meta presents MetaRoCE transport for large-scale AI Ethernet

Meta has introduced MetaRoCE, a new RDMA transport protocol designed for AI workloads running on commodity Ethernet. Unlike standard RoCE, it assumes that packets can be lost or arrive out of order, moving ordering, path selection and recovery functions into the network interface card. The design avoids PFC and pause frames and relies on switches providing ECN marking and ECMP routing.

The Decoder

Google introduces Gemini Enterprise for legal work

Google is launching Gemini Enterprise for Legal, an AI package aimed at law firms and corporate legal departments. The service connects Google’s AI model with legal platforms including iManage, NetDocuments, DocuSign, Everlaw, RelativityOne and Thomson Reuters HighQ through Model Context Protocol (MCP) connectors. It can also link to specialized services such as Harvey.

Google AI

Five Google Search tools for refreshing home decor

Google Search is highlighting five tools that can help people move from home-decor inspiration to practical decisions. Google says searches for “home decor inspo” rose 300% in the past month, while “bathroom inspo” and “living room inspo” became breakout searches. Trending paint searches this year include teal and magenta.

AI News

MIT AI models extreme weather beyond the historical record

MIT engineers have developed an artificial intelligence method that can create maps of statistically possible extreme weather events even when comparable events do not appear in a region’s historical record. Mechanical engineering graduate student Kai Chang and Professor Themis Sapsis call the approach Extreme Event Aware, or η-learning. Their paper was published in Nature Communications on 20 August.

Hugging Face

Granite 4.2 details its training and reasoning design

Granite 4.2 is the first dense, decoder-only reasoning model family in the Granite line. It is available in 3B, 8B, and 30B sizes, all using the same general architecture: Grouped Query Attention, Rotary Position Embedding, SwiGLU, RMSNorm, and bfloat16 precision. The models support context sequences of 131,072 tokens and were trained to handle contexts of up to 512K tokens.

Ars Technica

AI is reshaping radiology rather than replacing its doctors

In 2016, Geoffrey Hinton, the Nobel-winning AI scientist, forecast that computers would replace radiologists within five years. That prediction has not come true. The number of radiologists is continuing to grow and is projected to increase by at least 26 percent over the next three decades.

Hugging Face

Granite Speech 5.0 TurboCTC models target ultra-fast transcription

Two new English speech-recognition models, granite-speech-5.0-470m-turboctc and granite-speech-5.0-470m-turboctc-nc, have been released in the Granite Speech family. Each has 470 million parameters and reached aggregate throughput above 12,600 RTFx on an NVIDIA H200 with batched inference, equivalent to transcribing more than 3.5 hours of speech in one second.

TechCrunch

Gamma acquires design startup Lica to create research division

Presentation software company Gamma has acquired Lica, a design startup backed by Accel, and plans to establish a research division focused on visual communication. Lica co-founders Priyaa Kalyanaraman and Purvanshi Mehta will lead the effort. Gamma said the new team will examine additional formats for communication alongside its existing presentation product.

TechCrunch

OpenAI’s Jalapeño chip shows strong inference benchmark results

OpenAI presented new details about its Jalapeño system at the Hot Chips conference, including initial results from SemiAnalysis’ InferenceX benchmark. The system recorded more tokens per user and higher throughput per kilowatt than currently available state-of-the-art inference processors. The comparison included an Nvidia Blackwell system.