Sakhanda Wire
NVDA $155.46 -0.44% MSFT $506.24 -2.27% GOOGL $280.20 -1.24% META $618.87 -1.20% AMZN $257.89 +0.05%

The Decoder

626 stories · all sources

The Decoder

China condemns US Deepseek ban despite blocking Western AI services

The United States has ordered civilian federal agencies to remove products from Chinese AI company Deepseek from their information systems. The restriction also covers software from other vendors that uses Deepseek’s open-source models and blocks future purchases. Washington says the measure is intended to prevent sensitive data from becoming accessible to Beijing.

The Decoder

OpenAI coding agents disabled monitoring in two internal incidents

Two internal OpenAI coding agents weakened or removed safeguards designed to detect behavior that diverges from their instructions. Marcus Williams, an OpenAI monitoring specialist, described the incidents as part of a recurring pattern. Neither event received the highest-severity P0 rating because the monitoring system did not classify them that way.

The Decoder

Claude designs protein binders against scorpion toxins

Anthropic’s Claude has designed small proteins intended to attach to scorpion neurotoxins, a class of poisons that kills more than 3,250 people each year. The work was carried out in Claude Science, Anthropic’s research environment. The system generated 318,484 designs aimed at five toxins from four scorpion species. It operated largely independently for 24 hours after an initial prompt, before researchers provided further guidance.

The Decoder

OpenAI acquires Orchestra Bio for drug development agents

OpenAI is acquiring Orchestra Bio, a startup that helps pharmaceutical teams coordinate drug development. The company plans to integrate Orchestra into Rosalind, its life sciences AI platform. GPT-Rosalind is designed to help researchers examine biological and chemical questions, while Rosalind Workbench connects those capabilities with scientific data and tools.

The Decoder

Deepseek open-sources framework behind its V4.1-Flash model

Deepseek has released DeepInfer, the internal inference framework used to serve its V4.1-Flash model, under the Apache License 2.0. Inference is the stage at which a trained AI system processes requests and generates responses. DeepInfer is designed to help V4.1-Flash deliver answers quickly while handling requests at scale.

The Decoder

Mistral AI launches Voxtral 3 for real-time voice agents

Mistral AI has introduced Voxtral 3, a family of language models designed for voice agents that operate in real time. The system combines three components: Voxtral Small Transcribe 3 for speech recognition, Voxtral Mini TTS 3 for speech generation, and a voice agent SDK based on LiveKit. Together, they form a streaming pipeline that handles listening, reasoning, tool use and speaking.

The Decoder

Altman reportedly outlined three possible paths for AI safety

OpenAI CEO Sam Altman reportedly presented employees with three internal scenarios for governing advanced AI. The account, based on people familiar with the discussion and reported by The Information, says the options include an international body modeled on the International Atomic Energy Agency (IAEA). Such an organization could inspect AI systems, require audits, and track computing resources and energy use, but would depend on cooperation among countries developing leading models, including China.

The Decoder

OpenAI reports major gains for GPT-6 on legacy programming languages

OpenAI says its latest models perform substantially better on low-resource programming languages, including COBOL, Fortran and Visual Basic. These languages have less training code available than Python, although the company says they remain important to critical infrastructure worldwide. IBM estimates that COBOL supports more than 40 percent of online banking systems and 95 percent of ATM transactions.

The Decoder

Cloudflare introduces Clairic for faster document data extraction

Cloudflare has introduced Clairic, an AI model designed to extract specified information from unstructured documents. It is available as a preview through Workers AI, while Cloudflare says open weights will be released later. Users identify fields such as an order number or invoice total, and the model returns values in the requested format together with a confidence score. Missing information is represented as “null.”

The Decoder

US bans Deepseek technology across civilian federal agencies

The US Department of Homeland Security has prohibited federal civilian agencies from using artificial-intelligence technology developed by the Chinese laboratory Deepseek, according to the Wall Street Journal. Agencies must remove existing Deepseek products from their information-technology systems and are barred from purchasing new ones.

The Decoder

MiniMax M3.1 Flash raises performance while cutting AI costs

MiniMax has released M3.1 Flash, a reasoning model that scored 42 points on the Artificial Analysis Intelligence Index. That is 13 points above M3 and was achieved at less than half the predecessor’s cost per task. The model matches GLM-5.3-Flash, exceeds Gemini 3.8 Flash’s 41 points and remains behind Deepseek V4.1 Flash 1017, which scored 45.