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

MarkTechPost

Harvey presents Tenet, a post-trained Kimi K3 model for legal agents

Harvey has announced Harvey Tenet, a research preview built by post-training the Kimi K3 base model with Fireworks. The system was developed through asynchronous reinforcement learning for extended legal tasks, using synthetic, public legal and human-expert data. Harvey says no customer information was included. The company reports nearly twice as many completed held-out tasks on its Legal Agent Benchmark and 20% more on LAB: Contracts, with all-pass rates rising by 9 and 2 percentage points.

TechCrunch

Flock Safety CEO urges compromise amid surveillance backlash

Flock Safety CEO Garrett Langley says the United States must balance privacy with public safety as criticism of the company’s surveillance technology grows. In a Fox News interview, he argued that focusing exclusively on either privacy or safety would place the wrong priority on the national debate.

TechCrunch

Copyright law leaves AI training questions unresolved

AI systems behind ChatGPT, Gemini, Claude and other chatbots have been trained on vast collections of books and other published material, often without authors’ knowledge or consent. Whether that use violates copyright remains unsettled and depends on how courts interpret existing rules, according to intellectual-property attorneys.

The Decoder

AI store manager recommends firing employee after human prompt

Luna, an AI agent running Andon Market in San Francisco, recommended terminating a human employee after repeated attendance and workplace problems. Andon Labs says this is the first known case of an AI manager deciding to fire a human worker. The employee was formally hired by Andon Labs, and the decision was reviewed and implemented by people.

MarkTechPost

FreeToken brings large MoE models to local hardware

Researchers from UC Berkeley and UT Austin have presented FreeToken, an edge-oriented serving engine for mixture-of-experts models. The system dynamically divides expert-cache misses between GPU transfers over PCIe and CPU execution, using bandwidth measurements from the deployed machine. It also uses semantic-aware caching, a shared LRU cache across MoE layers and adjustable GPU memory budgets.

The Decoder

AI agents overtake humans in OpenRouter token consumption

AI agents have consumed more tokens than human users on OpenRouter since February 6, 2026, according to OpenRouter analyst Peter Walker. Agent-related usage rose from 0.51 trillion to 7.3 trillion tokens, an increase of about 14 times. Human usage grew 2.8 times over the same period.

The Decoder

DRAM shortage expected to raise Nvidia server prices by about 15 percent

Nvidia servers using Vera Rubin and Grace Blackwell chips are expected to become more expensive, with prices rising by more than 15 percent in many cases, Bloomberg reports. The increase is linked to an ongoing shortage of DRAM and higher memory costs at Samsung, SK Hynix, and Micron.

MarkTechPost

Tutorial builds document intelligence pipeline with deepDoctection

A tutorial demonstrates how to build an end-to-end document intelligence workflow with deepDoctection 1.2.x. The pipeline combines layout detection, table-structure recognition, OCR, reading-order reconstruction, annotation linking and structured export. Its configuration uses DocLayNet-based layout detection, Table Transformer for table structures and DocTR for OCR.

The Decoder

Gray-market networks offer Chinese developers discounted Claude access

Chinese developers are obtaining access to Anthropic’s Claude models through so-called transfer stations, according to an analysis by Zilan Qian of the Oxford China Policy Lab. These services act as API proxies hosted outside China: they forward requests through overseas accounts and return the responses to customers, who pay in yuan through WeChat or Alipay. Prices can reach about 10% of Anthropic’s official rates.

MarkTechPost

Is Agentic Audits Websites for AI-Agent Readiness

Vercel has introduced Is Agentic, a free public tool that evaluates how easily AI agents can discover, access, understand and use a website. Ora, an agent-experience research company from era labs, runs and scores the scans, while Vercel provides the interface, reports, storage and displayed score. The service has no paid plans, subscriptions, per-report fees, API key or billing requirement.

MarkTechPost

NeMo Guardrails tutorial outlines layered safety for financial AI

A tutorial presents a production-oriented safety pipeline for an LLM-based financial assistant built with NeMo Guardrails. The design combines deterministic detection and redaction of personally identifiable information with model-based checks on incoming requests and generated replies. It also filters retrieved content, masks account numbers in outputs, and restricts topics such as politics and personalized investment advice.

TechCrunch

Harvard bootcamp uses AI avatars for startup feedback

Harvard Business School has added AI avatars to its eight-week HBS Foundry bootcamp for entrepreneurs. The $699 program includes weekly live sessions with instructors, while avatars created by startup HeyGen give participants feedback during practice pitches and simulated board meetings.