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

TechCrunch

Anthropic explains how Claude’s text watermarking will work

Anthropic has provided more details about its plan to watermark text generated by Claude. The company says the system will encode a pattern through low-stakes word choices, such as selecting one of several similarly suitable descriptions. The pattern should be invisible to readers but detectable with the appropriate key, and Anthropic says it will not affect the quality or appearance of responses.

TechCrunch

SpaceX completes acquisition of AI coding startup Cursor

AI coding startup Cursor has officially become part of SpaceX, according to an announcement on Cursor’s blog. The companies had announced a technology-development agreement in April that included an option for SpaceX to acquire Cursor for $60 billion. Two months later, after SpaceX became a public company, they said they would proceed with the acquisition.

The Decoder

Nvidia cuts planned OpenAI data-center guarantee amid investor concerns

Nvidia has reduced its planned financial guarantee for OpenAI’s data-center project in Ohio from $250 billion to just under $120 billion, according to the Wall Street Journal. The revised commitment would cover the project’s first construction phase, expected to provide about five gigawatts of capacity. OpenAI is separately negotiating a lease for the full 10-gigawatt facility being developed by SB Energy, a SoftBank subsidiary.

MarkTechPost

Guide outlines LoRA fine-tuning for tool-calling language models

A tutorial presents an end-to-end supervised fine-tuning workflow for the XYZ-Aquila-SFT dataset using Hugging Face Transformers, PyTorch and PEFT. The process streams and inspects examples, parses multi-turn tool-use trajectories, extracts structured calls, and examines corpus characteristics such as message depth, trajectory size and tool usage.

The Decoder

Connecticut court strips plaintiff of e-filing access after hidden AI prompts

Matthew Elliott, a self-represented plaintiff in Connecticut, placed hidden instructions in court filings in an apparent attempt to influence a possible automated review. The text was set in 3-point white type on a white background, making it difficult for people to see while remaining readable to language models. It told a hypothetical system to support the filing’s position and reconsider an earlier clerk’s denial.

The Decoder

New benchmark finds major visual weaknesses in leading AI models

Moonshot AI has introduced PerceptionBench, an evaluation designed to measure visual perception in multimodal language models without relying on reasoning or outside knowledge. Its tasks test ten abilities identified from real model mistakes: visual relations, counting, attributes, depth and 3D understanding, localization, comparison, fine-grained recognition, context integration, OCR, and hallucination.

The Decoder

Anthropic to offer API for detecting Claude-generated text

Anthropic says it will soon release an API that allows third-party developers to check whether text was produced with Claude. The system is based on a version of Google DeepMind’s SynthID Text method, published in Nature in 2024. It changes the randomness used when Claude selects words, creating a detectable pattern that Anthropic says does not alter the text’s content, creativity or readability.

Ars Technica

Connecticut judge identifies hidden AI prompts in court filing

A Connecticut judge has identified what appears to be the first case in which a US plaintiff concealed instructions in a court filing for an artificial intelligence system to read. The plaintiff, who accused a healthcare provider of improperly withholding access to records, embedded text that was invisible to human readers but legible to software processing the document.

The Decoder

Qwen3.8 models released with open Apache 2.0 weights

Alibaba’s Qwen AI team has released open weights for the Qwen3.8 model family under the Apache 2.0 license. The main model, Qwen3.8-27B, is a dense multimodal system with 27 billion parameters. Qwen says it surpasses the larger Qwen3.7-Plus in coding and office-related tasks, while also offering stronger agent capabilities for independent planning and more reliable task completion.

The Verge

Instagram introduces a redesigned wordmark

Instagram has introduced a redesigned wordmark this week, changing a visual element that had become familiar to users. The new lettering differs noticeably from the previous design, and the change has prompted questions about whether the platform’s name remains immediately legible. The source article does not provide an explanation from Instagram for the redesign.

The Decoder

Computer History turns Mac activity into a searchable ChatGPT timeline

OpenAI’s Computer History records interaction events across apps and websites on Mac computers, creating a searchable timeline that ChatGPT and Codex can use as context. It replaces the earlier Chronicle research preview, which used screenshots. The feature reads clicks, keystrokes, keyboard shortcuts and app changes through macOS accessibility tools. OpenAI says it does not capture screenshots, screen recordings, microphone input or system audio, and excludes private browsing activity.

The Verge

Google makes visible watermarks optional in Gemini and Flow

Google is introducing an option to remove visible watermarks from images, videos and music created with its artificial-intelligence tools. Users can switch off a new “Media watermark” setting in Gemini and in Flow, Google’s AI video generator.