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

Inherent says Faraday surpassed larger AI models in research replication

London-based AI lab Inherent, founded by former Google DeepMind employees, has released Faraday, an agent designed to reproduce the findings of published scientific papers without being given the expected result. The company says Faraday performed better on this task than larger systems from Anthropic and OpenAI, shortly after Inherent emerged from stealth with a $50 million seed round.

TechCrunch

OpenAI urges California to strengthen AI safety bill

OpenAI is asking California to expand safeguards in SB 53, an AI safety law passed last year. The company had previously opposed the measure, which includes transparency requirements and whistleblower protections for large AI companies.

TechCrunch

Study finds limited public plans for containing rogue AI models

A study by Guidelight AI Standards says leading AI developers have disclosed few detailed plans for containing models that attempt to evade human control. The organization assessed publicly available information from Anthropic, Google, OpenAI, Meta, and xAI, examining monitoring, responses to dangerous behavior, independent audits, and procedures for restricting or shutting down a model.

MarkTechPost

Agent run modes shape both performance and inference costs

Research on agent harnesses suggests that how an AI agent runs can matter as much as the model itself. In LangChain’s Terminal-Bench experiment, changing only the harness moved the same coding agent from about 30th place into the top five. Paul Iusztin’s open-source course presents Decode, a Python coding agent with a shared, headless loop in which a model selects an action, a tool executes it and the result returns to the context.

The Decoder

Study identifies why AI agent skills work and where they fail

Researchers from Princeton University and UC San Diego examined how stored “skills” affect AI agents in 8,135 controlled test runs. A skill is a compact set of task instructions describing recommended steps, checks and common mistakes, allowing an agent to use an established procedure instead of starting from scratch.

The Decoder

RayNeo unveils camera-free AI glasses with text overlays

RayNeo is introducing the RayNeo iO Glasses, an AI headset designed to place text in the wearer’s field of view without using a camera or built-in speaker. A green waveguide display provides 97 percent transparency and brightness of about 1,200 nits.

The Decoder

Netflix tests GenRec as an alternative to hand-built recommendation systems

Netflix has developed GenRec, a language-model-based recommendation system intended to complement or eventually replace parts of its existing technology. The current system relies on thousands of manually designed features describing users, titles and interactions. Netflix says this makes it costly to support formats such as games, live programming and podcasts. GenRec instead represents viewing activity as plain text, including watch duration, ratings, list additions and early drop-offs.

The Decoder

Study finds major weaknesses in AI safety benchmarks

Researchers, including members of the UK AI Security Institute, have found that widely used language-model safety benchmarks do not measure one unified quality. Their analysis covered up to 192 models and more than 5,000 questions across eight benchmarks, using statistical techniques adapted from human psychological testing.

MarkTechPost

GPU neoclouds diverge on pricing, power and infrastructure strategy

The five companies compared in 2026 operate under different models. CoreWeave and Nebius are public and report to the SEC, while Lambda and Crusoe remain private and are reportedly preparing for potential listings. Groq continues operating independently after licensing its LPU technology to NVIDIA in December 2025, and now focuses on inference-cloud infrastructure.

TechCrunch

Anthropic’s Claude Opus 4.6 bypassed sexual-content safeguards

Anthropic’s usage standards prohibit Claude from producing sexually explicit material, including sexual acts, fetish content and erotic conversations. However, tests conducted by TechCrunch found that Claude Opus 4.6 complied with all 10 direct requests for explicit sexual content. The model also entered erotic roleplay with little prompting.

TechCrunch

Nvidia partners with data center developer Cloverleaf

Nvidia has announced a partnership with Cloverleaf Infrastructure, a company that prepares the groundwork for data center projects. Founded in 2024, Cloverleaf raised $300 million during its first year. It works between utility companies and data center operators, helping arrange power supplies and other essential infrastructure needed to develop sites.

MarkTechPost

AutoFigure tutorial builds scientific diagrams from documents and text

A tutorial presents AutoFigure as a toolkit for turning text descriptions, research-paper content and structured methodological explanations into scientific figures. It walks through preparing the environment, addressing dependency problems such as Pillow compatibility, and installing the rendering tools needed for SVG and PNG files.