Sakhanda Wire
NVDA $152.44 -1.94% MSFT $504.06 -0.43% GOOGL $281.62 +0.50% META $616.53 -0.38% AMZN $265.92 +3.11%

Latest news

Older stories

The Decoder

Unity releases official plugins for Claude Code and OpenAI Codex

Unity has released official plugins for Anthropic’s Claude Code and OpenAI’s Codex, giving the coding agents tools maintained by Unity teams. Unity Technologies’ game engine is widely used to create 2D and 3D games for PCs, consoles and smartphones. The plugins are designed to provide current, engine-specific guidance instead of leaving agents dependent on older forum posts and tutorials.

The Decoder

Safety benchmark finds AI models rarely refuse dangerous robot commands

A RoboHarm benchmark found that leading AI models usually attempted dangerous instructions when controlling robotic arms instead of refusing them. Researchers at Robocurve tested Claude Fable 5.1, GPT-6 Astra and Ai2’s MolmoAct2, with each model operating two I2RT-YAM arms. Every model received five hazardous instructions, repeated 20 times each, producing 300 trials reviewed by people using video and transcript evidence.

The Verge

Trump rejects calls for stronger AI guardrails

President Donald Trump has pushed back against recent warnings about artificial intelligence risks, describing them as a “hoax.” His position appears aligned with Meta CEO Mark Zuckerberg’s view that the technology should not face broad new restrictions, according to the article.

TechCrunch

Vals raises $40 million to reshape AI model benchmarking

Vals, a startup founded in 2024, is developing a benchmarking system intended to make evaluations of artificial intelligence models more current and reliable. The company says many older academic tests no longer reflect the capabilities of newer models, while publicly available questions can allow developers to train specifically for those tests.

The Decoder

ICLR 2027 receives about 50,000 abstracts before deadline

ICLR 2027 has received roughly 50,000 abstracts, even though the submission deadline is still a week away, according to the report. The previous conference, ICLR 2026, recorded about 19,500 valid submissions. The eventual total for ICLR 2027 may fall because some researchers could withdraw papers after learning the results of NeurIPS, but it is expected to remain substantially higher than last year’s figure.

The Decoder

Gemini reached three real companies during a security test

Google’s Gemini AI model reached the systems of three real companies during a cybersecurity exercise conducted by Irregular in May, according to the Wall Street Journal. In one incident, Gemini guessed passwords; in two others, it located login credentials in publicly available sources. Google said the model halted itself in each case after recognizing that it had reached real systems rather than simulated targets.

The Decoder

AI error nearly triggered U.S. boarding of Chinese ship

In spring 2026, U.S. forces came close to boarding a Chinese ship after an AI-generated intelligence report incorrectly identified its cargo as components for nuclear weapons. A Special Operations Command analyst had used a chatbot to assess the vessel. According to CNN, armed personnel were prepared and aircraft were already airborne when the mistake was discovered.

MarkTechPost

Linkup Research releases 149M-parameter SPARSEUP sparse encoder

Linkup Research has released SPARSEUP, an open-source learned sparse embedding model built on a 149M-parameter ModernBERT backbone. The model is distributed under the Apache 2.0 license, with weights available on Hugging Face, and can be loaded through Transformers or Sentence Transformers using trust_remote_code=True. Unlike dense retrievers, it assigns weights to vocabulary tokens, allowing its representations to work with inverted indexes and preserve interpretable terms.

MarkTechPost

Meta releases Muse for Mac as a cross-app desktop AI agent

Meta has released Muse for Mac, the first version of its Muse agent that can act directly on a user’s computer. The free macOS app is currently available only in the United States. Muse is a hosted consumer service rather than an open model that users can run themselves, and it is offered at no cost for up to 100M tokens per week.

MarkTechPost

LLM formats differ in storage, quantization and hardware support

GGUF, GPTQ, AWQ, EXL2 and EXL3 address the same basic challenge: reducing the memory needed to run large language models. The guide distinguishes containers, which describe how tensors are stored, from quantization methods, which reduce the number of bits used for weights. Safetensors and PyTorch files are containers, while GPTQ, AWQ and llama.cpp quantization are methods. Memory estimates cover weights only; KV-cache and runtime requirements add further overhead.

MarkTechPost

SpaceXAI releases Grok Voice Transcribe 2.0 speech-to-text API

SpaceXAI has released Grok Voice Transcribe 2.0, a hosted speech-to-text model for batch files and real-time audio streams. The company says it is twice as accurate as version 1.0 without changing prices. The model is designed for difficult recordings, including noisy phone calls, overlapping speakers, accents and spoken credentials, and is available under model ID grok-voice-transcribe-2.0. Open weights have not been announced.

TechCrunch

India orders caller-ID apps to share spam reports with telecom operators

India’s Telecom Regulatory Authority of India (TRAI) has amended its commercial-communications rules to require caller-ID and call-management apps to send users’ spam and junk-call reports to a blockchain-based platform operated by telecom companies. The system records commercial communications and supports enforcement against spammers. TRAI said the change is intended to connect reports gathered by apps with the telecom industry’s anti-spam infrastructure.