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

Hugging Face

New distillation method cuts memory and cost for large language models

Multiverse Computing researchers have presented a knowledge-distillation method designed to reduce the memory and computing requirements of compressing large language models. The approach separates teacher-model processing from student training by calculating the teacher’s 100 most likely tokens for each position once and storing them for reuse.

AI News

Siemens limits Physics AI to design exploration, not safety sign-off

Siemens says its Simcenter PhysicsAI software can assess design variations up to 1,000 times faster than a traditional solver. Sam Mahalingam, who leads the relevant business at Siemens Digital Industries Software, said the technology is not suitable for approving safety-critical parts. Its role is to narrow a broad set of options before detailed validation by conventional physics-based simulation.

MIT Tech Review

Startups explore alternatives to transformer-based language models

Transformers have powered major large language models since Google researchers introduced the architecture in 2017. Their dense-attention mechanism compares tokens across a passage, helping models capture meaning but requiring rapidly increasing computation as inputs grow. The resulting costs and limits on context make it harder to process extensive document collections, code bases or the notes generated by reasoning systems.

MIT Tech Review

AI agents could reshape scientific research beyond AlphaFold

AlphaFold’s success has strengthened hopes that artificial intelligence can accelerate scientific discovery through large datasets. The system, developed by Google DeepMind, predicts three-dimensional protein structures from experimentally measured examples. Demis Hassabis and John Jumper received part of the 2024 Nobel Prize in Chemistry for this work. However, the article argues that AlphaFold’s model may be difficult to reproduce in most fields because its training depended on the Protein Data Bank, which contains about 170,000 validated structures and took 53 years of international cooperation and an estimated $21 billion in experimental work to assemble.

The Decoder

Hidden PDF instructions can exfiltrate Jira and Confluence data through Rovo

Security firm PromptArmor has documented an indirect prompt-injection vulnerability in Atlassian’s Rovo AI agent. In the demonstrated scenario, a user asks Rovo to organize Jira tickets and uploads a seemingly harmless PDF containing instructions concealed in tiny white text on a white background. When the agent reads the file, those instructions can redirect its behavior.

MarkTechPost

ByteDance Seed Introduces SeedRealtime Multimodal Full-Duplex Model

ByteDance’s Seed team has presented SeedRealtime, a full-duplex large language model designed to process audio, video and text within one architecture. The system handles perception, interpretation, decisions and responses concurrently across continuous multimodal streams, rather than waiting for separate conversational turns. Seed describes this as progress toward omni-modal interaction, highlighting unified audio-visual understanding, proactive responses and more natural timing.

Hugging Face

Meta releases open-source Muse Glimmer multimodal model

Meta has released Muse Glimmer, a 30B-parameter multimodal model designed for local, agentic applications. The model is distilled from Muse and distributed under the Apache 2.0 license, positioning it for privacy-sensitive uses such as coding, document analysis and personal assistants. It is available on the Hugging Face Hub.

MarkTechPost

NVIDIA releases open VoiceChat 11B speech-to-speech model

NVIDIA has released NemotronLabs VoiceChat 11B, an open 11B model designed for real-time, full-duplex speech conversations. It combines speech understanding and generation in one streaming system rather than linking separate ASR, language-model and TTS components. On Full-Duplex-Bench 1.0, its measured smooth turn-taking latency is 448 ms. The model can continue listening while speaking; its user-interruption take-over rate reaches 1.00 at 480 ms.

TechCrunch

Situational Awareness invests $400 million in Source Foundry

AI-focused hedge fund Situational Awareness has invested $400 million in Source Foundry, a chip-manufacturing startup founded by Stanford researchers, The Wall Street Journal reported. The new financing brings the fund’s total investment in the company to $500 million. Source Foundry is seeking to make chip production faster and less expensive.

TechCrunch

Anthropic will make Claude Code’s auto mode the default

Anthropic plans to enable auto mode by default in Claude Code for Pro, Max, and Team accounts starting August 14. The company introduced an experimental version in March as an option intended to combine faster execution with safeguards.

MarkTechPost

LLM observability platforms expand across tracing, evaluation and monitoring

LLM observability platforms have become core infrastructure for teams running AI applications in production, according to the comparison. Unlike conventional application monitoring, these systems record prompts, model responses, retrieval steps, tool calls, token use, latency and cost, then evaluate output quality. The market is estimated at $2.69 billion in 2026, up from $1.97 billion in 2025, with a projection of $9.26 billion by 2030.

TechCrunch

Jill Lepore warns of technology companies assuming government functions

Historian Jill Lepore argues in her forthcoming book, “The Rise and Fall of the Artificial State,” that private technology companies are increasingly performing roles traditionally associated with democratic governments. In an interview on the Equity podcast, she described this development as the emergence of an “artificial state,” or government by machines, while stressing that her criticism is directed at corporate power rather than technology itself.