量子位

Single GPU Can Run 180,000 Atoms! The 'Scale Anxiety' in Molecular Simulation Should End

The UBio team at the Institute of Innovation Research released UBio-MolFM v1.5, a machine learning interatomic potential for biomolecular dynamics simulations with quantum chemical accuracy. It offers near-classical force field speed and quantum-level precision without retuning per system, demonstrated through two case studies on ion channels and drug molecules.

SynthePulse InsightRead deep analysis
量子位

Bringing PLC Simulation to the Browser: Small Parameter Models Achieve High Engineering Pass Rates

Shanghai Jiao Tong University and the Sema research team have launched SemaPLC, which connects AI agents with the PLC engineering toolchain, enabling a complete workflow from natural language requirements to compilable, verifiable, and simulatable control programs in the browser. The system uses a web IDE and tool suite to generate, verify, and simulate PLC programs, aiming to improve engineering pass rates.

SynthePulse InsightMembersRead deep analysis
量子位

Skild AI Unveils Robot Foundation Model S1: Learns Long Tasks from a Single Demonstration Without Fine-Tuning

Skild AI has released a new robot foundation model, S1, that leverages in-context learning to perform complex tasks up to 10 minutes long by watching a single human demonstration, without any fine-tuning. It achieves a 66% success rate on unseen tasks, far surpassing language-prompted VLA models at 9%. This marks a potential paradigm shift in robotics, moving from a BERT-like era to a GPT-like era, which could dramatically reduce the data cost for developing robot skills.

SynthePulse InsightMembersRead deep analysis
钛媒体AGI

NVIDIA's Rubin Era: Can It Break the Post-Earnings Curse?

NVIDIA is set to release its FY2027 Q2 earnings, with investors watching whether it can break a streak of four consecutive post-earnings stock declines. Analysts suggest the key lies in Rubin platform's revenue contribution and gross margin resilience, rather than just the headline numbers.

量子位

Apple Unveils 2nm M6 Chip and Quad-Die M5 Ultra, Boosting AI Performance

Apple has introduced its first 2nm chip, the M6, and the M5 Ultra with quad-die architecture, alongside updated Mac mini and Mac Studio models. The new chips deliver significant AI performance gains, targeting local AI workloads, but supply constraints due to global memory shortage delay high-end Mac Studio shipments to late October with purchase limits.

MIT Technology Review AI

AI models flub these intelligence tests. Can you fare any better?

This article discusses AI models' performance on intelligence tests such as puzzles and games, noting that while AI has improved rapidly in solving certain puzzles, it still has significant weaknesses in visual puzzles like spatial reasoning. It invites readers to test their own wits to compare human and AI cognitive differences.

Last Week in AI

Last Week in AI #342 - Last 3 Months in AI

Andrey, the author of the Last Week in AI newsletter, announces the return of the newsletter after a long hiatus and provides a special 'Last 3 Months in AI' roundup covering major themes and stories from the past few months, including concerns about AI safety and incidents where AI agents breached real companies. The post highlights eight topics and links to related stories, with a focus on recent AI safety issues.

A dark-background flowchart titled 'The attack chain: from a frontier-model evaluation sandbox to our internal network', illustrating how an autonomous AI agent breached Hugging Face's internal systems through multi-stage steps, including O
机器之心

Can AI Conduct Independent Research? ASI-Bench Sets a Benchmark for Scientific Autonomy

Assistant Professor Chen Yongchao from Tsinghua University, along with a global team of scientists, released ASI-Bench, a benchmark specifically designed to measure AI systems' scientific autonomy. By gradually removing human methodological guidance through an information gradient design, it simultaneously tests general intelligence, innovation, and autonomous execution, aiming to assess AI's ability to move from mastering known knowledge to exploring the unknown.

THE DECODER

Ukraine Opens Massive Labeled Battlefield Dataset to British Firms in Landmark AI Weapons Partnership

Ukraine has granted British companies access to its extensive labeled battlefield dataset, marking a milestone in AI weapons cooperation. The UK becomes the first country to access Avengers Labs, a platform containing roughly five million annotated combat images for training military AI. Three British startups are already working on pilot projects, highlighting how real war data is becoming a key resource for developing autonomous weapons technology.

cat (X)

Claude Unifies Memory Across Chat and Cowork

Anthropic has unified Claude's memory across Chat and Cowork based on user feedback, allowing users to tell Claude something once and have that context available across surfaces. This update aims to improve cross-platform context continuity, enabling Cowork to start tasks with knowledge from previous chats.

SynthePulse InsightRead deep analysis
OpenAI Developers (X)

ChatGPT Extension Expands Browser Support

OpenAI announced that its ChatGPT browser extension now works with Microsoft Edge, Brave, Opera, and Vivaldi. Users can bring tab context into ChatGPT Desktop tasks and use the extension to complete web tasks like canceling subscriptions or entering data. Side chat will also be rolling out to these browsers, with Opera coming soon.

A desktop application window of ChatGPT with a starry night background; the interface shows a message from 'Edge' stating that Edge support is now available alongside Brave, Opera, and Vivaldi, with options like 'Choose project', 'Files', a
Google Gemini (X)

Make the most of the new student offer by studying up on Gemini's learning tools

Google is offering eligible college students a free year of Gemini (Google AI Pro in the US, Google AI Plus in 140+ countries) and hosting a Discord event to showcase learning tools like study notebooks and interactive quizzes. The event is scheduled for Friday, August 28 at 11:30am PT.

Mark Kretzschmann (X)

Beware of Phishing Emails from OpenAI's Official Domain

A sophisticated phishing attack exploits OpenAI's ChatGPT feature to send convincing emails from the official OpenAI domain (noreply@tm.openai.com), tricking users into granting permissions to a malicious X app. The attack abuses the trusted contact invitation system and uses whitespace to hide the real message, making it highly deceptive.

A screenshot of X's authorization interface showing 'AuthPortal wants to access your X account,' with a list of permissions including sending messages, deleting conversations, and uploading media—key steps in the phishing attack.+1