OpenAI

How Endava is redesigning software delivery around AI agents

OpenAI

This article introduces the practice of technology service provider Endava restructuring its software delivery model around AI agents: the core takes AI agents as the core support of the delivery chain, paired with the implemented technical capabilities of ChatGPT Enterprise and Codex. It not only achieves the business goals of faster software delivery and automated redundant workflows, but also drives the establishment of an AI-native organizational culture within the enterprise, providing a reference for the intelligent transformation of similar enterprises.

Dreaming: Better memory for a more helpful ChatGPT

OpenAI

This research addresses the pain points of ChatGPT easily losing user preferences across sessions and weak context relevance, and proposes a brand new memory system called “Dreaming”. The system can long-term retain personalized preferences revealed in user interactions, break the limit of single-session context, and output more relevant responses that fit user needs even across multiple independent sessions, greatly improving ChatGPT’s interaction friendliness and practical value.

Anthropic News

Introducing Claude Opus 4.8

Anthropic

The official newly released Claude Opus 4.8 is an upgraded version of the Opus-class large model, with multi-dimensional improvements in core performance: it shows significantly enhanced performance in three core scenarios of code development, agent tasks, and professional field work, while optimizing long-running consistency. It can stably undertake longer-cycle, more complex continuous work, and is more suitable for high-demand professional use cases.

Expanding Project Glasswing

Anthropic

The core measure of the newly disclosed Expanding Project Glasswing is to expand the scope of the original project: the implementation coverage now adds more than 15 countries, and around 150 new cooperative organizations of various types are connected. This expansion will broaden the project’s geographic coverage breadth and multi-stakeholder collaboration depth, effectively improving the project’s implementation capacity in different regions and cross-domain radiation influence.

Google DeepMind

We’re launching the Google DeepMind Accelerator program in Asia Pacific to tackle environmental risks

Google DeepMind

Google DeepMind recently announced the launch of a special accelerator program in the Asia-Pacific region, focusing on environmental risk response scenarios. The project will rely on its leading AI technical capabilities, link up with regional scientific research teams, tech startups, relevant institutions and other entities, and collaboratively develop solutions for climate disaster early warning, ecological protection, carbon emission control and other directions, providing technical support for environmental governance in the Asia-Pacific region and helping address various climate and ecological challenges.

Fast-tracking genetic leads to reverse cellular aging

Google DeepMind

This study titled Fast-tracking genetic leads to reverse cellular aging uses a core method where biologists carry out genetic target screening with the Co-Scientist tool, successfully identifying new regulatory factors and achieving effective rejuvenation of human cells. This achievement not only greatly improves the efficiency of aging-related genetic research, but also provides new candidate targets for the subsequent development of anti-aging therapies and regenerative medicine.

Hugging Face Blog

Five labs, five minds: building a multi-model finance drama on small models

Hugging Face

Aiming at the pain points of high cost and hard-to-control compliance risks of large models in the financial content generation field, this article innovatively adopts a multi-small model collaboration route, uniting five laboratories to train exclusive small models adapted to financial logic verification, script creation, audio and video rendering, compliance review, and user interaction respectively, building a multi-modal financial drama generation system through distributed scheduling. Compared with large model solutions with the same effect, deployment cost is reduced by 83%, and compliance pass rate is increased by 32 percentage points, which can quickly generate customized financial popular science content suitable for investor education scenarios.

Nemotron 3.5 Content Safety: Customizable Multimodal Safety for Global Enterprise AI

Hugging Face

This article launches the Nemotron 3.5 content safety solution, built for global enterprise AI with customizable multi-modal security capabilities, supporting multi-modal risk content recognition for text, images and more, and can flexibly adapt to different regional regulatory rules and enterprise personalized risk control requirements. It has higher recognition accuracy than general security tools, can effectively reduce the compliance cost of enterprise AI global deployment, and meets the management requirements of diverse scenarios.

The Gradient

After Orthogonality: Virtue-Ethical Agency and AI Alignment

The Gradient

This AI alignment research carried out from the perspective of virtue ethics challenges the default goal-oriented assumption, pointing out that human rationality is not anchored to a fixed final goal, but actions adapt to a self-promoting practice network including action tendencies, evaluation rules and other elements. It proposes that if AI is to collaborate with humans, its decision-making logic needs to be isomorphic with human practical action logic, which can cover both ethical and core safety alignment requirements at the same time.

QbitAI

教你用AI一节课收17万，华尔街精英排着队付费

QbitAI

Former SoftBank financial practitioners Felipe and Dave Wang founded Wall Street Prompt, providing paid AI implementation training for Wall Street, with a single course priced at $25,000 (approximately 170,000 RMB). The course teaches how to use AI to analyze roadshow videos, mine stock price-sensitive information in earnings call recordings, and deconstruct ambiguous management statements into data usable for financial models. Top institutions such as Citigroup and Bank of America have invested heavily in AI layout but struggle with implementation, and are scrambling to pay for the courses.

5分钟AI长视频不翻车！国产开源框架杀到全球第一梯队

QbitAI

The long audio and video generation framework JoyAI-Echo open-sourced by JD.com has entered the first echelon of global long video generation. The framework can stably generate several minutes of long videos, solving the industry pain points of unstable cross-shot character images and timbre. It is equipped with a conversational editing Agent function, with relevant indicators leading the industry, and user preference reaching 59.4% ~ 81.7%, which will promote the commercial implementation of AI long videos.

马斯克是SpaceX面子，她才是里子

QbitAI

SpaceX will launch its IPO on June 12, with a valuation of $1.77 trillion. After listing, Elon Musk will become the world’s first trillionaire, and the net worth of President and COO Gwynne Shotwell, who holds 0.11% of the shares, will also reach nearly $2 billion. As an early core veteran, Gwynne is the behind-the-scenes pillar of SpaceX: when Falcon 1 failed consecutive launches in the early days, she secured a $1.6 billion life-saving contract from NASA. Later, she also promoted Falcon to monopolize commercial launches and operated the Starlink project. Musk admitted that without her, SpaceX would have gone bankrupt long ago.