AI News

Follow the latest AI and automation news

29 Aug 202617 Rabiʻ I 1448 AH

Google's WikiSkill gives AI agents a persistent memory of past mistakes to sharpen future performance

Google has announced the launch of WikiSkill, a framework that provides a persistent knowledge base for AI agents. Instead of discarding information after each run, agents document both successes and failures in a wiki-like structure, helping them improve their performance over time. Larger models benefit more from WikiSkill, but smaller models can achieve similar performance to larger ones without it. This innovation marks a significant step towards enhancing the capabilities of AI agents, allowing them to learn from past experiences.

Models

29 Aug 202617 Rabiʻ I 1448 AH

LAION drops massive open video dataset with 10 million hours of footage for AI research

LAION has launched the Big Video Dataset (BVD), which includes 80 million videos and 10 million hours of runtime. Models trained on BVD outperform the previous benchmark, InternVid, by up to 2.1 percentage points. Legally, LAION can reference a 2024 Hamburg court ruling that permits the collection of copyrighted content for non-commercial research. This ruling strengthens researchers' ability to utilize the dataset legally.

Agents

29 Aug 202617 Rabiʻ I 1448 AH

Anthropic wants to do for physical hardware what its Model Context Protocol did for software

Anthropic has announced the launch of its Model Hardware Standard (MHS), which provides a unified interface for AI agents to interact with physical devices like robotic arms. Initial tests revealed that integration time decreased from weeks to hours, showcasing a significant efficiency improvement. However, human oversight remains crucial, as Claude sometimes struggled to understand physical cause and effect. This standard aims to replicate the enhancements achieved by the Model Context Protocol in software, thereby improving the ability of AI agents to operate more seamlessly with physical devices.

💡

OpenAI Halts AI Model Access for Cursor Owned by SpaceX

What are the implications of OpenAI stopping model access for Cursor?

Models

China leverages short-video ecosystem, lower costs to carve out lead in AI video

29 Aug 202617 Rabiʻ I 1448 AH

Models

29 Aug 202617 Rabiʻ I 1448 AH

China leverages short-video ecosystem, lower costs to carve out lead in AI video

Chinese companies are taking the lead in AI video generation, capitalizing on vast short-video ecosystems and more flexible copyright laws. While US giants like OpenAI and Anthropic dominate large language models, Chinese firms are emerging as formidable competitors in the video space. These companies benefit from aggressive pricing and a favorable legal environment, allowing them to achieve significant successes in content generation. Their access to vast data from short-video platforms enhances their competitive capabilities.

Business

Neocloud Lambda secures $1B in debt to buy more chips

28 Aug 202616 Rabiʻ I 1448 AH

Business

28 Aug 202616 Rabiʻ I 1448 AH

Neocloud Lambda secures $1B in debt to buy more chips

Neocloud Lambda has raised $1 billion in private debt to purchase Nvidia AI chips and lease them to Microsoft. This move is part of a series of loans highlighting the high costs associated with the AI boom. The company aims to capitalize on the increasing demand for AI technologies, with Nvidia chips being a crucial component in this sector. This funding underscores the financial challenges companies face amid the rapid growth of artificial intelligence.

Agents

An Anthropic researcher just gave us a peek at self-improving AI

28 Aug 202616 Rabiʻ I 1448 AH

Agents

28 Aug 202616 Rabiʻ I 1448 AH

An Anthropic researcher just gave us a peek at self-improving AI

Automated systems have successfully improved their performance on all ten specified benchmarks for misaligned behaviors without degrading overall performance. This achievement demonstrates the effectiveness of the methods used to address misaligned behaviors, reflecting significant advancements in AI technologies. These results open up new possibilities for intelligent systems, enabling them to achieve better performance in diverse environments, thereby enhancing their applicability across various fields.

Models

Google Deepmind's AI Co-Scientist now plans experiments, runs lab equipment, and writes scientific papers

28 Aug 202616 Rabiʻ I 1448 AH

Models

28 Aug 202616 Rabiʻ I 1448 AH

Google Deepmind's AI Co-Scientist now plans experiments, runs lab equipment, and writes scientific papers

Google DeepMind has announced the expansion of its Co-Scientist system into an integrated research platform utilized across three disciplines. This Gemini-based system has delivered experimentally validated results in areas such as materials synthesis and the autonomous development of medical AI architectures. The new capabilities of Co-Scientist include planning experiments, operating lab equipment, and writing scientific papers, showcasing a significant advancement in the application of AI in scientific research. These developments position the system as a powerful tool for researchers across various fields.

Models

AI Has Human Doctors Asking: What’s Left for Us?

28 Aug 202616 Rabiʻ I 1448 AH

Models

28 Aug 202616 Rabiʻ I 1448 AH

AI Has Human Doctors Asking: What’s Left for Us?

A recent paper claims that artificial intelligence outperforms doctors in diagnosis, sparking mixed reactions within the medical community. The study compares the diagnostic performance of AI systems against that of human doctors, revealing that AI shows higher accuracy in many cases. This finding highlights the potential of AI to enhance medical diagnostics significantly. Such advancements could transform healthcare delivery, prompting physicians to reassess their roles in diagnosis and how they interact with this advanced technology.

Models

AI benchmarks have a trust problem and Google wants to fix it

28 Aug 202616 Rabiʻ I 1448 AH

Models

28 Aug 202616 Rabiʻ I 1448 AH

AI benchmarks have a trust problem and Google wants to fix it

Google DeepMind is conducting a double-blind evaluation of an advanced AI model for the first time. This pilot project, in collaboration with the Singapore AI Safety Institute, employs cryptographic protection through Confidential Space to prevent Google from accessing test questions and evaluators from seeing the model weights. The project utilizes the Gemini Flash Lite model and could establish a new standard for tamper-proof AI benchmarks. This initiative comes at a time when AI benchmarks face trust issues, highlighting the urgent need for reliable evaluations in the field.