AI News

Follow the latest AI and automation news

1 Apr 202613 Shawwal 1447 AH

ADeLe: Predicting and explaining AI performance across tasks

Microsoft researchers and collaborators have introduced ADeLe, a new methodology for evaluating AI models based on 18 core abilities. This method predicts performance on new tasks with approximately 88% accuracy, including for models like GPT-4o and Llama-3.1. ADeLe builds ability profiles that highlight strengths and weaknesses, linking outcomes to task demands. This approach explains performance variations as task complexity increases, offering deeper insights into model capabilities.

Infrastructure

31 Mar 202612 Shawwal 1447 AH

Efficiency at Scale: NVIDIA, Energy Leaders Accelerating Power‑Flexible AI Factories to Fortify the Grid

At CERAWeek, NVIDIA and Emerald AI announced a new model for AI factories, treating them as flexible grid assets. This approach, based on NVIDIA's Vera Rubin DSX design and Emerald AI's Conductor platform, enhances the connection and efficiency of AI operations with the grid. The collaboration involves companies like AES and NextEra Energy to develop optimized energy generation strategies for these AI factories. This model aims to improve grid reliability and reduce the need for excessive infrastructure to meet rising energy demands.

Models

30 Mar 202611 Shawwal 1447 AH

Latest open artifacts (#20): New orgs! New types of models! With Nemotron Super, Sarvam, Cohere Transcribe, & others

This post highlights the diversity of open models recently released, including NVIDIA-Nemotron-3 and cohere-transcribe-03-2026. These models cover a wide range of applications, from optical character recognition to RAG-based search, showcasing innovation in the field. The emergence of domain-specific, affordable models is increasingly vital, particularly against the backdrop of dominant large models. New releases come from a broader array of open model developers, indicating a promising future for open models.

💡

AsgardBench: A benchmark for visually grounded interactive planning

What impact could AsgardBench have on the development of intelligent robots?

Agents

GroundedPlanBench: Spatially grounded long-horizon task planning for robot manipulation

26 Mar 20267 Shawwal 1447 AH

Agents

26 Mar 20267 Shawwal 1447 AH

GroundedPlanBench: Spatially grounded long-horizon task planning for robot manipulation

VLM-based robot planners face challenges with long, complex tasks due to ambiguities in natural-language plans. GroundedPlanBench has been developed to evaluate whether these models can effectively plan actions and determine their locations in diverse real-world scenarios. The Video-to-Spatially Grounded Planning (V2GP) framework converts robot demonstration videos into spatially grounded training data, enhancing both planning and grounding capabilities.

Models

Gemini 3.1 Flash Live: Making audio AI more natural and reliable

26 Mar 20267 Shawwal 1447 AH

Models

26 Mar 20267 Shawwal 1447 AH

Gemini 3.1 Flash Live: Making audio AI more natural and reliable

The company has announced the launch of a new voice model that offers enhanced precision and lower latency, improving voice interactions. This model features significant performance improvements, with reduced response times allowing for more natural and effective user experiences. This advancement is a crucial step towards enhancing user experience in voice applications, potentially leading to greater adoption of this technology across various sectors.

Agents

Firecrawl + n8n: real-time web data for your AI workflows

26 Mar 20267 Shawwal 1447 AH

Agents

26 Mar 20267 Shawwal 1447 AH

Firecrawl + n8n: real-time web data for your AI workflows

Firecrawl is offering 100,000 credits for users connecting through n8n Cloud, simplifying web data integration into workflows. This partnership allows users to create accounts directly within the platform, enhancing user experience with recent improvements in credential management. Firecrawl addresses traditional data extraction issues by converting web pages into structured, LLM-ready outputs. Reports indicate this technology reduces token waste by 60-90% compared to conventional scraping methods.

Governance

Protecting people from harmful manipulation

25 Mar 20266 Shawwal 1447 AH

Governance

25 Mar 20266 Shawwal 1447 AH

Protecting people from harmful manipulation

Google DeepMind has announced new research focused on the risks of harmful manipulation of AI, particularly in finance and health sectors. This research includes the development of new safety measures aimed at mitigating risks associated with AI usage in these sensitive areas. The announcement comes amid growing concerns about AI's impact on financial and medical decision-making.

Models

Lyria 3 Pro: Create longer tracks in more

25 Mar 20266 Shawwal 1447 AH

Models

25 Mar 20266 Shawwal 1447 AH

Lyria 3 Pro: Create longer tracks in more

Lyria has announced the release of Lyria 3 Pro, which allows for longer track management due to its structural awareness. This update enhances system performance and enables users to create more complex content. Lyria aims to expand its integration with more Google products, reflecting its commitment to providing advanced tools for users. This expansion includes the integration of Lyria into new platforms, making it more accessible.

Models

Will machines ever be intelligent?

23 Mar 20264 Shawwal 1447 AH

Models

23 Mar 20264 Shawwal 1447 AH

Will machines ever be intelligent?

In the first episode of the series "The Shape of Things to Come," Doug Berger from Microsoft engaged with experts like Nicolo Fusi and Subutai Ahmad to examine whether current artificial intelligence systems are truly intelligent. The discussion focused on comparing transformer-based large language models with human brain architecture, exploring differences in efficiency and representation. The conversations also addressed how current systems learn, contrasting this with the continuous learning characteristic of the human brain.