AI news with context

Clarity in the AI noise.

The business, policy and practice of artificial intelligence — reported precisely, with a European point of view.

Latest reporting

Updated continuously

Tools & practice

Google Launches Sheets Canvas – Gemini-Powered Mini-Apps from Spreadsheet Data

Google announced Sheets Canvas on August 13, 2026, a Gemini-powered feature that transforms spreadsheets into interactive mini-apps without programming knowledge.

Aug 25, 2026

Small Language Models with Memory Systems: New Approaches for Edge-Computing Agents

AI-generated

Models & research

Small Language Models with Memory Systems: New Approaches for Edge-Computing Agents

Researchers are investigating small language models with edge computing for virtual agents. External memory systems are designed to compensate for the limitations of small models in long interactions.

Aug 25, 2026

Hamilton-Zero: Foundation Model Computes Quantum Ground States for Up to 8100 Qubits

AI-generated

Models & research

Hamilton-Zero: Foundation Model Computes Quantum Ground States for Up to 8100 Qubits

Researchers presented Hamilton-Zero on August 12, 2026, a foundation model with 0.5 billion parameters that computes quantum ground states for up to 8100 qubits – 127 times more than during training.

Aug 25, 2026

LittleLearner: Researchers Train 5B Language Model on Controlled Elementary Curriculum

AI-generated

Models & research

LittleLearner: Researchers Train 5B Language Model on Controlled Elementary Curriculum

A research team released LittleLearner on August 13, 2026 – a 5-billion-parameter language model trained exclusively on 88 billion tokens of curated US elementary school materials to investigate knowledge acquisition under controlled conditions.

Aug 25, 2026

GitHub Copilot Boosts Code Contributions in Open Source by 5.9%—Coordination Time Up 8%

AI-generated

Models & research

GitHub Copilot Boosts Code Contributions in Open Source by 5.9%—Coordination Time Up 8%

A recent arXiv study shows that GitHub Copilot increases code contributions in open source projects by 5.9 percent, while simultaneously raising coordination time by eight percent.

Aug 25, 2026

New Benchmark Shows: Standard Metrics Overestimate Language Models by Up to 13 Percentile Points

AI-generated

Models & research

New Benchmark Shows: Standard Metrics Overestimate Language Models by Up to 13 Percentile Points

A study published on arXiv on August 18, 2026, demonstrates that standard accuracy metrics systematically overestimate the actual performance of language models on exams.

Aug 25, 2026

CangjieBench Reveals LLM Weaknesses with New Programming Languages

AI-generated

Models & research

CangjieBench Reveals LLM Weaknesses with New Programming Languages

A new benchmark examines code generation in Cangjie, a resource-scarce language in the HarmonyOS ecosystem. Direct generation fails in over 80 percent of cases, and most failures involve foreign syntax rejected by the compiler.

Aug 25, 2026

ERSkill: New Framework for Evolvable Memory Retrieval Mechanisms in AI Agents

AI-generated

Models & research

ERSkill: New Framework for Evolvable Memory Retrieval Mechanisms in AI Agents

Researchers present ERSkill, a framework for evolvable memory retrieval mechanisms in LLM agents featuring a dual frontier architecture and oracle-coverage guarantee.

Aug 25, 2026

OpenAI Cuts GPT-5.6 Prices in Kiro by Up to 82 Percent

AI-generated

Models & research

OpenAI Cuts GPT-5.6 Prices in Kiro by Up to 82 Percent

OpenAI optimized the integration of GPT-5.6 into the Kiro development environment on August 24, 2026 and achieved approximately 82 percent cost reduction for successful task completion according to their own tests.

Aug 25, 2026

Reinforcement Learning for Concentrated Liquidity: New Research on DeFi Strategies

AI-generated

Models & research

Reinforcement Learning for Concentrated Liquidity: New Research on DeFi Strategies

A research paper published on August 19, 2026 on arXiv models optimal liquidity provision on decentralized finance platforms as a stochastic impulse control problem and uses Reinforcement Learning to solve high-dimensional optimization tasks.

Aug 24, 2026

Stream4D: 4D-Consistency for Streaming-Autoregressive Diffusion Video Models

AI-generated

Models & research

Stream4D: 4D-Consistency for Streaming-Autoregressive Diffusion Video Models

Stream4D addresses spatiotemporal consistency in streaming-based autoregressive diffusion models for video generation – a problem that previous architectures have not sufficiently solved.

Aug 24, 2026

Multi-Agent Architecture Uses LLM Hallucinations for Hypothesis Generation

AI-generated

Models & research

Multi-Agent Architecture Uses LLM Hallucinations for Hypothesis Generation

Researchers at IES Parquesol developed an architecture that transforms speculative language model outputs into testable scientific hypotheses while strategically leveraging hallucinations as source material.

Aug 24, 2026

New Bayesian Framework Models Alzheimer Severity Continuously Rather Than in Discrete Stages

AI-generated

Models & research

New Bayesian Framework Models Alzheimer Severity Continuously Rather Than in Discrete Stages

An international research team introduced Disease Continuum Positioning on August 19, 2026 – a Bayesian framework that continuously estimates Alzheimer severity from DTI data, overcoming discrete diagnostic stages.

Aug 24, 2026

Neural Networks Learn to Predict Deck Strength in Magic: The Gathering

AI-generated

Models & research

Neural Networks Learn to Predict Deck Strength in Magic: The Gathering

Researchers from Kyoto University have developed a neural model that predicts the strength of Magic: The Gathering decks during the drafting process, showing measurable improvements over baseline methods.

Aug 23, 2026

Three-Dimensional Typology of Agency for Advanced AI Systems in Development

AI-generated

Models & research

Three-Dimensional Typology of Agency for Advanced AI Systems in Development

Researchers develop multidimensional frameworks for systematically categorizing agency in advanced AI systems to make autonomy, decision authority, and accountability measurable.

Aug 23, 2026

Transformer Models for Text Summarization: BART Shows Structural Advantages Over BERT and RoBERTa

AI-generated

Models & research

Transformer Models for Text Summarization: BART Shows Structural Advantages Over BERT and RoBERTa

An arXiv study from June 2026 compares BART, BERT, and RoBERTa for text summarization. BART demonstrates structural advantages through its sequence-to-sequence architecture.

Aug 23, 2026

Research: New Watermarking Method for Autoregressive Image Generation Published

AI-generated

Models & research

Research: New Watermarking Method for Autoregressive Image Generation Published

Researchers developed a token-based watermarking method for autoregressive image generation that improves reverse-cycle consistency through tokenizer finetuning and progressive augmentations.

Aug 23, 2026

New Evaluation Protocol Reveals Blind Spots in Multimodal AI Models

AI-generated

Models & research

New Evaluation Protocol Reveals Blind Spots in Multimodal AI Models

Researchers show that high-performing multimodal AI models systematically fail at reasoning over their own outputs—a new evaluation protocol reveals blind spots.

Aug 23, 2026

Semantic Structured Segmentation Improves Patch-Based Time Series Forecasting

AI-generated

Models & research

Semantic Structured Segmentation Improves Patch-Based Time Series Forecasting

New research approaches to adaptive patching methods address the fragmentation of semantic information in multivariate time series. Multi-scale architectures like M2Patch achieve linear complexity while improving forecast accuracy.

Aug 22, 2026

Embodied AI: How LLMs, Knowledge Bases and Reasoning Enable Physical AI Agents

AI-generated

Models & research

Embodied AI: How LLMs, Knowledge Bases and Reasoning Enable Physical AI Agents

The integration of large language models with physical agents and structured knowledge bases is intended to enable the next generation of embodied AI – with vision-language-action models and agent-based architectures.

Aug 22, 2026

DeepMind Tests AI Agents in EVE Online After 15 Years of Game Research

AI-generated

Models & research

DeepMind Tests AI Agents in EVE Online After 15 Years of Game Research

Google DeepMind expands its game research programme, which has run since 2010, and is using the MMORPG EVE Online from May 2026 as a test environment for AI agents in persistent virtual worlds.

Aug 22, 2026

Multimodal Foundation Models Improve Object Detection in Underwater Robots in Turbid Water

AI-generated

Models & research

Multimodal Foundation Models Improve Object Detection in Underwater Robots in Turbid Water

Researchers combine vision-language models with sonar data to stabilize perception of autonomous underwater vehicles under degraded lighting conditions.

Aug 22, 2026

Study Reveals Declining Diversity of LLM Outputs Over Three Years

AI-generated

Models & research

Study Reveals Declining Diversity of LLM Outputs Over Three Years

A new investigation on arXiv demonstrates that Large Language Models are increasingly producing similar content – a trend with implications for creative collaboration between humans and AI.

Aug 22, 2026

Mistral Introduces Agentic Search – AI Agents Search Enterprise Data Three Times More Accurately

AI-generated

Tools & practice

Mistral Introduces Agentic Search – AI Agents Search Enterprise Data Three Times More Accurately

On August 20, 2026, Mistral unveiled Agentic Search, a system that enables AI models to autonomously search complex enterprise data. For financial reports, correctness rises from 26.7 to 86 percent.

Aug 21, 2026

OpenAI Launches AI Futures: New Initiative Examines Power Concentration Through Transformative AI

AI-generated

Tools & practice

OpenAI Launches AI Futures: New Initiative Examines Power Concentration Through Transformative AI

OpenAI has launched AI Futures, a research program on power concentration risks from transformative AI. In parallel, the ChatGPT Futures Class of 2026 honors students who became the first generation to complete their studies with AI.

Aug 21, 2026

MedUAG: Researchers from China Release Training Dataset with 6 Million Medical Images

AI-generated

Models & research

MedUAG: Researchers from China Release Training Dataset with 6 Million Medical Images

A research team from China released MedUAG on August 19, 2026 – the largest unified training dataset for medical multimodal AI – featuring over 6 million instances across 14 imaging modalities.

Aug 21, 2026

Replit Launches Free Mode with GPT-5.6 Luna

AI-generated

Tools & practice

Replit Launches Free Mode with GPT-5.6 Luna

Replit and OpenAI enable software development without usage limits starting August 18, 2026. The cost-efficient GPT-5.6 Luna model reduces costs by 80 percent.

Aug 20, 2026

Google Search: Five New AI Learning Features for Students

AI-generated

Models & research

Google Search: Five New AI Learning Features for Students

On August 19, 2026, Google unveiled new AI-powered learning tools in Search and Gemini, including interactive visualizations, customizable quizzes, and real-time support via video input.

Aug 20, 2026

OpenAI Introduces Private Safety Processing for Zero Data Retention

AI-generated

Safety & society

OpenAI Introduces Private Safety Processing for Zero Data Retention

OpenAI unveiled on August 19, 2026 a new security system that detects abuse patterns without accessing customer data. The technology extends Zero Data Retention with safeguards across multiple interactions.

Aug 20, 2026

OpenAI Tests Zero-Data-Retention Security System for Frontier Models

AI-generated

Safety & society

OpenAI Tests Zero-Data-Retention Security System for Frontier Models

OpenAI tests Private Safety Processing, a system for detecting abuse in frontier models without data retention, while Anthropic requires 30-day retention.

Aug 20, 2026

OpenAI Launches ChatGPT for Teens with Study Mode and Content Filters

AI-generated

Tools & practice

OpenAI Launches ChatGPT for Teens with Study Mode and Content Filters

OpenAI launched ChatGPT for Teens on August 18, 2026, a version for ages 13 to 17 with safeguards against problematic content and a Study Mode for progressive learning.

Aug 19, 2026

Asana Replaced Test System in Two Weeks—Originally Estimated at Five Years

AI-generated

Models & research

Asana Replaced Test System in Two Weeks—Originally Estimated at Five Years

Asana used OpenAI Codex to migrate a legacy test system in two weeks. The project was originally estimated at five years and cost approximately 12,000 US dollars.

Aug 19, 2026

Democratic Control of AI in National Security: US Lawmakers Demand Transparency

AI-generated

Policy & regulation

Democratic Control of AI in National Security: US Lawmakers Demand Transparency

US representatives demand clarification from OpenAI and Anthropic on cybersecurity incidents. In parallel, federal agencies and states are developing oversight frameworks for AI use in national security.

Aug 19, 2026

OpenAI Slows Frontier Model Training Due to Cybersecurity Concerns

AI-generated

Safety & society

OpenAI Slows Frontier Model Training Due to Cybersecurity Concerns

On August 18, 2026, OpenAI paused development of its most powerful AI models after internal tests of the Astra model showed signs of critical cybersecurity capabilities.

Aug 19, 2026

OpenAI Secures 8 Gigawatt Computing Capacity in Ohio

AI-generated

Models & research

OpenAI Secures 8 Gigawatt Computing Capacity in Ohio

On August 17, 2026, OpenAI announced an agreement for 8 gigawatts of IT capacity at the PORTS-Pike Technology Campus in Ohio. The project on the grounds of a former uranium enrichment facility is being developed in partnership with SB Energy, NVIDIA, and the U.S. Department of Energy.

Aug 18, 2026

OpenAI Warns of Closing Window for AI-Powered Cyber Defense

AI-generated

Models & research

OpenAI Warns of Closing Window for AI-Powered Cyber Defense

OpenAI President Greg Brockman describes in a policy essay a critical window before AI models with cyber capabilities become widely available. Companies should now close security gaps with AI agents.

Aug 18, 2026

Same Cluster, 33 Percentage Points More Utilization: What Changed Was the Order

AI-generated

Models & research

Same Cluster, 33 Percentage Points More Utilization: What Changed Was the Order

GPU clusters currently run at only 30-50% utilization while the industry suffers hardware shortages. A July 30, 2026 Dharma AI analysis shows that orchestration strategies like bin-packing and fractional allocation can double capacity without new hardware.

Aug 18, 2026

Google Partners Gemini and Pixel with Five European Top Football Clubs

AI-generated

Models & research

Google Partners Gemini and Pixel with Five European Top Football Clubs

Google partners with Arsenal, Barcelona, Bayern München, Liverpool and PSG starting August 2026. Gemini is positioned as an AI assistant for match analysis, Pixel provides behind-the-scenes content.

Aug 18, 2026

Synthesizer-Folding Framework Improves Multilingual Multi-Hop Question Answering Through Question Decomposition

AI-generated

Models & research

Synthesizer-Folding Framework Improves Multilingual Multi-Hop Question Answering Through Question Decomposition

A new framework for multilingual multi-hop question answering relies on improved question decomposition and streamlined aggregation. The research fits into current approaches that break complex questions into manageable sub-questions.

Aug 17, 2026

Language-Dependent Safety: Claude Recommends Nuclear Weapon Use Less Often in Japanese

AI-generated

Safety & society

Language-Dependent Safety: Claude Recommends Nuclear Weapon Use Less Often in Japanese

A study shows that Claude Sonnet 4.6 recommends nuclear weapon use up to 76 percentage points less often in Japanese than in English when faced with simulated nuclear scenarios – security assessments in only one language fall short.

Aug 17, 2026

Research Team Uses AI System to Narrow Grothendieck Constant

AI-generated

Models & research

Research Team Uses AI System to Narrow Grothendieck Constant

A team led by Rahul Saha published a case study on 11 August 2026 on human-AI collaboration that improves the bounds of the Grothendieck constant KG and determines the first decimal place for the first time.

Aug 16, 2026

Behavioral Reprogramming of Open Language Models: New Study Maps Cognitive Plasticity Limits

AI-generated

Models & research

Behavioral Reprogramming of Open Language Models: New Study Maps Cognitive Plasticity Limits

Researchers published a systematic analysis of behavioral reshaping of open language models on August 13, 2026. The study shows that with 41.9 million trainable parameters, the behavior of an 8-billion-parameter model can be deliberately shifted.

Aug 16, 2026

Edge-aware GAT improves binding site prediction for RNA-protein interactions

AI-generated

Models & research

Edge-aware GAT improves binding site prediction for RNA-protein interactions

Graph Attention Networks with edge-awareness enable more precise predictions of binding sites between RNA and proteins – an active research field with over 15 documented deep learning tools.

Aug 15, 2026

Study Shows Fragile Multilingual Capabilities in Vision-Language Models

AI-generated

Models & research

Study Shows Fragile Multilingual Capabilities in Vision-Language Models

Recent research reveals weaknesses in vision-language models for multilingual tasks – from security risks to semantic limitations.

Aug 15, 2026

OpenAI Launches Ultrafast Mode for GPT-5.6 Sol with 14x Speed

AI-generated

Tools & practice

OpenAI Launches Ultrafast Mode for GPT-5.6 Sol with 14x Speed

On August 13, 2026, OpenAI introduced Ultrafast Mode for GPT-5.6 Sol, which processes the model up to 14 times faster while delivering 750 tokens per second – initially for selected API customers only.

Aug 14, 2026

OpenAI Expands EU AI Act Compliance in Europe

AI-generated

Policy & regulation

OpenAI Expands EU AI Act Compliance in Europe

OpenAI is working with the EU AI Office on implementing transparency obligations that take effect on 2 August 2026. The company already joined the EU Code of Practice in July 2025.

Aug 13, 2026

Data product

The tracker

Availability and regulation at a glance.

Open tracker

Data may be freely cited with attribution –

Clarity in the AI noise.

Google Launches Sheets Canvas – Gemini-Powered Mini-Apps from Spreadsheet Data

Small Language Models with Memory Systems: New Approaches for Edge-Computing Agents

Hamilton-Zero: Foundation Model Computes Quantum Ground States for Up to 8100 Qubits

LittleLearner: Researchers Train 5B Language Model on Controlled Elementary Curriculum

GitHub Copilot Boosts Code Contributions in Open Source by 5.9%—Coordination Time Up 8%

New Benchmark Shows: Standard Metrics Overestimate Language Models by Up to 13 Percentile Points

CangjieBench Reveals LLM Weaknesses with New Programming Languages

ERSkill: New Framework for Evolvable Memory Retrieval Mechanisms in AI Agents

OpenAI Cuts GPT-5.6 Prices in Kiro by Up to 82 Percent

Reinforcement Learning for Concentrated Liquidity: New Research on DeFi Strategies

Stream4D: 4D-Consistency for Streaming-Autoregressive Diffusion Video Models

Multi-Agent Architecture Uses LLM Hallucinations for Hypothesis Generation

New Bayesian Framework Models Alzheimer Severity Continuously Rather Than in Discrete Stages

Neural Networks Learn to Predict Deck Strength in Magic: The Gathering

Three-Dimensional Typology of Agency for Advanced AI Systems in Development

Transformer Models for Text Summarization: BART Shows Structural Advantages Over BERT and RoBERTa

Research: New Watermarking Method for Autoregressive Image Generation Published

New Evaluation Protocol Reveals Blind Spots in Multimodal AI Models

Semantic Structured Segmentation Improves Patch-Based Time Series Forecasting

Embodied AI: How LLMs, Knowledge Bases and Reasoning Enable Physical AI Agents

DeepMind Tests AI Agents in EVE Online After 15 Years of Game Research

Multimodal Foundation Models Improve Object Detection in Underwater Robots in Turbid Water

Study Reveals Declining Diversity of LLM Outputs Over Three Years

Mistral Introduces Agentic Search – AI Agents Search Enterprise Data Three Times More Accurately

OpenAI Launches AI Futures: New Initiative Examines Power Concentration Through Transformative AI

MedUAG: Researchers from China Release Training Dataset with 6 Million Medical Images

Replit Launches Free Mode with GPT-5.6 Luna

Google Search: Five New AI Learning Features for Students

OpenAI Introduces Private Safety Processing for Zero Data Retention

OpenAI Tests Zero-Data-Retention Security System for Frontier Models

OpenAI Launches ChatGPT for Teens with Study Mode and Content Filters

Asana Replaced Test System in Two Weeks—Originally Estimated at Five Years

Democratic Control of AI in National Security: US Lawmakers Demand Transparency

OpenAI Slows Frontier Model Training Due to Cybersecurity Concerns

OpenAI Secures 8 Gigawatt Computing Capacity in Ohio

OpenAI Warns of Closing Window for AI-Powered Cyber Defense

Same Cluster, 33 Percentage Points More Utilization: What Changed Was the Order

Google Partners Gemini and Pixel with Five European Top Football Clubs

Synthesizer-Folding Framework Improves Multilingual Multi-Hop Question Answering Through Question Decomposition

Language-Dependent Safety: Claude Recommends Nuclear Weapon Use Less Often in Japanese

Research Team Uses AI System to Narrow Grothendieck Constant

Behavioral Reprogramming of Open Language Models: New Study Maps Cognitive Plasticity Limits

Edge-aware GAT improves binding site prediction for RNA-protein interactions

Study Shows Fragile Multilingual Capabilities in Vision-Language Models

OpenAI Launches Ultrafast Mode for GPT-5.6 Sol with 14x Speed

OpenAI Expands EU AI Act Compliance in Europe

The tracker

ChatGPT

Claude

Le Chat