AI Technology & Research

Articles in the AI Technology & Research category. Practical tips, news and guides for working with AI.

• Claude Science: Anthropic Introduces a Dedicated AI Workbench for Scientists and Researchers — Anthropic has introduced Claude Science, a specialized workspace designed specifically for scientific research. The new platform consolidates fragmented databases, analytical tools, and computational resource management into a single application. The goal is to streamline complex processes, from data analysis to preparing materials for publication. How does this new tool, currently in beta, work in practice, and who can access it?
• Clarivate AI50 reveals who is really driving AI innovation in the world — Analytics firm Clarivate has published its new AI50 ranking, which uses patent data to identify 52 organizations with the strongest AI inventions in the world. Unlike conventional rankings, it is not concerned with revenue or company size, but with measurable quality and global innovation intent. The result? Four distinct models of AI leadership and a surprising finding about where real value concentrates.
• Google's Gemma 4 runs on your phone and beats models twenty times its size — Google DeepMind has unveiled Gemma 4, its most capable family of open AI models to date. Four variants cover everything from smartphones to powerful servers, work with text, images, video, and audio, and are available under the fully open Apache 2.0 license. The largest model ranks among the top three open models in the world on benchmarks.
• Why Choosing the Right AI Model Now Matters More Than Knowing How to Prompt — The era of needing to be an expert at prompting is slowly coming to an end. The best AI models today guide you to the result themselves. But beware — choosing the right model is more important than ever. A simple car wash test showed that even a trivial logical task is reliably solved by only 5 out of 53 tested models. Which one should you choose and why does it matter?
• LLM, RAG, AI Agent, and Agentic AI — Four Layers of Artificial Intelligence Worth Understanding — Language model, RAG, AI agent, agentic system — it sounds like four different things, but in reality they are layers of one ecosystem. In this article, we'll explain what each layer does, where one ends and another begins, and most importantly, when you need each one. Clearly and without clichés.
• Searching for Miracles in Algorithms — Is Artificial Intelligence Becoming a New Form of Faith? — You've certainly experienced it. You write a complex prompt, press enter, and in that brief second of silence you catch yourself inwardly hoping it will "land right." It's a strange moment of tension that closely resembles a silent prayer. Is artificial intelligence becoming an entity we look up to with sacred reverence, or do we simply not understand what's happening under the hood? Let's take a look at why technology awakens these feelings in us — and how to replace them with certainty.
• A New Model from France Shows Great Potential Even for Demanding Programmers — For a long time, it looked like Europe was only gasping for breath in the AI race, watching the backs of American and Chinese runners. That may be about to change. A French company is coming with a new solution that isn't just trying to catch up with the competition through brute force, but bets on cleverness and openness. If you work with code or are interested in whether Europe has a chance at technological sovereignty, this new development shouldn't pass you by.
• The Largest Neural Dataset in History Is Being Built in an Ordinary San Francisco Basement — Imagine writing an email or controlling a computer just by thinking about it. That sounds like distant sci-fi, but one San Francisco startup is working intensively on it in rather unconventional conditions. In a basement laboratory, they've collected thousands of hours of data directly from human brains to teach AI to convert neural activity into text. How did this unique experiment unfold, and why did participants have to wear two-kilogram helmets on their heads? Let's take a look under the hood of the project that could replace keyboards.
• Just 250 Manipulated Documents Are Enough to Make a Large Language Model Vulnerable — Imagine someone being able to sabotage a chatbot with just a few hundred manipulated texts. Anthropic, in collaboration with British security institutes, has found that as few as 250 malicious documents are sufficient to introduce a backdoor into a large language model. The size of the model or the volume of training data makes no difference. What does this mean for AI security?
• Samsung Proved That Successful AI Doesn't Need Billions of Parameters — What if it turned out that size isn't what matters in artificial intelligence? Samsung has introduced the TRM model, which has only 7 million parameters, yet in complex reasoning tasks it has literally outperformed models that are ten thousand times larger. This breakthrough proves that efficient architecture can matter more than raw computing power. What approach did Samsung use, and what does it mean for the future of AI?
• Why can longer thinking time worsen the output from artificial intelligence? — Researchers from Anthropic have uncovered a strange paradox in AI behavior: the longer AI models think, the worse results they often produce. This discovery challenges the prevailing belief that more computing power and time automatically lead to better understanding and decision-making. This article summarizes the essence of this phenomenon and its implications for the future development of AI.
• What Is the Difference Between Quick Search and Deep Research — Distinguishing between casually chatting with AI and using its capabilities for deeper, systematic research is essential. This article explains why regular AI queries may not provide current and verified information, how Deep Research works in real time, and what options GPT Search and advanced reasoning offer. You'll learn when to use which approach so AI can truly help you.
• New AI Architecture Enables 100x Faster Reasoning Than Large Models with Minimal Training Data — A new AI model from Singapore-based startup Sapient Intelligence introduces a breakthrough architecture capable of solving complex tasks up to 100 times faster than current large language models (LLMs). The hierarchical approach simulates human thinking by combining slow strategic planning with rapid detail processing, while requiring only a minimal amount of training data. This model opens new possibilities for businesses with limited resources and data.
• What If We Have Been Training AI Wrong All Along? — How do we measure and develop cognitive abilities — and how does this affect AI training? A new study brings a surprising finding: important abilities we consider universal may be strongly shaped by the school environment. This has major implications for how we should approach the development of AI systems.
• AI and Performance Paradoxes: Sometimes Less Is More — Do you think the more time you give artificial intelligence to think, the better answer you get? New Anthropic research shows the exact opposite. Find out why longer AI reasoning can mean lower quality results and how to make the most of this knowledge for your company.
• MemOS: The Innovation That Gives AI the Ability to Remember Like a Human — Imagine an AI that remembers what you told it – even when you return a week or a month later. The new MemOS system promises AI with memory similar to our own, maintaining details, preferences, and context over the long term. What does this breakthrough mean for the future of smart assistants? Find out below.
• AI Memory Under the Microscope: A New Look at How It Works — Memory in artificial intelligence is often wrapped in myths and misunderstandings. A new scientific framework brings a clearer division of memory types and defines six fundamental operations that determine how AI stores, updates, and forgets information. What does this mean for the future of AI?
• AI Basics for Seniors: First Steps, Tips and Practical Uses — Artificial intelligence has long ceased to be the domain of only the young and tech-savvy. Seniors too can easily enter its world and use it to simplify everyday life. How to get started, where to begin, and what to watch out for? Read this practical guide that will walk you through the basic concepts and first steps.
• Analysis Suggests AI Reasoning Progress May Slow Down — AI models focused on logical reasoning have been achieving stunning results in recent months, particularly in mathematics and programming. However, a new analysis warns that these rapid advances could slow down significantly very soon. What is behind this outlook and what consequences might it have?
• Continuous Thinking in AI: Sakana Introduces a Model Inspired by the Human Brain — Sakana AI comes with a revolutionary Continuous Thought Machines (CTM) architecture that pushes the boundaries of artificial intelligence closer to human thinking. How does the model that mimics the timing and synchronization of neurons in the brain work, and why could it represent a breakthrough in AI interpretability and adaptability? Read on.
• How Latency Affects the Usability of AI in the Real World — Latency is an often overlooked but crucial parameter that determines how quickly and efficiently artificial intelligence will respond to your requests. In this article you'll learn why low latency is key in various areas, when it matters most, and how it can be optimized.
• ZeroSearch: AI That Learns to Search Without Google and Saves Money — Alibaba comes with the revolutionary ZeroSearch method, which enables artificial intelligence to learn how to search for information without dependence on real search engines. This innovation not only increases control over AI training but mainly reduces costs by up to 88%. How does it work and what does it mean for the future of AI?
• Why NLWeb Is the Key to the Future of AI on the Web — Microsoft introduces NLWeb, an open protocol that enables ordinary websites to become interactive AI applications with a conversational interface. What does this mean for businesses, why should they pay attention, and what is the future of the agentic web? Find the answers in our article.
• Why Artificial Intelligence Still Doesn't Understand the Word "No" in Visual-Language Tasks — Artificial intelligence systems that combine images and language can now recognize objects and generate descriptions. But add a single word — "no" — and their performance drops to chance. Why do VLM models struggle so much with negation, how does it affect practice, and what can be done about it? The new NegBench benchmark has the answers.
• With Google's AI Agents, You No Longer Have to Search — Summarized Web Information Comes to You
• The AI World Has a New Force: Contextual AI — The emergence of Contextual AI represents a significant milestone in the development of artificial intelligence, with the potential to fundamentally change the way we work with information and data. This new model, focused on accuracy and reliability, opens new possibilities for companies, researchers, and individuals in areas where absolute information accuracy is critical.
• ESM3: The Artificial Intelligence That Designed a New Building Block of Life — American startup EvolutionaryScale has created the artificial intelligence ESM3. The generative AI model ESM3, drawing on an extensive database of already-known proteins, designed a new protein sequence and structure now known as the green fluorescent protein esmGFP. It differed from the closest known proteins as if it had undergone 500 million years of evolution. How was this achieved and what will it be used for? Read on in the article below.