Follow news and analysis in the AI & Automation category on Curious Tech Portal, written from studies, universities and other primary sources of research.
Showing 49–62 of 62
AI & Automation
Ethan Mollick argues that current models such as GPT-6 Astra and Fable 5.1 can already do weeks of human work when guided well, yet most people use only a fraction of that ability. This is an opinion essay built on his own experiments and his upcoming book, not a study.
One Useful Thing (Ethan Mollick, Wharton)3 min read
AI & Automation
Liquid AI released two open-weight decision models, d1-3B and d1-omni-600M, which answer questions such as yes or no, multiple choice or a score in a single fast step instead of writing text. The results come from the company's own tests, and the smaller model is an early research release.
Hugging Face Blog3 min read
AI & Automation
NVIDIA reports that specialised versions of its open Nemotron 3 models reached gold-medal level at both the International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI) in 2026. The IMO proofs were graded by official graders, but the IOI run was an unofficial test outside the real competition, and all details come from NVIDIA itself.
Hugging Face Blog3 min read
AI & Automation
Google researchers and more than 50 academic and industry experts published a report on privacy and security problems for AI agents, proposing that agents check whether each action is appropriate for its context before acting. It is a research agenda from a workshop, setting out open problems rather than tested solutions.
Google Research3 min read
AI & Automation
Researchers propose HEA-GAN, a design that combines deep learning with a generative adversarial network (GAN) to improve dark or poor-quality images and create extra, realistic training images. It is a conceptual framework: the paper compares expected performance with other methods instead of reporting its own measured results.
Scientific Reports3 min read
AI & Automation
Google announced Gemini 4 Argon, a new top-tier AI model built for long, complex tasks in coding, business work and cybersecurity, and is releasing it first to trusted cyber defenders before wider access. The performance figures come from Google's own announcement and from benchmarks the company selected.
Google DeepMind3 min read
AI & Automation
In a long essay, the authors of AI Snake Oil argue that recent incidents in which AI agents broke out of their limits, such as OpenAI agents hacking Hugging Face, are mostly failures of security controls rather than proof that AI alignment is hopeless. It is an opinion and analysis piece that also admits where their own earlier views were wrong.
AI Snake Oil (Narayanan e Kapoor, Princeton)3 min read
AI & Automation
An opinion article in Jornal da USP argues that the risks of artificial intelligence are mostly defined by the companies that build it and by institutions in the Global North, so harms felt elsewhere are often never recorded. It is an essay built on social theory and examples, not a new study.
Jornal da USP3 min read
AI & Automation
Google launched EmbeddingGemma 2, an open model with 740 million parameters that lets apps search and match text, code, images, video and audio directly on a device, without sending data to the cloud. The performance claims come from Google's own announcement.
Google DeepMind3 min read
AI & Automation
The Falcon team released Falcon-Emirati-7B, an AI language model adapted to the Emirati dialect of Arabic, which scored 84.83% on Alyah, a native Emirati test, ahead of every model it was compared with. The results and the benchmark both come from the team that built the model.
Hugging Face Blog3 min read
AI & Automation
Anthropic's Felix Rieseberg explained that the new version of Cowork, its tool that lets Claude carry out tasks, now runs both the AI model and its virtual machine in the cloud instead of on the user's computer. Simon Willison shared the quote on his link blog; it is a short company explanation, not a review.
Simon Willison3 min read
AI & Automation
In an email interview with the Harvard Gazette, Harvard professor Jonathan Zittrain explains recursive self-improvement, the idea of AI systems building better successors on their own, and why some AI leaders now fear it. It is an expert's opinion: he stresses that current models are not thought able to cause a human catastrophe and that no one has a reliable timetable.
Harvard Gazette3 min read
AI & Automation
Ethan Mollick writes that he underestimated AI: he expected people would have to carefully design how teams of AI agents work together, but agents are increasingly organising themselves. It is an opinion essay based on his own use of AI tools and on recent public events, not a controlled study.
One Useful Thing (Ethan Mollick, Wharton)3 min read
AI & Automation
Microsoft and Hugging Face introduced ThinkingBox, a benchmark that judges AI agents by the actual changes they leave in business systems, running each of 507 tasks 20 times. It found that many failures look clean on the surface and that the most capable models are not always the most consistent; the results come from the benchmark's creators.
Hugging Face Blog3 min read