Tudomány Hírek: 21 augusztus 2025

November 9, 2025, 5:06 AM EST. This column revisits knowledge elicitation as a practical path to endowing generative AI and LLMs with true domain expertise. By revisiting GOFAI-era methods, it argues that surface-level documents alone often miss tacit knowledge and best practices embedded in experts' heads. The piece outlines how to surface hidden knowledge, codify it, and feed it into modern models-using strategies like retrieval-augmented generation (RAG) and other data-sourcing techniques-without waiting for exhaustive formal datasets. It explores turning a general LLM into an expert in fields such as medicine, law, or CBT in mental health and compares codified knowledge to tacit know-how. The takeaway: combine structured elicitation with current AI tools to transform LLMs into trusted, domain-smart assistants-and support ongoing coverage in AI breakthroughs.

November 9, 2025, 5:04 AM EST. AI's hardware backbone is evolving. Early progress depended on high-performance processors to run platforms like OpenAI's ChatGPT and Google's Gemini. Nvidia NVDA popularized this with the DGX-1, the first true deep-learning supercomputer, powering today's AI workloads. But the market is shifting: data-center operators are increasingly building their own, custom AI processing chips. These application-specific integrated circuits, or ASICs, are designed for the exact workloads AI demands, offering potential gains in cost and efficiency. The era of Nvidia's near-monopoly may be giving way to a broader ecosystem of bespoke processors optimized for AI tasks. If this trajectory continues, the competitive landscape could unlock dramatic investment opportunities across AI infrastructure and chip startups.

November 9, 2025, 4:42 AM EST. Apple's Vision Pro introduces a new category called a spatial computer, blending augmented reality (AR) and virtual reality (VR). The headset uses cameras to map the real world for AR overlays and can switch to immersive VR by dimming the view, controllable with an on-device Digital Crown. Design resembles ski goggles with a laminated front, aluminum frame, and a soft Light Seal. Two micro-OLED displays deliver over 4K per eye (about 23 million pixels total) and an external EyeSight display shows your eyes. The latest model adds an M5 chip, and accessories like the Dual Knit Band with a Fit Dial improve comfort and weight distribution. No physical controllers are used, and Zeiss Optical Inserts attach magnetically for prescription wearers. A notable step for Apple's ecosystem toward spatial computing and immersive experiences.

November 9, 2025, 4:38 AM EST. 2025 is shaping up as a tipping point for GenAI in accounting. The Thomson Reuters Institute report shows 68% of tax and accounting professionals are hopeful about GenAI, with 21% already using it and 53% planning to adopt. Only 25% have no plans, down from 49% in 2024. Firms are using AI to automate routine tasks and expand advisory work. Notably, many rely on open-source tools (about 52%), while only 17% currently use industry-specific GenAI solutions. The Big 4-Deloitte, EY, PwC, and KPMG-lead the charge with substantial AI investments to deliver deeper client insights and transformation.

November 9, 2025, 4:36 AM EST. Google Research on November 7, 2025 unveiled a new machine learning paradigm called Nested Learning to address catastrophic forgetting in AI models. The approach treats AI as a system of nested learning processes that update at different rates, mirroring human memory. As proof-of-concept, the team introduced Hope, a self-modifying architecture that can continually learn and adapt without restarting from scratch. The work highlights the classic stability-plasticity dilemma and the limitations of networks trained with standard backpropagation on distributed representations. By enabling layered updates, this paradigm aims to preserve old knowledge while integrating new information, improving robustness in dynamic environments and moving AI closer to true continual learning.