[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"nav-categories":3,"article-2025-the-year-large-language-models-revealed-their-true-nature":70},{"data":4},[5,37,57,64],{"name":6,"slug":7,"categories":8},"Productivity","productivity",[9,13,17,21,25,29,33],{"id":10,"title":11,"slug":12},17,"Branding","branding",{"id":14,"title":15,"slug":16},19,"Marketing","marketing",{"id":18,"title":19,"slug":20},20,"Work","work",{"id":22,"title":23,"slug":24},34,"Community","community",{"id":26,"title":27,"slug":28},21,"For newbies","for-newbies",{"id":30,"title":31,"slug":32},24,"Investment","investment",{"id":34,"title":35,"slug":36},22,"Finance","finance",{"name":38,"slug":39,"categories":40},"Tech","tech",[41,45,49,53],{"id":42,"title":43,"slug":44},28,"Technology","technology",{"id":46,"title":47,"slug":48},32,"Artificial Intelligence","artificial-intelligence",{"id":50,"title":51,"slug":52},26,"Security and protection","security-and-protection",{"id":54,"title":55,"slug":56},31,"YouTube Blog","youtube-blog",{"name":58,"slug":59,"categories":60},"News","news",[61],{"id":62,"title":58,"slug":63},18,"quasanews",{"name":65,"slug":66,"categories":67},"Business","business",[68],{"id":69,"title":65,"slug":66},16,{"post":71,"published_news":92,"popular_news":151,"categories":215},{"title":72,"description":73,"meta_title":72,"meta_description":74,"meta_keywords":75,"text":76,"slug":77,"created_at":78,"publish_at":78,"formatted_created_at":79,"category_id":46,"links":80,"view_type":83,"video_url":84,"views":85,"likes":86,"lang":87,"comments_count":86,"category":88},"2025: The Year Large Language Models Revealed Their True Nature","As 2025 draws to a close, it's clear that this was a pivotal year for large language models (LLMs). Far beyond incremental metric improvements on benchmarks, the field witnessed fundamental shifts in training paradigms, our understanding of AI \"intelligence,\" and how we build applications on top of these systems.","As Vasipenok observed, the field is wide open: a sea of work and ideas awaits. The next breakthroughs won't just scale models; they'll redefine how we think, create, and collaborate with AI.","Progress will accelerate, but challenges remain - regulatory scrutiny, data limits, and ethical alignment chief among them.","\u003Cp>As 2025 draws to a close, it's clear that this was a pivotal year for large language models (LLMs). Far beyond incremental metric improvements on benchmarks, the field witnessed fundamental shifts in training paradigms, our understanding of AI "intelligence," and how we build applications on top of these systems.\u003C/p>\n\n\u003Cp>Drawing from insights shared by industry leaders like \u003Ca href=\"https://www.linkedin.com/in/slava-vasipenok/\">Slava Vasipenok, CEO of QUASA\u003C/a>, alongside broader developments, here are the standout themes that defined the year.\u003C/p>\n\n\u003Chr />\n\u003Ch4>\u003Cstrong>1. RLVR: The New Pillar of LLM Training\u003C/strong>\u003C/h4>\n\n\u003Cp>\u003Cimg alt=\"\" class=\"image-align-left\" height=\"447\" src=\"https://quasa.io/storage/photos/00/image - 2025-12-22T135353.257.jpg\" width=\"300\" />The traditional LLM training stack - pretraining on vast data, followed by supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF) - got a powerful upgrade in 2025 with \u003Cstrong>Reinforcement Learning from Verifiable Rewards (RLVR)\u003C/strong>.\u003C/p>\n\n\u003Cp>RLVR trains models not on subjective human preferences but on objective, automatically verifiable outcomes in domains like mathematics, coding, and logic puzzles. Models learn to generate reasoning traces, break down problems, test hypotheses, and self-correct - all driven by clear "correct/incorrect" signals.\u003C/p>\n\n\u003Cp>This approach proved remarkably efficient. Labs shifted significant compute from endless pretraining to extended RLVR runs, yielding dramatic gains in reasoning capability per dollar spent. Pioneered in models like \u003Ca href=\"https://quasa.io/media/deepseek-r1-soars-to-top-3-outshining-gemini-an-open-source-triumph-in-coding-math-and-logic\">DeepSeek R1\u003C/a> and scaled in \u003Ca href=\"https://quasa.io/media/openai-s-o3-model-uncovers-critical-vulnerability-in-linux-kernel-module\">OpenAI's o3 series\u003C/a> (released April 2025), RLVR enabled breakthroughs on tough benchmarks like AIME math competitions and SWE-Bench coding tasks.\u003C/p>\n\n\u003Cp>A key side effect: \u003Cstrong>test-time compute\u003C/strong> emerged as a new performance lever. Models like o3 demonstrated that allocating more inference-time "thinking" (longer chain-of-thought reasoning) intuitively boosts accuracy, turning compute into a tunable dial for real-world use.\u003C/p>\n\n\u003Chr />\n\u003Ch4>\u003Cstrong>2. Ghosts, Not Animals: Embracing Jagged Intelligence\u003C/strong>\u003C/h4>\n\n\u003Cp>\u003Cimg alt=\"\" class=\"image-align-right\" height=\"372\" src=\"https://quasa.io/storage/photos/00/image - 2025-12-22T135351.307.jpg\" width=\"250\" />2025 forced the industry to confront a hard truth: LLMs don't exhibit the smooth, general competence we associate with human (or animal) intelligence. Instead, their capabilities are \u003Cstrong>jagged - \u003C/strong>brilliant in narrow, verifiable domains but surprisingly naive elsewhere.\u003C/p>\n\n\u003Cp>Optimized for text imitation and reward maximization on formal tasks, LLMs spike dramatically in areas amenable to RLVR (e.g., advanced math or code) while remaining vulnerable to simple tricks, jailbreaks, or out-of-distribution scenarios.\u003C/p>\n\n\u003Cp>This "ghostly" intelligence - summoned through optimization rather than evolved for survival - explains why benchmarks lost much of their credibility.\u003C/p>\n\n\u003Cp>That shift has made \u003Ca href=\"https://www.deepchecks.com/llm-evaluation/\">LLM Evaluation\u003C/a> less about leaderboard wins and more about stress-testing models for robustness, safety, and real-world generalization.\u003C/p>\n\n\u003Cp>Most are verifiable by design, making them easy to "overfit" via synthetic data or targeted RLVR, without broader generalization.\u003C/p>\n\n\u003Cp> \u003C/p>\n\n\u003Cp>As Vasipenok noted, beating benchmarks no longer signals proximity to AGI; it often just highlights optimized jaggies.\u003C/p>\n\n\u003Chr />\n\u003Ch4>\u003Cstrong>3. Cursor and the Rise of Specialized LLM Layers\u003C/strong>\u003C/h4>\n\n\u003Cp>Tools like Cursor redefined LLM applications in 2025, creating a new category: "\u003Cstrong>Cursor for X\u003C/strong>" platforms that transform general models into domain experts.\u003C/p>\n\n\u003Cp>These aren't mere chat interfaces. They handle advanced context engineering, orchestrate multi-call workflows (as directed acyclic graphs), balance cost/quality tradeoffs, and provide tailored UIs with "autonomy sliders."\u003C/p>\n\n\u003Cp>\u003Ca href=\"https://quasa.io/media/cursor-s-meteoric-rise-how-vibecoding-fueled-a-200m-arr-startup-without-a-dollar-on-marketing\">\u003Cstrong>The vision\u003C/strong>\u003C/a>: Frontier labs build versatile base models ("universal students"), while applications layer on domain data, tools, sensors, and feedback loops to create specialists.\u003C/p>\n\n\u003Cp>Cursor's explosive growth - acquiring code-review startup Graphite and adding visual design tools - underscored this shift toward composable, agentic ecosystems.\u003C/p>\n\n\u003Chr />\n\u003Ch4>\u003Cstrong>4. Local Agents: Claude Code and the Personal AI Companion\u003C/strong>\u003C/h4>\n\n\u003Cp>\u003Ca href=\"https://quasa.io/media/anthropic-launches-claude-ai-agent-for-chrome-browser\">\u003Cstrong>Anthropic's Claude Code marked a breakthrough\u003C/strong>\u003C/a>: the first compelling local AI agent that lives on your machine, accessing files, terminals, and context directly.\u003C/p>\n\n\u003Cp>\u003Cimg alt=\"\" class=\"image-align-left\" height=\"169\" src=\"https://quasa.io/storage/photos/00/image - 2025-12-22T135318.165.jpg\" width=\"300\" />Running in CLI or IDEs (via partnerships like JetBrains), it proved far more practical than cloud-based agent swarms for everyday tasks. Developers gained a constant "companion" for coding, debugging, and research - blurring the line between tool and teammate.\u003C/p>\n\n\u003Cp>This form factor shift - from web chats to ambient, always-on assistants - highlighted how local deployment addresses privacy, latency, and integration needs in a jagged-capability world.\u003C/p>\n\n\u003Chr />\n\u003Ch4>\u003Cstrong>5. Vibe Coding: Programming Without the Code\u003C/strong>\u003C/h4>\n\n\u003Cp>Coined by Andrej Karpathy, \u003Cstrong>vibe coding\u003C/strong> exploded in 2025, democratizing software creation. Users describe intentions in natural language ("vibes"), and LLMs generate full applications - no manual syntax required.\u003C/p>\n\n\u003Cp>Enabled by tools like \u003Ca href=\"https://quasa.io/projects/lovable\">Lovable\u003C/a> and Opal (integrated into Gemini), it let non-coders build apps, while pros iterated faster and bolder. Code became cheap, experimental, and disposable: spin up a program for one-off tasks, then discard it.\u003C/p>\n\n\u003Cp>This paradigm reshaped software economics and professions, making complex development accessible and accelerating innovation.\u003C/p>\n\n\u003Chr />\n\u003Ch4>\u003Cstrong>6. Toward True Multimodal Interfaces: Gemini's Nano Banana\u003C/strong>\u003C/h4>\n\n\u003Cp>\u003Cimg alt=\"\" class=\"image-align-right\" height=\"169\" src=\"https://quasa.io/storage/photos/00/image - 2025-12-22T135316.983.jpg\" width=\"300\" />Chats proved limiting - text is machine-friendly but not human-optimal. 2025 saw early steps toward richer LLM GUIs, with Google's \u003Ca href=\"https://quasa.io/media/mini-guide-unlocking-crisp-production-ready-art-from-nano-banana-pro-with-art-director-level-prompts\">\u003Cstrong>Gemini Nano Banana\u003C/strong>\u003C/a> (image generation/editing models) as a standout.\u003C/p>\n\n\u003Cp>Blending text, images, diagrams, and animations seamlessly, it hinted at interfaces where knowledge flows visually and interactively. Nano Banana Pro excelled at precise visuals with legible text, paving the way for intuitive, multimodal AI experiences.\u003C/p>\n\n\u003Chr />\n\u003Ch4>\u003Cstrong>Looking Ahead\u003C/strong>\u003C/h4>\n\n\u003Cp>2025 revealed LLMs as a distinctly new form of intelligence: sharper in structured reasoning than anticipated, yet fundamentally uneven. We've tapped perhaps 10% of their potential, with vast untapped applications in agents, interfaces, and hybrid human-AI workflows.\u003C/p>\n\n\u003Cp>Progress will accelerate, but challenges remain - regulatory scrutiny, data limits, and ethical alignment chief among them. As \u003Ca href=\"https://quasa.io/media/interview-ceo-and-founder-of-quasa-slava-vasipenok\">Vasipenok\u003C/a> observed, the field is wide open: a sea of work and ideas awaits. The next breakthroughs won't just scale models; they'll redefine how we think, create, and collaborate with AI.\u003C/p>\n\n\u003Cp>\u003Cstrong>\u003Cimg alt=\"\" class=\"image-align-left\" height=\"169\" src=\"https://quasa.io/storage/photos/00/image - 2025-12-22T135302.177.jpg\" width=\"300\" />Also read:\u003C/strong>\u003C/p>\n\n\u003Cul>\n\t\u003Cli>\u003Ca href=\"https://quasa.io/media/how-machine-learning-improves-accuracy-in-legal-and-delivery-management-tasks\">How Machine Learning Improves Accuracy in Legal and Delivery Management Tasks\u003C/a>\u003C/li>\n\t\u003Cli>\u003Ca href=\"https://quasa.io/media/monero-surges-ahead-while-zcash-faces-headwinds-in-the-privacy-coin-arena\">Monero Surges Ahead While Zcash Faces Headwinds in the Privacy Coin Arena\u003C/a>\u003C/li>\n\t\u003Cli>\u003Ca href=\"https://quasa.io/media/the-world-s-50-most-profitable-companies-in-2025-tech-titans-banking-behemoths-and-energy-empires\">The World's 50 Most Profitable Companies in 2025: Tech Titans, Banking Behemoths, and Energy Empires\u003C/a>\u003C/li>\n\u003C/ul>\n\n\u003Cp>Author: \u003Ca href=\"https://www.linkedin.com/in/slava-vasipenok/\">Slava Vasipenok\u003C/a>\u003Cbr />\nFounder and CEO of QUASA (quasa.io) - Daily insights on Web3, AI, Crypto, and Freelance. Stay updated on finance, technology trends, and creator tools - with sources and real value.\u003C/p>\n\n\u003Cp>Innovative entrepreneur with over 20 years of experience in IT, fintech, and blockchain. Specializes in decentralized solutions for freelancing, helping to overcome the barriers of traditional finance, especially in developing regions.\u003C/p>","2025-the-year-large-language-models-revealed-their-true-nature","2025-12-22T12:56:18.000000Z","22.12.2025",{"image":81,"thumb":82},"https://quasa.io/storage/images/news/KuM4738B9tYoH7u6yPWcQENvwPSLraPHG4o3jP2t.jpg","https://api.quasa.io/thumbs/news-thumb/images/news/KuM4738B9tYoH7u6yPWcQENvwPSLraPHG4o3jP2t.jpg","small",null,3040,0,"en",{"id":46,"title":47,"slug":48,"meta_title":47,"meta_description":89,"meta_keywords":89,"deleted_at":84,"created_at":90,"updated_at":91,"lang":87},"Artificial Intelligence, ai, ml, machine learning, chatgpt, future","2024-09-22T08:08:27.000000Z","2024-09-23T12:49:38.000000Z",[93,106,117,129,140],{"title":94,"description":95,"slug":96,"created_at":97,"publish_at":98,"formatted_created_at":99,"category":100,"links":101,"view_type":83,"video_url":84,"views":104,"likes":86,"lang":87,"comments_count":86,"is_pinned":105},"Figure AI Founder Brett Adcock Launches Hark: A New Lab Building True Personal AGI","Brett Adcock, the founder of humanoid robotics company Figure AI, has quietly launched his next ambitious venture: Hark, an AI laboratory dedicated to creating the world’s most advanced personal intelligence.","figure-ai-founder-brett-adcock-launches-hark-a-new-lab-building-true-personal-agi","2026-03-25T20:23:25.000000Z","2026-04-09T09:11:00.000000Z","09.04.2026",{"title":43,"slug":44},{"image":102,"thumb":103},"https://quasa.io/storage/images/news/C1UuYbaKmb90FPNMq9HupZ84jzRUQmO6HBbSPEPh.jpg","https://api.quasa.io/thumbs/news-thumb/images/news/C1UuYbaKmb90FPNMq9HupZ84jzRUQmO6HBbSPEPh.jpg",47,false,{"title":107,"description":108,"slug":109,"created_at":110,"publish_at":111,"formatted_created_at":99,"category":112,"links":113,"view_type":83,"video_url":84,"views":116,"likes":86,"lang":87,"comments_count":86,"is_pinned":105},"Pokémon Go’s 10-Year Legacy: How Millions of Players Accidentally Trained Robots to Deliver Your Food","Ten years after its explosive launch, Pokémon Go continues to deliver surprises — this time, not in the form of rare Pikachu, but in the form of a powerful new navigation system for delivery robots.","pokemon-go-s-10-year-legacy-how-millions-of-players-accidentally-trained-robots-to-deliver-your-food","2026-03-25T20:06:36.000000Z","2026-04-09T06:59:00.000000Z",{"title":43,"slug":44},{"image":114,"thumb":115},"https://quasa.io/storage/images/news/ceiFSOJQ3diMLwMI4POUSJwIjscAuwf5TWE2DXlN.jpg","https://api.quasa.io/thumbs/news-thumb/images/news/ceiFSOJQ3diMLwMI4POUSJwIjscAuwf5TWE2DXlN.jpg",126,{"title":118,"description":119,"slug":120,"created_at":121,"publish_at":121,"formatted_created_at":122,"category":123,"links":124,"view_type":127,"video_url":84,"views":128,"likes":86,"lang":87,"comments_count":86,"is_pinned":105},"Jaxx Liberty for Cross-Device Crypto Management","A practical article about Jaxx Liberty as a cross-platform blockchain wallet with portfolio tools, exchange access, and block explorer features for daily crypto management.","jaxx-liberty-for-cross-device-crypto-management","2026-04-08T12:21:43.000000Z","08.04.2026",{"title":35,"slug":36},{"image":125,"thumb":126},"https://quasa.io/storage/images/news/2BoYmDUS7F9wHE0dwpD2KUpX702LkL1Spp2KC5Vm.jpg","https://api.quasa.io/thumbs/news-thumb/images/news/2BoYmDUS7F9wHE0dwpD2KUpX702LkL1Spp2KC5Vm.jpg","large",803,{"title":130,"description":131,"slug":132,"created_at":133,"publish_at":134,"formatted_created_at":122,"category":135,"links":136,"view_type":83,"video_url":84,"views":139,"likes":86,"lang":87,"comments_count":86,"is_pinned":105},"Nvidia CEO Jensen Huang Proposes Paying Engineers with AI Tokens","Nvidia CEO Jensen Huang has put forward a radical new idea for compensating tech talent: paying engineers part of their salary in AI tokens.","nvidia-ceo-jensen-huang-proposes-paying-engineers-with-ai-tokens","2026-03-25T19:56:16.000000Z","2026-04-08T11:48:00.000000Z",{"title":19,"slug":20},{"image":137,"thumb":138},"https://quasa.io/storage/images/news/KL5Yqt5ZzW3R4cu7XmE15grrRHdKZ1WBoKbC6Zu3.jpg","https://api.quasa.io/thumbs/news-thumb/images/news/KL5Yqt5ZzW3R4cu7XmE15grrRHdKZ1WBoKbC6Zu3.jpg",856,{"title":141,"description":142,"slug":143,"created_at":144,"publish_at":145,"formatted_created_at":122,"category":146,"links":147,"view_type":83,"video_url":84,"views":150,"likes":86,"lang":87,"comments_count":86,"is_pinned":105},"The AI Scientist Hits Nature: Scaling Scientific Discovery Like Code","The dream of a fully autonomous laboratory has moved from the realm of \"experimental demo\" to a validated scientific reality. The AI Scientist, a pioneering system designed to automate the entire lifecycle of research, has officially been published in Nature.","the-ai-scientist-hits-nature-scaling-scientific-discovery-like-code","2026-03-25T19:23:02.000000Z","2026-04-08T09:15:00.000000Z",{"title":47,"slug":48},{"image":148,"thumb":149},"https://quasa.io/storage/images/news/QZJ74ZuxTmCA52qWq4HZkP2WikrOYpg6WBusDdyn.jpg","https://api.quasa.io/thumbs/news-thumb/images/news/QZJ74ZuxTmCA52qWq4HZkP2WikrOYpg6WBusDdyn.jpg",924,[152,165,178,190,203],{"title":153,"description":154,"slug":155,"created_at":156,"publish_at":157,"formatted_created_at":158,"category":159,"links":160,"view_type":83,"video_url":84,"views":163,"likes":164,"lang":87,"comments_count":86,"is_pinned":105},"The Anatomy of an Entrepreneur","Entrepreneur is a French word that means an enterpriser. Enterprisers are people who undertake a business or enterprise with the chance of earning profits or suffering from loss.","the-anatomy-of-an-entrepreneur","2021-08-04T15:18:21.000000Z","2025-12-14T06:09:00.000000Z","14.12.2025",{"title":65,"slug":66},{"image":161,"thumb":162},"https://quasa.io/storage/images/news/mVsXPTMuHZuI7UXCsENgL1Qwp1uSOf7Rz3uVPMfm.webp","https://api.quasa.io/thumbs/news-thumb/images/news/mVsXPTMuHZuI7UXCsENgL1Qwp1uSOf7Rz3uVPMfm.webp",67910,2,{"title":166,"description":167,"slug":168,"created_at":169,"publish_at":170,"formatted_created_at":171,"category":172,"links":173,"view_type":127,"video_url":84,"views":176,"likes":177,"lang":87,"comments_count":86,"is_pinned":105},"Advertising on QUASA","QUASA MEDIA is read by more than 400 thousand people a month. We offer to place your article, add a link or order the writing of an article for publication.","advertising-on-quasa","2022-07-06T07:33:02.000000Z","2025-12-15T17:33:02.000000Z","15.12.2025",{"title":58,"slug":63},{"image":174,"thumb":175},"https://quasa.io/storage/images/news/45SvmdsTQbiyc3nxgbyHY1mpVbisYyub2BCHjqBL.jpg","https://api.quasa.io/thumbs/news-thumb/images/news/45SvmdsTQbiyc3nxgbyHY1mpVbisYyub2BCHjqBL.jpg",67299,4,{"title":179,"description":180,"slug":181,"created_at":182,"publish_at":183,"formatted_created_at":184,"category":185,"links":186,"view_type":83,"video_url":84,"views":189,"likes":177,"lang":87,"comments_count":86,"is_pinned":105},"What is a Startup?","A startup is not a new company, not a tech company, nor a new tech company. You can be a new tech company, if your goal is not to grow high and fast; then, you are not a startup. ","what-is-a-startup","2021-08-04T12:05:17.000000Z","2025-12-17T13:02:00.000000Z","17.12.2025",{"title":65,"slug":66},{"image":187,"thumb":188},"https://quasa.io/storage/images/news/EOsQhSW3VXyG7a6NPdE1oZd00xfJXe3bjY5aJGb7.webp","https://api.quasa.io/thumbs/news-thumb/images/news/EOsQhSW3VXyG7a6NPdE1oZd00xfJXe3bjY5aJGb7.webp",65576,{"title":191,"description":192,"slug":193,"created_at":194,"publish_at":195,"formatted_created_at":196,"category":197,"links":198,"view_type":83,"video_url":84,"views":201,"likes":164,"lang":87,"comments_count":202,"is_pinned":105},"Top 5 Tips to Make More Money as a Content Creator","Content creators are one of the most desired job titles right now. Who wouldn’t want to earn a living online?","top-5-tips-to-make-more-money-as-a-content-creator","2022-01-17T17:31:51.000000Z","2026-01-17T11:30:00.000000Z","17.01.2026",{"title":19,"slug":20},{"image":199,"thumb":200},"https://quasa.io/storage/images/news/gP8kiumBPpJmQv6SMieXiX1tDetx43VwFfO1P4Ca.jpg","https://api.quasa.io/thumbs/news-thumb/images/news/gP8kiumBPpJmQv6SMieXiX1tDetx43VwFfO1P4Ca.jpg",40038,1,{"title":204,"description":205,"slug":206,"created_at":207,"publish_at":208,"formatted_created_at":209,"category":210,"links":211,"view_type":127,"video_url":84,"views":214,"likes":164,"lang":87,"comments_count":86,"is_pinned":105},"8 Logo Design Tips for Small Businesses","Your logo tells the story of your business and the values you stand for.","8-logo-design-tips-for-small-businesses","2021-12-04T21:59:52.000000Z","2025-05-05T03:30:00.000000Z","05.05.2025",{"title":15,"slug":16},{"image":212,"thumb":213},"https://quasa.io/storage/images/news/Wbx2NtS1CnTupgoQbpFMGspJ5jm4uob2hDOq33r0.jpg","https://api.quasa.io/thumbs/news-thumb/images/news/Wbx2NtS1CnTupgoQbpFMGspJ5jm4uob2hDOq33r0.jpg",39443,[216,217,218,219,220,221,222,223,224,225,226,227,228],{"title":23,"slug":24},{"title":47,"slug":48},{"title":55,"slug":56},{"title":43,"slug":44},{"title":51,"slug":52},{"title":31,"slug":32},{"title":35,"slug":36},{"title":27,"slug":28},{"title":19,"slug":20},{"title":15,"slug":16},{"title":58,"slug":63},{"title":11,"slug":12},{"title":65,"slug":66}]