{"id":9360,"date":"2025-01-25T20:12:29","date_gmt":"2025-01-26T01:12:29","guid":{"rendered":"https:\/\/www.inthacity.com\/blog\/uncategorized\/new-ai-real-time-self-learning-scary\/"},"modified":"2025-04-13T08:56:09","modified_gmt":"2025-04-13T13:56:09","slug":"new-ai-real-time-self-learning-scary","status":"publish","type":"post","link":"https:\/\/www.inthacity.com\/blog\/tech\/ai\/new-ai-real-time-self-learning-scary\/","title":{"rendered":"This New AI Is Honestly SCARY: Real-Time Self-Learning!"},"content":{"rendered":"
\n

The Rise of Sakana AI: Transformer Squared and Real-Time Adaptation<\/h2>\n

Let\u2019s start with Sakana AI<\/a>, a research lab that\u2019s taking inspiration from nature to build super-flexible AI systems. Their latest creation, the Transformer Squared<\/strong> model, is nothing short of revolutionary. Unlike traditional language models<\/a> that require extensive retraining for new tasks, Transformer Squared can adapt on the fly. Yes, you read that right\u2014no retraining, no extra data, just instant adaptation.<\/p>\n

How does it work? The secret lies in a technique called Singular Value Decomposition (SVD)<\/strong>. By breaking down the model\u2019s weight matrix into smaller components, Sakana AI has created what they call Z vectors<\/strong>. These vectors allow the model to adjust its \u201cskill dials\u201d in real-time, optimizing itself for specific tasks like math, coding, or language understanding. The result? A model that\u2019s not only more efficient but also incredibly versatile.<\/p>\n

In experiments, Transformer Squared outperformed traditional models across a range of tasks, from math and coding to visual question answering. And the best part? The code is open-source<\/a>, so you can try it out for yourself. If you\u2019re an enterprise user, this means faster development, lower costs, and a more practical approach to handling specialized data. It\u2019s a win-win for everyone.<\/p>\n

Why This Matters: The Bigger Picture<\/h3>\n

Over the past year, the AI community has been obsessed with improving how language models perform at inference time. Whether it\u2019s extending context windows or reducing the need for extra data, the goal is to make these models more efficient and user-friendly. Transformer Squared fits perfectly into this narrative, offering a seamless way to adapt existing models to new challenges. It\u2019s not just an improvement\u2014it\u2019s a paradigm shift.<\/p>\n

DeepSeek R1: Open-Source AI That Rivals OpenAI<\/h2>\n

Now, let\u2019s talk about DeepSeek<\/a>, a Chinese AI startup that\u2019s making waves with its latest model, DeepSeek R1<\/strong>. This open-source powerhouse is not just competing with OpenAI\u2019s flagship models\u2014it\u2019s beating them, and at a fraction of the cost. How? By focusing on advanced reasoning and robust Chain of Thought processes.<\/p>\n

DeepSeek R1 is based on a Mixture of Experts (MoE)<\/strong> model called DeepSeek V3. The team has made the entire thing open-source under an MIT license, allowing developers to supercharge their own models. In benchmarks, DeepSeek R1 has achieved near-parity with OpenAI\u2019s GPT-4, all while running at 90-95% lower costs. That\u2019s a jaw-dropping reduction that could be a game-changer for startups and enterprises alike.<\/p>\n

Performance Metrics That Speak for Themselves<\/h3>\n

Let\u2019s break it down:<\/p>\n