avatar
歸藏(guizang.ai)
90,871 Twitter followers
Follow
关注人工智能、LLM 、 AI 图像视频和设计（Interested in AI, LLM, Stable Diffusion, and design） AIGC 周刊主理人｜公众号：歸藏的AI工具箱
Posts
avatar
歸藏(guizang.ai)
Alibaba released the Qwen 3.6 Plus model yesterday. The key improvement is the significant enhancement in Agent and coding capabilities! They've been very productive lately: first 3.5 Omni, then Wanxiang 2.7, and now Qwen 3.6 Plus, with Max seemingly coming soon as well. In evaluations regarding development and Agent performance, the 3.6 Plus model shows a significant improvement over 3.5. Qwen 3.6 Plus also boasts enhanced capabilities in image and document understanding. For example, it performs exceptionally well in the following areas: mathematical image recognition experience, general VQA real-world question answering, and OCR capabilities. These were also strengths of Qwen previously. Even more impressively, it now supports 1 million context values by default. The longest output is close to 991K, while the input is 64K. The performance with 1 million context values is top-notch. The development experience is significantly better than the previous 256. In terms of pricing, Qwen 3.6 Plus costs 2 yuan per million tokens for input and 12 yuan per million tokens for output. In addition, some tools (such as webpage extraction) are currently free for a limited time. It's already available on Bailian. Alibaba's Qoder and Wukong will also be available later; hopefully, they'll be on the code plan soon. If you're unsure about the quality, you can also try it out for free on OpenRouter to see how it performs. twitter.com/op7418/status/2039...
MAX
5.63%
avatar
歸藏(guizang.ai)
Thread
Karpathy shared how he built his local AI knowledge base. The method I used is similar, but there are still many aspects worth learning from, so I recommend checking it out. Both used Obsidian, pure local Markdown, and then connected them using backlinks and indexes. He built a personal Wiki knowledge base driven by a large language model, and then put all the raw materials into a directory called RAW. Then the large language model compiled these raw materials into a Markdown Wiki, achieving the following functions: automatically creating summaries, creating indexes, creating concept entries, adding crosslinks, generating visualizations, etc. I had already done this during content collection, using the Obsidian clipping plugin. During content collection, the AI automatically performs the following processing: tagging, automatic summarization, translation, and summary creation; currently, there are no backlinks yet. After its Wiki is built, you can ask questions on this Wiki to ensure the credibility of the data sources. For example, the large language model will automatically search the index, read relevant documents, and write answers or reports, not just search on web pages. In this way, the information it acquires is primarily relevant to you. Its output is also designed to be more than just a single sentence; it generates new documents, visual web pages, or PowerPoint presentations, which are then fed back into the knowledge base, making it increasingly rich and comprehensive. However, this presents a problem, as the authors of Obsidian have stated, as it can pollute the knowledge base. Therefore, it's best to separate reliable sources from AI-generated content. Another positive aspect is that it allows a large model to perform health checks on the Wiki. This includes identifying contradictions, supplementing missing information, discovering new connections, and improving consistency. Many companies are already doing this, and I've incorporated this approach into CodePilot. Regarding the choice of assistant folder, I generally recommend using the Obsidian folder. If you have your own Obsidian folder, the AI will directly access all your context and knowledge. For example, here I had it search my Obsidian folder and the internet for articles on UI design principles for the AI era—the quality was very high. This way, you directly obtain an AI assistant with complete memory. If you're unsure how to implement this, I recommend trying Codepilot's assistant. Put your Obsidian folder inside and let it organize it for you, while also having it write these principles into Claude.md. twitter.com/op7418/status/2039...
avatar
歸藏(guizang.ai)
Thread
Google released Gemma 4 yesterday, and it's incredibly powerful! It's specifically designed to run on local devices (like phones and computers) and supports agents and tools. Four parameter sizes: E2B: Primarily for mobile phones/IoT/edge devices. E4B: Designed for mobile devices + Jetson/Raspberry Pi. 26B MoE: 3.8B per activation, very small effective parameters, focusing on high TPS and low latency. 31B Dense: Fully dense 31B, primarily for desktop workstations/single-card H100, etc. This time, they prioritized Agency Workflows support: native support for Function Calls, JSON and structured output, and System Instructions. Even stronger, it's a native multimodal model, supporting image and video understanding, speech-to-text, and can function as a local voice assistant. Furthermore, this time it's truly Apache 2.0 open source, allowing commercial use, redistribution, embedding in products, and private deployment without additional terms. twitter.com/op7418/status/2039...
IOT
0%
loading indicator
Loading..