Original): Towards the AI editorial team
Originally published in the direction of artificial intelligence.
Good morning, AI enthusiasts,
This week, the edition returns to what is happening when you go deeper by one layer, regardless of whether it is MCP for the integration of smarter tools or hand -coding backprop to finally understand what your model is doing under the hood.
In AI, we spread the model contextual protocol (MCP), a new standard that can save hours of repetitive work of integration between tools and orgs. Then we go practically: building vision transformers in Pytorch, testing LLM at bytes level without tokenization, comparing neural optimizers and assessment when Open Source alternatives are good enough to use.
Also in the mix: tools for enriching metadata for rust, new necessary collabów and a meme, which can hit a bit too close for anyone who has fallen a gradient this week.
Let's get it.
What is a week AI
This week What is AII will immerse myself in the context protocol of the Anthropic model. How many times have you set up a new AI project and discovered the same github, slack or sql? This annoying repetition of a copy of the code and the lack of standards between organizations and people, is exactly the reason why there is a model contextual or MCP protocol. Read about what it is and why it matters in your everyday life or Watch the video on YouTube.
-Louis-François Bouchard, towards the co-founder of AI and the head of the community
Learn a community section!
Presented post from disagreement
Superuser666_30897 He developed a system of collecting, enriching and analyzing metadata for rust, using observations powered by AI, scraping of websites and analysis of dependencies. It combines internet scraping, AI driving analysis and load tests to ensure comprehensive insight into rust ecosystem packages. Check it out on github and support another member of the community. If you have any questions or suggestions, Reach to him in the thread!
Ai Poll week!
More than half of this community entered AI after chatgpt, and this is not a bad thing. This means a clear generational change: from the first research to the product, from academic documents to API. If you joined in front of chatgpt, how do you think, how do you think about thinking? And if you joined later, what is one thing, what do you think the earlier crowd underestimated? Tell us in the thread!
Opportunities for cooperation
The community of science and together is flooded with cooperation opportunities. If you are excited to immerse yourself in the artificial intelligence, you want a learning partner and even want to find a partner for your passion project, Join the cooperation channel! Keep an eye on this section, we share cool possibilities every week!
1. Skaggsllc He builds Vera AI, the AI system in the field of predictive vehicle maintenance and fleet diagnostics, and is looking for programmers who may be interested in the development of this platform. If it sounds like your niche Connect in the thread!
2. Vergil727 He is looking for someone who will help integrate the advanced planning and planning system (APS) in their ERP/MES environment. You will support data mapping, planning configuration and system integration (SQL, ERP, MES). If this is in line with your skill set, please Reach in the thread!
Meme of the week!
Meme made available by Bigbuxchungus
TAI section
Article
From pixels to forecasts: building a transformer for photos through Vicki Y
Because generative AI is still developing, it is necessary to understand fundamental models such as the Vision transformer. This review describes in detail the process of building a vision transformer (VIT) from zero in Pytorch, explaining the theory of transforming images into patches and processing them with many self -translations. The author implements the model, trains it on the Cifar-10 data set and analyzes the results, reaching an accuracy of 60%. It also includes VIT restrictions and mentions more advanced architecture.
Ours must read articles
1. The week I spent manually coding neural networks to finally understand reverse propagation By Abduldttijo
The author describes his experience in building a neural network from scratch with only NumPy, despite the fact that he had many years of experience in such as Pytorch. This process caused by the vulnerability in understanding of reverse propagation includes challenges regarding manual calculation of the gradient, the implementation of reverse pass and building the optimizer. This exercise gave a deeper, practical understanding of how neural networks operate, which caused better debugging skills and better ability to understand complex architecture.
2. From bytes to ideas: LLM without tokenization By Sq m
The author analyzes the autoregression U-Net Meta (AU-NET), architecture designed to overcome the restrictions of traditional tokenization in language models. Instead of using predefined tokens, AU-NET processes a raw text at bytes level, learning to build understanding from letters to concepts. This method improves the support of typos and new languages. The benchmark performance shows that AU-NET is competitive with standard models, showing special strength in multilingual translation and tasks at the character level. However, the current version is optimized primarily for Latin languages.
3. Understanding the contextual protocol model (MCP): The future of the integration of AI tools By Mahendramedapati
This blog explains the context model of the model (MCP), a standardized method of connecting AI models with external systems such as databases, API interfaces and files. He acts as a universal translator, eliminating the need for complex, non -standard integration for each tool. Using the hotel analogy, the author illustrates how MCP sets data and forms data for AI. The text presents benefits such as reduced costs and better security, and provides guides with examples of code to implement.
4. The best optimization algorithm for your neural network By Riccardo Andreoni
The author presented a guide to algorithms for optimizing neural networks designed to shorten the training time. He browses basic methods, such as descent from the party gradient and mini, before explaining more advanced techniques. It included a momentum that uses past gradients for faster convergence, and RMSSProp, which adapts learning speed for each parameter. Then the discussion went to Adam, an algorithm connecting both approaches, whose excellent performance was shown in practical comparison with multist fashion. The summary also noticed the distribution of learning speed as a complementary technique of improving the training process.
If you want to publish with AI, check our guidelines and register. We will publish your work on our network if it meets our editorial rules and standards.
Published via AI
