How good or bad are LLMs at programming tasks?
More topics: NVIDIA has released a Python library for Cuda-X, and AI can recreate images from brain activity
Hi AI Enthusiasts,
Welcome to this week’s Magic AI News, where we present you the most exciting AI and Tech news of the week. We organize the updates for you and share our thoughts with you!
This week’s Magic AI tool is an all-in-one app for newsletters, blogs, and YouTube channels. Imagine enjoying all the content of your favorite creators and major newspapers in a single app with AI support! Stay curious! 😎
Let’s explore this week’s AI news together. 👇🏽
Top AI news of the week
👩🏽💻 How good or bad are LLMs at programming tasks?
Symflower, a provider of automated test generation software, has examined how well LLMs can generate code. They have documented their results in various blog articles. The results provide interesting insights into the performance of state-of-the-art LLMs.
The authors assigned various LLMs to create tests for 23 programming examples in Java and Go. The experiment showed that on average only 57.53 % of the generated source code was compiled, with only ten models reaching more than 80 %. So the programmers had to spend some time fixing the code. A common error was missing or unnecessary imports. The automatic management of imports is a common feature in today’s IDEs, so it’s easy to fix these compile errors.
According to the study, the models can generate Java code better than Go code because of more existing training data for Java. The three models (gpt-4o, deepseek-coder, claude-3-opus) have always generated executable code for Java.
Our thoughts
It is interesting to see how LLMs perform with compiler-based programming languages. The results show that programmers are currently indispensable.
AI will not replace you. People who use AI will. For this reason, every programmer should use AI as a personal assistant. And you can run an AI Coding Buddy for free and locally on your computer. We have written an article about this.
More information
- DeepSeek v2 Coder and Claude 3.5 Sonnet are more cost-effective at code generation than GPT-4o! - Symflower Blog
- eval-dev-quality - Symflower GitHub repo
🐍 NVIDIA has released a Python library for Cuda-X!
NVIDIA has released the open-source library nvmath-python as beta. The Python library for Cuda-X speeds up calculations of matrices and vectors.
Here’s all you need to know:
- Access to the mathematical core operations of Cuda-X, without additional C/C++ libraries
- Very good performance, which should come close to that of the native C libraries
- Ideal for the development of hardware-accelerated applications, libraries, frameworks, or deep-learning compilers
- Powerful set of APIs to perform n-dimensional discrete Fourier Transformations
You should also check out the demo video.
Our thoughts
According to Nvidia, the library works seamlessly with the Python ecosystem, with other GPU packages (CuPy, PyTorch, or RAPIDS), as well as with CPU libraries such as SciPy, scikit-learn, or NumPy. This is the right step and simplifies the implementation of hardware-accelerated applications.
More information
- nvmath-python - NVIDIA Developers
- nvmath-python: NVIDIA Math Libraries for the Python Ecosystem - GitHub repo
🤖 AI can recreate images from brain activity
Researchers at Radboud University have developed an AI system that can reconstruct images of what a person is seeing based on their brain activity. The AI learns which parts of the brain to pay attention to, and so the AI becomes significantly better at reconstruction.
In the following image, you can see an example:
In the top row, you can see the original picture, and in the bottom row, you see the reconstructed image of the AI based on the brain activity of a monkey.
Our thoughts
It is impressive how well the AI can reconstruct the images so far. In the future, the technology will become even more precise, which will lead to many medical applications.
More information
Magic AI tool of the week
Imagine you can enjoy all your favorite newsletters, blogs, and YouTube channels in a single app. That is exactly what the app feeeed does. It supports popular newsletter services like Beehiiv*, Substack, Steady*, or Medium. We’ve been using the app for a few weeks now, and we think the app is great!
Here are the main features:
- Enjoy all your favorite newsletters, blogs, YouTube channels, and major magazines, for example, New York Times or The Guardian in a single app
- Daily Updates generated by AI
- AI Summary of blog posts
- Discover new newsletters on specific topics with an AI chatbot
- No account is required!
👉🏽 Try feeeed today and run your own news feed!
Articles of the week
- Mistral’s Codestral: Create a local AI Coding Assistant for VSCode
- Build a Local Chatbot in Minutes with Chainlit
- Responsible Development of an LLM Application + Best Practices
💡 Do you enjoy our content and want to read super-detailed articles about AI? If so, subscribe to our blog and get our popular data science cheat sheets for FREE.
Thanks for reading, and see you next time.
- Tinz Twins
P.S. Have a nice weekend! 😉😉
Leave a comment