Sometimes your code fails, you change absolutely nothing, and on the next run everything works again. If that sounds familiar, this week’s video is for you.

​

These kinds of failures usually have nothing to do with bad code. They come from the outside world: APIs that time out, networks that briefly misbehave, or LLMs that occasionally return something almost structured, but not quite. It’s frustrating, because the failure feels random, and those types of failures are really annoying to debug!

​

In the video, I break down the Retry Pattern step by step. You’ll see how to add simple retries, improve them with exponential backoff, turn the logic into a clean decorator, and finally introduce a fallback strategy for cases where retrying the same thing doesn’t help anymore.

​

I hope this gives you a few ideas to make your own projects more robust :).

​

Enjoy the video, and have a great weekend!

​

Cheers,

Arjan

NVIDIA Nemotron 3: Open Models Built for Real-World AI Systems

(sponsored)

​

If you’re building AI systems beyond simple demos (think agents, multi-step reasoning, or production workloads), NVIDIA just released something worth paying attention to: NVIDIA Nemotron 3.

​

Nemotron 3 is a new family of open models designed specifically for efficient, accurate, and scalable agentic AI. What I like about it is that NVIDIA isn’t treating this as a black box. The models, data, and training techniques are open, so you can actually understand why they behave the way they do and adapt them to your own use case.

​

A few things I found interesting:

• You get access to model weights, training recipes, and datasets. That makes it a great choice if you care about transparency, fine-tuning, or running models in your own environment.
• Nemotron 3 uses a hybrid mixture-of-experts approach to balance performance and cost. This means better throughput and accuracy without just throwing more compute at the problem.
• Nemotron 3 Nano is available to download on Hugging Face, and you can already try it today via OpenRouter. The model is optimized for long-context reasoning, tool usage, and multi-step decision-making, making it well suited for coding tasks, powering AI assistants, and building AI agents for targeted use cases such as transaction monitoring in banking, fraud detection in cybersecurity, and checkout processing in ecommerce.

​

From smaller, faster models to more powerful variants (with larger versions coming next), Nemotron 3 lets you choose the right trade-off for your application.

​

If you want a high-level overview of what Nemotron is, how the model family is structured, and where it fits in the broader NVIDIA AI ecosystem, click here to learn more:

And if you want to go deeper into the technical details (including some of the training techniques, data curation, and efficiency optimizations) NVIDIA also published a solid engineering deep dive (definitely worth a look!).