Alex Strick van Linschoten

Something I found interesting about reading this thread is that it made me wonder how many people think of ChatGPT as primarily a tool for writing things for you

I use ChatGPT/Claude/Gemini possibly 100+ times a day and NONE of those uses are to produce writing that I then share with other people

Blog posts and technical reports welcome if you've already written this up!

#DeepSeek #DataLabeling

Particularly valuable would be:

- Production deployment experiences
- Alternatives to system prompts (discouraged with R1)
- Better approaches than few-shot prompting (since this results in lower performance)
- Effective prompt structures and patterns
- R1 vs V3 prompting differences

🔍 Seeking DeepSeek Prompting Insights for ML Labeling

I'm gathering effective prompting strategies for DeepSeek models (R1/V3) specifically for large-scale data labeling operations.

Will maybe write this up more over on my blog as I dive into how all the pieces work. I fiddled a bit with the learning rate + # of epochs but seems still can improve quite a bit.

Also excited to continue on in the book since the text lessons were something I actually didn't ever get too deeply into when I was doing the course the first time round so excited to dive into LSTMs etc and this whole pre-transformer world!

I generated about 8000 samples (across 100 different prompts, with different temperatures etc) to create the dataset, did a bit of cleanup (to remove `<think>` sections etc) and actually it did quite well! Surprised.

training metrics for the classification model (a graph)

data distribution for the data that went into the model

a confusion matrix showing which samples / examples it did well at / not

per-family performance for the classification model

Have been participating in a group that is doing fastAI part 1 for the first time. Took the opportunity in lesson 1 to use the text dataloaders / trainers to do a little miniproject to see how well fastAI does at classifying between the outputs of different LLM families.

Alex Strick van Linschoten - First impressions of the new Gemini Deep Research (with 2.5 Pro) Some initial fast impressions of Google Deepmind's new iteration of Gemini Deep Research that uses their 2.5 Pro model.

Wrote up a few initial thoughts on the new iteration of Google's Deep Research

mlops.systems/posts/2025-0...

Been meaning to use the latest Gemini(2.x series) models for a while now. Perfect timing by @strickvl.bsky.social with this set of practical examples in this standalone site.

The only thing that could have made this even better for me would be direct curl examples, but maybe I'm asking too much. 😅

I actually implemented a version that spins up Pyodide / WASM and that lets you do actually run them in browser (CURL commands turned into Python 'requests') but I was uncertain if it would detract from the experience... Maybe I'll just push it and see...

Building this has been a great learning experience about Gemini's capabilities.

As someone who works on documentation at ZenML, I wanted to create something that complements the official resources with a different, example-first approach.

Contributions welcomed! github.com/strickvl/gem...

What you'll find:

- 📚 Concise, functional examples that focus on the 80/20 of what most developers need
- 🧠 Native support for llms.txt, bringing documentation directly into your IDE
- ⬅️ ➡️ Simple keyboard navigation (←/→) between examples
- 🌼 Clean, minimal design for quick reference

The fun part? I used Gemini (2.5-pro, obv) itself to help create many of the 30+ examples! My CLI tool asks for a feature and theme (yes, often cats), then Gemini generates code that I test and refine.

Little weekend project: 'Gemini By Example' – practical examples for building with Google's AI models 🚀

I built this simple site to complement Google DeepMind's Gemini SDK with focused, practical code examples - inspired by the "Go By Example" approach.

geminibyexample.com

50,000

Quarto

I'm genuinely excited to see how teams will use this in their workflows. If you're building ML pipelines, I'd love to hear your thoughts on how conversational interfaces might fit into your development process.

#MLOps #MachineLearning

The current implementation is intentionally read-only (with limited exceptions) as we navigate security considerations. MCP has real promise, but the ecosystem needs further development before reaching its full potential.

Screen recording showing me using the ZenML MCP server together with Claude Desktop

With all the recent excitement around Model Context Protocol (MCP), we've built something practical: a way to query your infrastructure, analyze pipeline performance, and trigger runs through simple conversation rather than CLI commands.

Chat With Your ML Pipelines: Introducing the ZenML MCP Server - ZenML Blog Discover the new ZenML MCP Server that brings conversational AI to ML pipelines. Learn how this implementation of the Model Context Protocol allows natural language interaction with your infrastructur...

Just published a blog post about our new ZenML MCP Server - a tool that lets you chat with your ML pipelines using natural language!

www.zenml.io/blog/chat-wi...

@thomasruttig.bsky.social did you read it yet? You’re quoted a LOT

So excited to read this (just arrived). Had been waiting a long time.

GenAI Loft Berlin - Day 6: Vertical Velocity – Accelerating GenAI in Specific Industries | AWS Startups GenAI Loft Berlin - Day 6: Vertical Velocity – Accelerating GenAI in Specific Industries

In Berlin for the Amazon Web Services (AWS) GenAI Loft Week they're putting on. I'll be speaking about the lessons learned from the ZenML LLMOps Database for what it means to put LLMs in production. Do lmk if you're in town!

Free registration for the Monday event:

aws.amazon.com/startups/eve...

Yes. classic Claude

GIF of Joe McMillan from Halt and Catch Fire saying "Computers aren't the thing. they're the thing that gets us to the thing."

@katganly.bsky.social

GitHub - strickvl/tinbox: Translator in a box Translator in a box. Contribute to strickvl/tinbox development by creating an account on GitHub.

And links to the tool here: github.com/strickvl/tin...

Alex Strick van Linschoten - Tinbox: an LLM-based document translation tool Explores an open-source tool I built that tackles the challenges of large-scale document translation using LLMs.

Wrote up a blog with more on this context: mlops.systems/posts/2025-0...