Continuum: Efficient and Robust Multi-Turn LLM Agent Scheduling with KV Cache… reminded me about the importance of KV Caches. I was thinking a lot about them a few months ago about the potential of bridging the inference infrastructure with a query engine that can optimize the plan execution to optimize the use of the KV Cache, then with agents I thought that optimizing the KV Cache won’t be as important but I was very wrong. Anthropic and OpenAI are doing a lot of work to optimize the cache based on the workloads and agents are in many respects a different workload on their own. This paper has a smart approach in optimizing the KV cache based on a specific characteristic of the agentic workload related to the back and forth with tool calling. Context engineering at the systems layer!

I made my first Claude Code plugin today to quickly build an agentic application I wanted to test and I realized that it’s a great platform to quickly test new ideas and agentic architectures, especially for multi-agent type of things. It does have it’s issues and limitations, e.g. you can’t really use a non anthropic model right? That’s exactly the reason I find it a great way to quickly put MVPs together to test ideas!

Just tried Anthropic’s Cowork and my first impression is that it indeed brings the experience of Claude Code on the desktop environment, especially if you combine it with the Claude extension for Chrome. It feels more and more that Anthropic is winning over OpenAI when it comes to building products and not only models. (Make sure you understand the security implications before you try Cowork!)

Keystatic is pretty cool. It turns your github repo into a CMS in a very elegant way. The experience of setting up my own GitHub app for oAuth was miserable though, but the keystatic.cloud service they offer is free and it works. Happy that I found this.

I've always loved building things from zero. The kind where you're not just writing code but you are figuring out everything: why it matters, who it's for, how to bring a team together and how to turn it into a product and a company. Two years ago I started working on something new based on a strong conviction that data, although already critical, was about to become exponentially more important.

[... 551 words]

Snowflake announced it would offer dbt Core as a native Snowflake feature. This prompted dbt Labs to modify dbt Fusion's licensing to gain more leverage over its IP. The partnership between these companies has been symbiotic. Snowflake benefits from tools like dbt because they drive workloads inside Snowflake, while dbt Labs gained customer access and revenue through Snowflake's ecosystem.

[... 210 words]

Snowflake's Cortex AISQL announcement got me excited, and I want to explain why. It represents a paradigm shift in integrating large language models into data systems as structured, composable functions rather than opaque tools. Distilling LLM capabilities into five well-defined operators that address 80% of use cases signals meaningful progress. This approach prioritizes reproducibility and composability over raw prompt flexibility.

[... 188 words]

Robust, maintainable, expressive and composable pipelines are critical for scaling synthetic data generation. This post advocates for abstractions that reduce boilerplate, avoiding ad-hoc scripts, and leveraging dataframe APIs with structured document representations. The concrete example involves fine-tuning a smaller model using synthetic QA pairs generated from SEC corporate reports by a frontier model, maintaining quality while reducing inference costs.

[... 260 words]

In this post, I explore how synthetic data is used to train and fine-tune large language models. I'll focus on Meta's open-source **synthetic-data-kit**, a tool built for exactly this purpose. LLMs owe their success to two factors: human ingenuity and the vast, annotated text of the internet.

[... 1,007 words]

Meta's **synthetic-data-kit** is a toolkit designed to generate high-quality synthetic datasets for fine-tuning Large Language Models. The tool streamlines the process of creating training data through an ETL-like pipeline with four key operations. The toolkit exposes a simple CLI interface with these commands:

[... 310 words]