DSB #150

Hi,

It’s been over a year since we last published a written version of DSB. During the break, we filled the gap with our podcast in CZ on data science topics, and recently we thought it might be a good idea to revive the bulletin. If you have tips for interesting articles, videos, or podcasts, be sure to share them with us!

And as always, enjoy your reading.

Science

https://www.anthropic.com/engineering/multi-agent-research-system – Anthropic describes how they put together a multiagent system (MAS) for research and what issues they ran into. Sadly, without clear answers on how to solve them.

https://steve-yegge.medium.com/welcome-to-gas-town-4f25ee16dd04 – An eccentric and captivating read on how to put together autonomous, self-sufficient MAS for development. Coined as Kubernetes for agentic development.

https://huggingface.co/papers/2601.23265 – PaperBanana, nice example of how to use multistep LLM processing for high-quality image generation grounded in a prepared dataset. It confirms a trend of differentiating roles for LLM steps (Retriever, Stylist, Critic, …).

https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents – A well-structured (and well-referenced) post about different categories of agentic evaluations, their comparison, and a solid process for achieving excellent evals.

https://github.com/karpathy/autoresearch – Andrej Karpathy released AutoResearch, an AI agent that acts as an LLM researcher. He lets it experiment with the LLM training pipeline in the background overnight and (hopefully) wake up to better results for your LLM training pipeline. If you have a verifiable problem (e.g., LLM performance on a benchmark), you can easily transfer this approach to other domains.

Pop

https://simonwillison.net/2025/Dec/31/the-year-in-llms/ – 2025, another year with LLMs nicely summed up. This post covers most of the topics, new terms, and hype we encountered throughout the year.

https://cepr.org/voxeu/columns/how-ai-affecting-productivity-and-jobs-europe – A beautiful study on the impact of AI in European companies. One of the highlights: “Firms must make complementary investments to unlock AI’s full potential. … An extra percentage point of investment in software and data infrastructure increases AI’s productivity effect by 2.4 percentage points. … an additional percentage point spent on training amplifies AI’s productivity gains by 5.9 percentage points.”

DORA 2025 – This year’s “State of AI-assisted software development” changes its methodology. Here is a decent commentary on the evolution from the old methodology (2024 and earlier) to the 2025 attempt at unifying DORA, SPACE, and DevEx metrics into one framework.

Joke

https://www.reddit.com/r/ProgrammerHumor/comments/1r8tkhq/returnfalseworksinprod/?tl=en#lightbox – Love it, accuracy at its best!