Towards Data Science

"For data engineers, this isn’t just market news. Instead, it’s a daily constraint." Jiayan Yin outlines several ideas for data engineers struggling to navigate memory constraints in their workflows. towardsdatascience.com/when-memory-be…

"Four parts, four kinds of redundant work, one underlying lesson: refusing to recompute beats every clever algorithm." Anubhab Banerjee brings his excellent series on agentic inference to a close. towardsdatascience.com/persistent-lat…

To gain a deeper understanding of the particular challenges of dealing with spatial data, and how its "spatial dependence, panel structure, and coverage shape the credibility of the results," don't miss @erikapatgs's insightful deel dive. towardsdatascience.com/why-powerful-m…

"So, how about we make retrieval iterative? What if the model can search, read, decide whether it has enough evidence, and search again when needed?" Shuai Guo walks us through an agentic-RAG workflow step by step. towardsdatascience.com/agentic-rag-le…

Why do Claude Code sessions gradually (and quietly) "rot" — and what can you do to keep them fresh and well-performing? Jake Minns's new technical deep dive addresses an emerging challenge for AI practitioners. towardsdatascience.com/governed-conte…

For her debut TDS article, Sneha Rani presents an incisive reflection on the differences between building models in an academic settings vs. in a real-world industry context. towardsdatascience.com/building-model…

Looking for a robust, cloud-based approach to building AI agents? @taupirho explains how to go about it using Strands and AgentCore. towardsdatascience.com/build-and-run-…

From adopting a schema instead of branching code to leaning into the power of an expert dictionary over embeddings, Angela Shi offers six counterintuitive insights for RAG question-parsing. towardsdatascience.com/the-untaught-l…

To help us better understand how time-series LLMs work and how well they perform, Sean Moran takes t0-alpha — a decoder-style patch transformer for probabilistic time-series forecasting — for a spin. towardsdatascience.com/time-series-ll…

"A check that rewards a confident wrong answer is doing exactly that — feeding the loop a reward signal that is correlated with fluency rather than truth." @jamarinval highlights the risks of having models judge their own performance, and proposes a more robust approach instead. towardsdatascience.com/design-loops-n…

Don't miss Gadi Singer's new article: a thoughtful deep dive outlining a framework for aligning agentic AI with enterprise intent to ensure consistent scenario‑wide autonomous behavior. towardsdatascience.com/the-three-dime…

When should you use RAG and when is it time to turn to fine-tuning? @m_mouschoutz shares an accessible guide to two powerful approaches and the tradeoffs they present. towardsdatascience.com/rag-vs-fine-tu…

Looking to scale up your Claude Code workflows? @EivindKjos's latest guide offers tips on orchestrating 100 agents — or more! towardsdatascience.com/how-to-orchest…

What if your LLM pipeline is failing because you made it remember too much? @emmimalpa explains how to add a prompt-pruning layer to remove redundant and low-value tokens you accumulate over time. towardsdatascience.com/long-context-i…

Shouldn't we have moved past hallucinating models by now? Omer Rosenbaum explains why even cutting-edge frontier models still hallucinate, and what we can do on our end to reduce that risk. towardsdatascience.com/that-is-embarr…

Learn to spin up 100+ agents, in my latest @TDataScience article: towardsdatascience.com/how-to-orchest…

Read Sam Black's new article to learn why tokenmaxxing is out, and why practitioners should instead focus on reducing costs without sacrificing AI effectiveness. towardsdatascience.com/tokenminning-h…

The next step after local inference is giving your model the ability to act. Shuai Guo explains how tool use transforms a local LLM into a practical research assistant. towardsdatascience.com/from-local-llm…

AI agents become useful when they can do more than generate text. @m_mouschoutzi explains how tool calling enables LLMs to decide when and how to interact with external systems. towardsdatascience.com/tool-calling-e…

"Cloud LLMs reason better, but we send our sensitive context outside. Local models keep that context private, but limited by the local compute, they can struggle when tasks get complex." What if you didn't have to choose? Shuai Guo explains in his recent deep dive. towardsdatascience.com/stop-choosing-…