Recent Changes

Here is a list of recent updates the site. You can also get this information as an RSS feed and I announce new articles on Fediverse (Mastodon), Bluesky, LinkedIn, and X (Twitter) .

I use this page to list both new articles and additions to existing articles. Since I often publish articles in installments, many entries on this page will be new installments to recently published articles, such announcements are indented and don't show up in the recent changes sections of my home page.

Building Custom Tooling with LLMs

Wed 14 May 2025 10:10 EDT

Tools that treat diagrams as code, such as PlantUML, are invaluable for communicating complex system behavior. But Unmesh Joshi often wished for an extension to walk through these diagrams step-by-step. Yet, extending established tools like PlantUML often involves significant initial setup - parsing hooks, build scripts, viewer code, packaging - enough “plumbing” to deter rapid prototyping.

He narrates now he used an LLM to build a small extension adding step-wise playback to PlantUML sequence diagrams. This illustrates how syntax design, parsing, SVG generation, build automation, and an HTML viewer were iteratively developed through a conversation with an LLM - turning tedious tasks into manageable steps.

Coding Assistants Threaten the Software Supply Chain

Tue 13 May 2025 09:11 EDT

We have long recognized that developer environments represent a weak point in the software supply chain. Jim Gumbley and Lilly Ryan explore how agentic coding assistants expand the attack surface of software supply chains and what measures we need to do to safeguard security when using them.

Function calling using LLMs

Tue 06 May 2025 09:02 EDT

While LLMs excel at generating cogent text based on their training data, they may also need to interact with external systems. Kiran Prakash describes how we get them to construct external function calls to do this. The LLM does not execute these calls directly, instead it creates a data structure that describes the call, passing that to a separate program for execution and further processing. The LLM's prompt includes details about possible function calls and when they should be used.

Building TMT Mirror Visualization with LLM

Wed 30 Apr 2025 09:54 EDT

Creating a user interface that visualizes a real-world structure — like the Thirty Meter Telescope's mirror — might seem like a task that demands deep knowledge of geometry, D3.js, and SVG graphics. But with a Large Language Model (LLM) like Claude or ChatGPT, you don't need to know everything upfront. Unmesh Joshi describes how working with an LLM gave him a working prototype even when he's unfamiliar with the underlying tech.

Additional explanatory material for the Deepseek Overview

Mon 21 Apr 2025 09:07 EDT

A couple of months ago, my colleague Shayan Mohanty published a technical overview of the series of papers describing the deepseek AI models. He's now gone through that article, adding more explanations to make it more digestible for those of us who don't have a background in building these kinds of models.

Guiding an LLM for Robust Java ByteBuffer Code

Thu 17 Apr 2025 09:39 EDT

AI editors like Cursor can generate code with remarkable speed using LLMs, handling boilerplate and providing functional snippets for various tasks. However, when building robust systems, functional correctness is only the starting point. Code must also be safe, predictable, maintainable, and free from subtle side effects. Unmesh Joshi demonstrates, through a dialogue between a developer and an LLM, how expert guidance is crucial to transform an initial, potentially unsafe code snippet into a robust, system-ready component.

Updating yesterday's post on social media engagement

Fri 04 Apr 2025 14:32 EDT

Some people asked about how many people clicked through the links on these social media posts. I've added some more to the article, partly to explain why I don't have that information, and partly to show overall source data for traffic to the site.

Social Media Engagement in Early 2025

Thu 03 Apr 2025 09:41 EDT

A few years ago, whenever I published a new article here, I would just announce it on Twitter, but since the Muskover its importance has declined, and now I post updates to several services. To compare engagement on these services, I've looked at reposts, likes, and replies to two dozen of my recent posts.

I've been kidnapped by Robert Caro

Tue 01 Apr 2025 17:22 EDT

I've always enjoyed reading, and for most of my life I've particularly enjoyed reading history. I've head many great things about Robert Caro's books, but was deterred by their size. Finally, with his first book newly available as an ebook, I decided to dip my toes in. The books are too good for me to escape.

The role of developer skills in agentic coding

Tue 25 Mar 2025 10:39 EDT

As agentic coding assistants get more capable, Birgitta Böckeler is trying them to change existing codebases. This has led to some impressive collaboration sessions, but she's needed to intervene, correct and steer all the time. In her latest post, she describes examples of these interventions, giving ideas of the types of skills we currently need to correct the tools' missteps.

Commenting on removing “X” on US passports

Wed 26 Feb 2025 11:49 EST

The new US administration has decided to eliminate the “X” option for gender/sex on passports. I have several non-binary friends, and I don’t see why they should have to select an option that makes no sense for them. I also don’t see how an “X” option on identity documents causes a material problem for anyone else. There’s not much I can do to object to this change, but one little thing is to add a comment on the Federal Register for such rule changes. There are three such pages:

DS-11 new passport application: (deadline 3/17/25) https://www.federalregister.gov/documents/2025/02/14/2025-02648/30-day-notice-of-proposed-information-collection-application-for-a-us-passport

DS-82 passport renewal: (deadline 3/20/25) https://www.federalregister.gov/documents/2025/02/18/2025-02697/30-day-notice-of-proposed-information-collection-us-passport-renewal-application-for-eligible

DS-5504 name change or data corrections: (deadline 3/20/25) https://www.federalregister.gov/documents/2025/02/18/2025-02696/30-day-notice-of-proposed-information-collection-application-for-a-us-passport-for-eligible

It didn’t take me more than a few minutes to make a brief comment (essentially the first three sentences of this post) on all three. Pretty inconsequential compared to the problems facing my non-binary friends.

GenAI Patterns: Fine Tuning

Tue 25 Feb 2025 07:20 EST

While RAG is the most common way to focus a foundation model on material outside its training set, Bharani Subramaniam and I now finish our current batch of patterns by examining our experience with Fine Tuning where we learned to prioritize curating high-quality data.

GenAI Patterns: Guardrails and RAG overview

Wed 19 Feb 2025 07:02

Gen AI systems are gullible, and can easily be tricked into responding in ways that are contrary to an enterprise's policies or leak confidential information. Bharani Subramaniam and I describe how we can counter this by adding guardrails at the boundaries of the request/response flow. We also conclude our discussion of RAG with an overview of how all the RAG component patterns fit together.

What role does LLM reasoning play for software tasks?

Tue 18 Feb 2025 07:02

Recent LLM models have provided “reasoning” capabilities. Birgitta Böckeler asks what role these play with coding tasks. She doesn't have an answer, but does have questions and thoughts - and has not found such capabilities worthwhile so far.

GenAI Patterns: Reranker

Thu 13 Feb 2025 21:16

LLMs struggle with large amounts of context. Bharani Subramaniam and I explain how to mitigate this common RAG problem with a Reranker which takes the document fragments from the retriever, and ranks them according to their usefulness.

GenAI Patterns: Query Rewriting

Wed 12 Feb 2025 07:58

Users often have difficulty writing the most effective queries. Bharani Subramaniam and I explain Query Rewriting: getting an LLM to formulate alternative queries to send to a RAG's retriever.

The DeepSeek Series: A Technical Overview

Thu 06 Feb 2025 09:17 EST

The appearance of DeepSeek Large-Language Models has caused a lot of discussion and angst since their latest versions appeared at the beginning of 2025. But much of the value of DeepSeek's work comes from the papers they have published over the last year. Shayan Mohanty provides an overview of these papers, highlighting three main arcs in this research: a focus on improving cost and memory efficiency, the use of HPC Co-Design to train large models on limited hardware, and the development of emergent reasoning from large-scale reinforcement learning

GenAI Patterns: RAG Limitations and Hybrid Retriever

Wed 05 Feb 2025 10:03 EST

Today Bharani Subramaniam and I outline four limitations to the simple RAG from yesterday, and the pattern that addresses the first of these: Hybrid Retriever. This tackles the inefficiencies of embeddings-based search by combining it with other search techniques.

Panel at goto Copenhagen: “Where is SW development Going

Tue 04 Feb 2025 10:23 EST

was on a panel at goto Copenhagen last September with Holly Cummings, Trisha Gee, Dave Farley, and Daniel Terhorst-North. We discussed the current state of software development and where it was heading. Given the timing, there was much discussion about the role AI would play in our profession's future.

GenAI Patterns: Retrieval Augmented Generation (RAG)

Tue 04 Feb 2025 09:57 EST

A pre-trained GenAI model lacks recent and specific information about a domain. Bharani Subramaniam and I explain how Retrieval Augmented Generation (RAG) can fill that gap.

Bliki: Forest And Desert

Thu 30 Jan 2025 00:00 EST

The Forest and the Desert is a metaphor for thinking about software development processes, developed by Beth Andres-Beck and hir father Kent Beck. It posits that two communities of software developers have great difficulty communicating to each other because they live in very different contexts, so advice that applies to one sounds like nonsense to the other.

The desert is the common world of software development, where bugs are plentiful, skill isn't cultivated, and communications with users is difficult. The forest is the world of a well-run team that uses something like Extreme Programming, where developers swiftly put changes into production, protected by their tests, code is invested in to keep it healthy, and there is regular contact with The Customer.

Clearly Beth and Kent prefer The Forest (as do I). But the metaphor is more about how description of The Forest and the advice for how to work there often sounds nonsensical to those whose only experience is The Desert. It reminds us that any lessons we draw about software development practice, or architectural patterns, are governed by the context that we experienced them. It is possible to change Desert into Forest, but it's difficult - often requiring people to do things that are both hard and counter-intuitive. (It seems sadly easier for The Forest to submit to desertification.)

In this framing I'm definitely a Forest Dweller, and seek with Thoughtworks to cultivate a healthy forest for us and our clients. I work to explain The Forest to Desert Dwellers, and help my fellow Forest Dwellers to make their forest even more plentiful.

Further Reading

The best short summary from Beth and Kent is on Kent's Substack. For more depth, take a look at their keynote at Øredev.

Acknowledgements

Kent Beck supplied the image, which he may have painstakingly drew pixel by pixel. Or not.

Embeddings in GenAI Products

Wed 29 Jan 2025 10:55 EST

GenAI systems, like many modern AI approaches, have to handle vast quantities of data, and find similarities between elements in an image or chunk of words. Bharani Subramaniam and I describe a key tool to do this - Embeddings - transforming large data blocks into numeric vectors so that embeddings near each other represent related concepts

Emerging Patterns in Building GenAI Products

Tue 28 Jan 2025 07:43 EST

Everyone is fascinated about using generative AI these days, and my colleagues are no exception. Some of them have had the opportunity to put these system into practice, both as proof-of-concept, and more importantly as production system. I've known Bharani Subramaniam for many years as a technology leader in India, he's been assembling the lessons we've learned and I've worked with him to describe them as patterns.

In this first installment, we look the limits of the base case of Direct Prompting, and how we might assess the capability of a system using Evals.

Podcast with Luca Rossi

Fri 24 Jan 2025 13:01 EST

Luca Rossi hosts a podcast (and newsletter) called Refactoring, so it's obvious that we have some interests in common. The tile comes from me leaning heavily on Beth Anders-Beck and Kent Beck's metaphor of The Forest and The Desert. We talk about the impact of AI on software development, the metaphor of technical debt, and the current state of agile software development.

Codemods in other languages

Wed 22 Jan 2025 10:54 EST

Juntao Qiu finishes his description of codemods by looking at some other approaches outside the JavaScript world such as JavaParser and OpenRewrite.

Fixing Common Pitfalls of Codemods

Wed 15 Jan 2025 11:43

So far the codemods that Juntao Qiu has described are fascinating, but rather straightforward. Real codebases offer more challenges. In this installment, he goes into how to tackle more complicated cases by composing codemods.

My favorite musical discoveries of 2024

Thu 09 Jan 2025 09:49 EST

I've got into the habit of starting the New Year by sharing six favorite albums I discovered during the last year. This years set includes Celtic jazz, trip-hop neo-fado, jazz-fusion for the 2020's, playful harmonies, and a vibrant collaboration between Indian classical musicians and a string quartet. (I was also unable to limit myself to six.)

A more complex codemod

Wed 08 Jan 2025 09:44 EST

Juntao Qiu moves onto a more complex example of a codemod, one that extracts a tooltip responsibility from a JSX component. This illustrates how to manipulate the AST in several steps.

Refactoring with Codemods to Automate API Changes

Tue 07 Jan 2025 09:34 EST

As a library developer, you may create a popular utility that hundreds of thousands of developers rely on daily, such as lodash or React. Over time, usage patterns might emerge that go beyond your initial design. When this happens, you may need to extend an API by adding parameters or modifying function signatures to fix edge cases. The challenge lies in rolling out these breaking changes without disrupting your users’ workflows. Juntao Qiu begins an article to explain how we can use codemods to tackle this. Codemods are a tool automating large-scale code transformations, allowing developers to introduce breaking API changes, and refactor legacy codebases.

Design Token-Based UI Architecture

Thu 12 Dec 2024 10:36

Design tokens are fundamental design decisions represented as data. Andreas Kutschmann explains how they work and how to organize them to balance scalability, maintainability and developer experience.