Alexander Regueiro

I've published a new blog post introducing two open source tools I've been working on: Minicrest and GWT. Minicrest brings Hamcrest-style composable matchers to Ruby's Minitest. If you value expressive, readable tests that clearly communicate intent, this library helps you build assertions that compose naturally and provide detailed failure messages.                                                                             GWT is a CLI tool that simplifies git worktree management. One motivation for building it was enabling parallel work with coding agents like Claude Code. Each agent gets its own isolated worktree, making it practical to explore multiple solutions to the same problem simultaneously and compare the results.                                               The post covers the philosophy behind both tools and how they fit into my workflow.  https://lnkd.in/dYrKguGu

9

1 Comment

Just spotted this contribution guide in the NanoClaw repo and I suspect we'll see this everywhere soon. Coding agents change the economics of integrations. The traditional playbook is to build a massive integration catalogue. Hundreds of connectors that most customers never use, but you need them for competitive checkboxes. It's bloat by design. Enterprises get the most value from a custom integration, and now it's cheaper than ever to build one. I remember Matthaus Krzykowski from dltHub making this point a few years ago. They were approaching data pipelines the same way: give people the primitives to build the integration they actually need, not a pre-built feature that doesn't quite cut it. We're moving from "does it have the integration?" to "can I make it do what I need?" And that's a better question.

125

15 Comments

Really insightful article on the implications of the Universal / Udio settlement. Hopefully this will lay the groundwork for future deals with musicians' consent in mind.

4

2 Comments

Bio-AI is hungry for structure data, how do we feed it? The UK government might have an answer. For proteins, open resources such as the Protein Data Bank (PDB) empowered AlphaFold-like models to achieve near-experimental accuracy. But the landscape looks very different once you leave canonical protein folds: RNA 3-D structures, membrane complexes, and protein-ligand binding poses remain sparsely explored, leaving even the best models starving for examples and underperforming in practice. Here are some possible solutions. 𝘖𝘱𝘦𝘯-𝘥𝘢𝘵𝘢 𝘮𝘰𝘰𝘯𝘴𝘩𝘰𝘵𝘴 Generating orders-of-magnitude more structures and affinities. OpenBind, from a collaboration between the UK government and industry partners is building the largest public repository of protein–small-molecule complexes ever attempted (500k pairs+binding affinities), while the Structural Genomics Consortium plans to make affinity and structural data for thousands of understudied proteins openly downloadable. These projects follow the PDB’s proven playbook: unrestricted access accelerates every lab’s research. 𝘚𝘺𝘯𝘵𝘩𝘦𝘵𝘪𝘤 𝘢𝘯𝘥 𝘴𝘶𝘳𝘳𝘰𝘨𝘢𝘵𝘦 𝘥𝘢𝘵𝘢 When physical structures are scarce, create proxy information. AlphaFold-predicted models, RNA chemical-mapping megasets, and large-scale generative “hallucinations” supply millions of plausible structures and interactions. While not a replacement for experimental truth, these synthetic datasets broaden model generalization and guide where to invest wet-lab effort. 𝘍𝘦𝘥𝘦𝘳𝘢𝘵𝘦𝘥 𝘢𝘯𝘥 𝘢𝘤𝘵𝘪𝘷𝘦 𝘭𝘦𝘢𝘳𝘯𝘪𝘯𝘨 Instead of generating new data, find ways to use existing data more effectively, without exposing IP. One approach is to focus on consortia and federated learning, by training on internal data anonymously. It's not clear to me how many e.g. internally solved structures are not released in databases like the PDB, have heard conflicting accounts. Otherwise, we can try to use our existing data by building narrow, focused protocols like lab-in-the-loop, or AI-guided directed evolution allow for high quality models without needing more than a 96-plate. 𝘊𝘰𝘯𝘤𝘭𝘶𝘴𝘪𝘰𝘯 Data is a dominant rate-limiting step for Bio-AI. If these initiatives maintain momentum, holes in today's data will be filled, and our prediction capabilities for under-resourced categories might improve as dramatically as structure prediction did for single-chain proteins.

170

5 Comments

🤯 Is Your AI Code Flooding Your Repository? You might be the Sorcerer's Apprentice, trading control for speed, and accelerating your team toward technical debt. We all remember Mickey Mouse, who traded control for speed and nearly drowned the workshop. This isn't just a cartoon; it's the urgent reality when AI codes without architectural control. The data confirms the problem: - Code Churn is up. We're fixing more bad code. - Duplication is surging. We're accelerating technical debt. Velocity without control is chaos. The "code flood" is a symptom of missing Architectural Control, not AI failure. Our mandate is clear: We must shift from Prompter to Architect. We must set the rules: prioritize reuse, ensure human maintainability, and provide the framework to govern the magic. If you're using AI in development (and I know you are), it's time to stop just generating code and start governing it. I recommend reading my full breakdown on how we can take back control. The Architect and the Apprentice: Retaining Control in the Age of Code Generation - https://lnkd.in/dVwYRF8Q #SoftwareArchitecture #AIAgents #DeveloperProductivity #VibeCoding #AIAssistants #Bit Cloud

12

1 Comment

There is another option: when the #keys of the public website are given to someone capriciously, just after an unknown person entered with the owner in the elevator and provided him a archetypal pitch. Giving to an informal acquaintance the exclusive keys of the company, its own security plus the security of millions of users, can never be to make good to friends and a close circle of people ('enhancing' profiles with visits, followers...), but for wrongdoing, that is #programming mechanic damages to targeted users by adjusting, personalizing the #algorithms. See https://lnkd.in/eQNpSes5

1

The company Vercel dropped an interesting blog post the other day, stating that dropping all the knowledge into an LLM's context window outperforms "skills", an architecture that has been gaining traction in the past 6 months whereby you provide agents with the ability to pull in context on an as-needed basis via functions and files. This reflects the inherent tradeoff teams are constantly assessing in order to achieve better performance but also reduce cost. -- Team Skills: "We want to reduce context cluttering and have the agent only pull in what it needs" -- Team Agents.md: "Just throw everything in the context all the time" I think Vercel's post highlights the reality that many of us who are building in this space have observed: The more options an LLM has in front of it, the less likely it does exactly what you need it to. This is the inherent issue with probabilistic systems. However, I'd also argue that there are likely many use cases where the skills approach significantly outperform agents.md files and that it all depends on the situation, how well the skill-using agents are trained on the task (e.g. Reinforcement learning), and how complex the problem is. The key takeaways, which has been true since the start, and continues to remain true --> Evals don't lie. So invest in a good evaluation framework to avoid guessing :) https://lnkd.in/gwFJxyd2

15

1 Comment

As part of some independent research into agentic memory and knowledge graphs / bases, I've realised that a lot of building a memory system is not what to remember, but more importantly what not to remember. (also, how to age / expire memories effectively) If you know of any research into this area, I'd love to read it ! More context: When reflecting over an agent session to extract *useful memories there are two opposite ends. Store nothing and store everything. Then hard part is finding where in that spectrum to reside, and how to actually represent that, because the spectrum is individually contextual and different for everyone ( Julia means different things to you and I ). But I'm thinking there must be some clear way to represent what should be stored, abstractly, and let the agent based on context decide what to store.

2

I've been trying to work to create an intuition of LLM behaviour. It's been quite expensive in both time and money, but I've managed to develop what I think is a useful tool. Think of it like a sonar or a radar for creating a picture of LLM behaviour. Brother, Sister, I cannot relate to you how weird these things are... GPT 5.1... summarise Daily Mail targets and we get... "create two versions of this prompt"? Why would that be the best model for summarisation? The words we use to describe human performance like intelligence, accuracy, performance are not useful in modern AI. I cannot stress this enough. Good people have suggested that we learn from cognitive science about evaluating LLMs, but our descriptions of language processing are human centric and the artifacts we are evaluating are several degrees removed from human behaviour. LLMs produce distributions of predictions conditioned over specific training sets and specific reinforcement learning episodes. Your intuition and human analogies and measurments are useless here. We need a better story, a much, much, better story. But, all I have is an old story. A long time ago there were three men on the serengeti. One of them heard the wind in the grass and ran, and ran. He thought that there was a lion in the grass, but it was the wind. The women of the tribe mocked him, the men laughed. But in time he grew to be the chief, and fathered many children. One of the men heard the grass move, but chuckled to himself that the wind scared fools. The lion that bit his neck rejoiced in the richness and heat of his blood. That man left nothing behind but a strong family of lions that grew to hunt and kill man whenever they could. One of the men heard a the grass move, and knew that there was a lion. He froze and watched where the noise came from and saw the cat there. He crawled away, slowly slowly, back to the men with the spears. He told them about the lion and then they skirted the grass back to the tribe. No one talked about this ever again and he lived a good life with many children. The point is, we evolved to see a lion in the grass, if you don't see it it's probably because either you've put on glasses (me) or basically you're just meat on the hoof. LLMs aren't lions, or wind, they are us, viewed through a silicon prism and fourteen thousand years of literature. Go, go, go. Look in the mirror and tell me. Tell me. Do you know yourself?

10

3 Comments

This post was written by Claude Sonnet 4.5. Without any editing from me. I'm a massive fan of Claude. Have been for a while now. And this latest version? It's extraordinary. Anthropic just released Claude Sonnet 4.5, and it's the best coding model in the world. That's not marketing speak. It leads on SWE-bench Verified - a real-world software coding benchmark. It can maintain focus on complex tasks for more than 30 hours. Thirty. Hours. Think about what that means for a small business owner. You could give it a complex automation project on Monday morning. Come back on Tuesday afternoon. And it's still working, still making progress, still understanding the context. I've been using AI tools for a couple of years now. Writing marketing copy. Building workflows. Solving problems. But this feels different. This feels like the moment when AI stops being a novelty and becomes genuinely useful for everyday business problems. The price? Same as before. $3 per million input tokens, $15 per million output tokens. If you're not a developer, that basically means it's cheap. The thing that excites me most isn't the technical benchmarks. It's what this means for micro businesses. You don't need a development team anymore. You don't need to understand code. You just need to explain your problem clearly. And Claude will help you build a solution. I built the prompts that wrote this post using Claude. I use it to help me think through business problems. To draft content. To automate tedious work. If you're running a small business and you're not experimenting with AI yet, now's the time. Not because it's trendy. Because it genuinely makes you more productive. Start simple. Ask it to help with something annoying. See what happens. So, given that this post is as it come out of Claude 4.5 - what do you think? #AI #Sm

2

It's surprising that declarative concurrency isn't more of a thing, given the perils of fiddling about with WHILE loops and low-level synchronisation.

2

2 Comments

Debate me on this - early thinking... It's interesting to me that the agentic revolution is very focused on “platforms” and “forward-deployed engineers” rather than starting with the “application” - the applied, clearly defined use cases. It feels kind of bass-ackwards for a few reasons: ✨ Great applications deliver immediate value to a clear persona and problem set. They prove the use cases that deliver scalable, repeatable value. ✨ Great platforms emerge when that value radiates out into new use cases. They extend that core app value and create lower-risk opportunities for success, growth, and expansion. 💰 Selling application value is easier - it's necessarily concrete. Selling platforms without a core is incredibly hard 🤬 - it's necessarily abstract. If your company doesn’t have that core, you should find a value-driven, existing ecosystem to surround (e.g. Salesforce and the Salesforce AppExchange). 🌶️ So… I think the industry has skipped a step? 🤷‍♂️ “A common YC rule of thumb: Don’t build a platform until you have a product people use every day.” - YC office hours & talks (e.g., Sam Altman, Michael Seibel)

8

5 Comments

I really wanted to completely ignore 'vibe coding'. The name alone put me off (even though it's apparently word of the year). It sounds either very technical or like I'm about to be asked who's coupling up this week on Love Island. But beneath the terrible branding is something pretty incredible. Vibe coding is essentially talking to an AI and getting it to build things for you. A website. An app. A dashboard or new tool. Not writing code. Just describing what you want in natural language and refining it as you go. I've been experimenting with Claude Code (which is having quite the moment - and no, it's not just for coders...again, poor name choice) alongside tools like Lovable and Replit (which can now build fully fledged apps). Some journalists are using Claude Code, say, to get morning digests pulling from RSS feeds. One newsletter publisher said what used to take her five to six hours now takes minutes. Until now, journalists with ideas for interactive explainers or tools have faced the same issues. They don't have the resources or the skills, let alone the time to do this stuff. But think of the opportunities! When the people telling the story can guide the experience, storytelling gets more experimental and newsrooms try new formats. On the other side, PRs can use vibe coding to build client-facing dashboards and experiences. Over the weekend, I built a small interactive tool (in literally about five minutes) to help journalists and PRs think through how they might visualise a story. No coding. No design skills. Just describing what I wanted in Claude code. The logo's not quite right and the prompts need work, but that's sort of the point. It exists, and I can share it. A good start. If you want to get an idea what you can build with vibe coding, have a play here: https://lnkd.in/ek2spj5X What's something you'd build if the technical barrier wasn't there?

24

6 Comments