Matt Hull

The most common argument against remote, "you can't build a real team without people in the same room," gets the diagnosis wrong. The culture problem most remote teams hit isn't really about remote. It's the result of offices producing team chemistry by accident, in ways no one designed and no one would think to replace until they were gone. Our co-founder Brant DeBow writes this week on running BiTE as a fully remote firm since 2008. The honest version of what eighteen years has taught us: what an office provides without anyone noticing, what we've built to replace it, what's still hard, and why the remote-vs-office binary is asking the wrong question. #RemoteWork #EngineeringLeadership #DistributedTeams #CTO #EngineeringCulture

3

1 Comment

I am now convinced that most teams are asking the wrong AI architecture question. I was recently at "Architecting the Future of AI with Google Cloud" when a medical transcription use-case came up. The company operates internationally, with 800+ physicians. Same language "English", but different accents and speaking patterns. The question being debated was: Should this be solved by fine-tuning a model, or by building a multi-agentic AI system? My first instinct was to compare the two. But the more I thought about it, the more I realised: Fine-tuning vs Agents isn't really the right question, and neither is: "Should we use RAG?" These technologies aren't interchangeable alternatives. They solve fundamentally different problems. A simpler way I now think about it: 1. Do you need the model to behave differently? Think fine-tuning. 2. Do you need the model to know information that changes or lives outside its weights? Go with RAG / retrieval. 3. Do you need the system to actually DO something: call APIs, use tools, maintain state, execute workflows? AI Agents are the way to go. And sometimes the right architecture uses more than one. For the physician transcription example, accent variation alone DOES NOT automatically justify a complex multi-agent system. You first need to ask: what exactly does the system need to know, learn, and do? That conversation stayed with me. So I went back, formalised the thinking, added the economics around volume, stability, accuracy, and cost, and turned it into a decision framework at MLDeep Systems Because enterprises don't need the most sophisticated AI architecture. They need the simplest architecture that reliably solves the problem. Full framework in the comments Oh, and one last thing! A python code which DOES NOT use an LLM to reason, and then take a call to act; is NOT AN AI AGENT. And you can not convince me otherwise. #ArtificialIntelligence #AIEngineering #GenerativeAI #MLDeepSystems

16

10 Comments

As Genius Scan continues to grow, keeping our documentation clear and up to date is critical. With GitBook, reviewing and improving content as a team is simply easy. They asked us a few questions about how the migration of our Knowledge Base to GitBook went. 👉 "We can create a new article in one section while simultaneously updating another topic in a different section. Each edit is reviewed independently, so we're never blocked waiting on a single review or worried about overwriting changes or losing base content.” If you’re scaling a product with a growing knowledge base, tooling matters more than you think. 👉 Check out the full story: https://lnkd.in/etsmp84i

16

2 Comments

Steve Yegge believes coding by hand will gradually disappear, grudgingly acknowledges that software engineers no longer need to understand Assembly (and is unhappy about this), and says most devs should NOT use his agent orchestrator, Gas Town. I always enjoy Steve's energy and him thinking two steps ahead: this conversation was no different. Watch or listen: • YouTube: https://lnkd.in/e4bDbPKp   • Spotify: https://lnkd.in/e7_j-4sd • Apple: https://lnkd.in/ec6w7je9 Brought to you by: • Statsig – ⁠ The unified platform for flags, analytics, experiments, and more. https://lnkd.in/eQkXqjMZ • Sonar – The makers of SonarQube, the industry standard for automated code review. https://lnkd.in/ekMzUJxY • WorkOS – Everything you need to make your app enterprise ready. https://workos.com/ Three interesting thoughts from Steve that we talked about in this conversation: 1. Reading ability is becoming a blocker for wider AI adoption. Some struggle with walls of text that current AI tools produce, and Steve predicts that in the very near future, most people will program by talking to a visual avatar, not reading terminal output because he observes that five paragraphs is already a lot to read for many devs. 2. What software engineers need to know keeps changing. In the 1990s, any decent software engineer knew Assembly, and today almost no decent developer knows it because Assembly has long been superseded by technical progress. What engineers “need” to know these days is different from the ‘90s and that process continues with AI, changing the parts of the craft that are essential for devs. We grumble about this but that won’t change anything by itself. 3. There’s a “Dracula Effect” where AI-augmented work drains engineers faster than traditional work. This is because AI automates the easy tasks, meaning that engineers are stuck doing high-intensity thinking all day. Steve says you may only get three daily productive hours at max speed, but during that time, you could produce 100x more output than before.

281

31 Comments

The introduction of MCP servers and Claude into our environment has been like unlocking a seemingly endless amount of possibilities. This week I have configured an MCP server that combines SQL tooling with a couple related vendor APIs and the results are incredible. Asking plain language questions and Claude being able to utilize the workspace and tooling available to formulate responses, then outputting that data into presentation grade materials is incredible to watch. Connecting these otherwise disparate systems with a lightweight app is like integrations on jet fuel.

4

I recently relaunched tablecloth.time — an (experimental) composable extension for time series analysis for the Clojure data processing library tablecloth. The redesign involved dropping the index, which led to some interesting reflections on abstraction and composability. I discuss those ideas here. https://lnkd.in/gRA-PuqH

21

What if I told you that a 1 GiB/s Kafka topic streamed directly into your S3 data lakehouse as an Iceberg table could cost you... $10/hr?  Bufstream does it. It's literally too good to be true. They are a core part of the massively deflationary wave that's exerting commoditization pressure and crushing Kafka margins. How did this Protobuf start up build their own diskless Kafka implementation in no time? 🤨 I spent 20 hours researching them. Here's their story 🧵 

200

19 Comments

I recently wrapped up a project where we rebuilt, stabilized, and added new features to a Drinking Water Explorer Tool, for the folks at the Environmental Policy Innovation Center. This tool is for anyone who wants insights on Drinking Water Data at the local level, across water systems in the United States. > TOOL: https://lnkd.in/gxXAf5Ka This is now an open-source codebase, written using a Rails and Hotwire framework. New contributors are always welcome - the GitHub repo and instructions can be found here. > REPO: https://lnkd.in/gYkEXXim Kudos and thank you to all of the great folks at Environmental Policy Innovation Center, Thrive, and The Bitfoot Company for making this happen, and for being awesome to work with!

34

1 Comment

I promise not to repost all of these 😄 But as I'm stumbling through marketing, I wanted to share for reach. 501see is all about the data, and I'm unearthing interesting, actionable information for #nonprofits a couple of times a week. For a anyone curious about the process: - Codex and I brainstorm the ideas, usually questions I'm curious about - An agent pulls the data and some (usually very rough) copy - I polish and usually rewrite about 80% of it - And we collaborate to make the infographic (HTML rendered to PNG) Follow there if you're curious about #giving trends, #grant insights, and other fascinating tidbits in IRS filings.

6

1 Comment

Earlier this year I had to downsize my staff, which sucked hard. However, fewer chefs in the kitchen means less worry about code conflicts between parallel projects, so I've been taking advantage of that. I have been running the gamut of AI skills related to security, design, architecture, design patterns, anti-patterns, testing, best practices, etc. and letting them run hog-wild on an enormous thirteen-year-old ninety-thousand-lines-of-code product. The results have been impressive. It found edge-case gaps in test coverage. If fixed some elusive "only breaks on a full moon" bugs that have been lingering for years. I have been able to optimize the speed of deployments and test suites. Performance of the product in general feels much improved even though that wasn't a target metric. Seeing a decade of accumulated tech debt get melted away so fast has been shocking.

7

3 Comments