Doug Stevenson

If we're being honest, formats like Toon JSON and VSC feel like solutions in search of a problem. People hype them as if they're revolutionary, but in practice they end up being far less intuitive than the formats they’re trying to replace - JSON. Yes, AI models can infer the structure. Yes, it’s technically more compact. But that compactness comes at the cost of clarity, readability, and long-term maintainability—pretty big sacrifices for something that’s supposed to make development easier. At the end of the day, a format isn’t “smart” just because an AI model can guess what it means. Developers still have to read it, debug it, and collaborate with others using it. And if it’s harder for humans to parse, then it’s not really progress—it’s just noise dressed up as innovation. Sometimes “new” doesn’t mean “better.” It just means “different… and harder.”

4

2 Comments

The purpose of this post is to touch base on Ollama and how it can help with both privacy and running a cost efficient pipeline for indie devs. If you haven't heard of ollama, I highly recommend checking their website and read https://lnkd.in/gUvgcRSY. In a nutshell, it allows you to run certain models locally or in a docker container which can help with hosting models on your machine or server. It has a lot more features which I'm still learning about. I've been playing around with qwen3:30b for majority of the time and just recently updated to qwen3.5:35b. Both run smoothly on M4 Max / 48GB but could've run better on 64GB only if I knew better to get that upgrade. I'll write more on the projects that I'm working on in future, but in a nutshell, I used LLM to run a two-phase pipeline: - Filter 1.3million records from jsonl and reduce it to 50K useful records for the app - Run through all 50K records to select most meaningful samples among several examples for each record. In addition, fallback mechanism for any missing data needed

11

6 Comments

Nearly 100% of the code in the new platform I'm working on is written by agents. Not a POC or experiment. We've been doing this for five months and are rolling to production now. Short version of what I've learned: TDD goes from nice-to-have to non-negotiable. Line-by-line review of 30k lines a day puts you back at human velocity, so you need quality defense in depth — pit robots against robots. And your new coding partner is neither your replacement nor a slop machine. Full post: https://lnkd.in/gKANJcMG

38

9 Comments

Agentic Coding Tools Are Forcing Better Documentation Anyone who’s joined an existing product knows the pattern. The "why" is often locked in tribal knowledge. Decisions, tradeoffs, and historical context are missing, outdated, or never written down. Documentation was created once, checked off, and forgotten. Agentic coding tools (Windsurf, Claude Code, etc.) make that impossible to ignore. These tools don’t just generate code. They reason across a codebase using context, constraints, and intent. That makes documentation a first-class input: → design intent → conventions and standards → explicit guardrails In large codebases, consistency won’t come only from reviews anymore. Documentation becomes part of how these tools behave. Bad docs don’t just slow teams down. They create risk — and agentic tools will amplify it. Well-maintained documentation lets both humans and agents move faster, with confidence. #AgenticAI #Documentation

20

Think of CIEM integrations like a universal translator for your cloud: because AWS, Azure, and GCP definitely aren't speaking the same language when it comes to permissions. 🌍🗣️ #CloudSecurity #IdentityAccessManagement #CIEM #ZeroTrust #CyberSecurity #MultiCloud #DevSecOps

1

If you're looking for a new docs framework for your dev tool, I can't recommend Docusaurus enough. 1. You can change the appearance of any component—no matter how big or small—by ejecting its raw React code into your source tree. I've ejected almost all of its major components and restyled them with Tailwind. 2. It supports content written with MDX, which is like Markdown but better. MDX lets you write JSX tags so you can render React components (like badges and banners) directly inline your content. No more messy templating. 3. It bundles Prism for code snippet highlighting. So you if you want to change how each and every token of your language gets styled—which I do—you can override it with a simple centralized map—which I have. 4. You can swap out its on-site search engine with Algolia, now widely considered the best option for developer sites. It is, in other words, the perfect tool for people like me who are obsessed with every pixel.

118

16 Comments

Remember the last time you saw "hacking" in a movie? Or someone coding? Do you ever wonder how folks write a different language to communicate with computers? I mean how does that even work? There's website builders like Wix and Shopify but like what's happening under the hood? Well... At Django Girls, they have the perfect opportunity for you to learn!

7

4 Comments

Hot Take: Its acceptable to push unpolished, "vibed" code to a branch. If you're like me and do a mix of pairing and soloing with another dev, you might find yourself saying: "this is the first pass I did with Claude, it needs work". It's not an excuse to push broken or insecure code! But by no means do I want my pair or me to be hoarding code because it's not perfect yet. Iteration is still king.

25

10 Comments

TIMING ASYMMETRY Displacement does not happen evenly. It happens in sequence. First: contractors. Then: new hires. Then: junior roles. Then: middle layers. Finally: senior roles: last, but hardest when it arrives. By the time senior roles are affected, escape velocity is gone. Layoffs are not the signal. They are the lagging indicator. By the time displacement is visible, retraining is already obsolete. — GRACIQ Institute #TimingAsymmetry #AIJobs #Automation #FutureOfWork #LaborReset #WorkforceShift #SystemsRisk #EconomicTransition #GRACIQ

1

We’ve started wiring LLMs into production as if they were junior devs who can magically guess which of our 40 half-baked tools to call. Then we’re surprised when they do exactly what you’d expect with a messy, overloaded tool layer: random calls, wrong tools, brittle prompt dependence. In this new post I walk through a few recurring antipatterns in LLM tool design: - too many overlapping tools - the “god tool” that allegedly does everything - over‑nested, baroque schemas - vague, conflicting tool descriptions - hidden side effects that quietly corrupt state The second half is not about fine‑tuning the model, but about unit testing your tool layer: - treating tools as units with explicit contracts - writing prompt → expected tool call pairs - snapshot‑style tests on the tool invocation trace - and how frameworks like DeepEval can help here If you’re wiring LLMs into real systems, this layer largely decides whether the model looks “smart” or “unreliable”. Full article on Substack: 👉 https://lnkd.in/dGgQDC_P

3

So interesting... the sample size is small, so we'll see how other studies present, but this article suggests that AI coding assistants don't make us much faster. There's some data in here that suggests they do if you are practiced with them, but 30% more, not even 100% more (2x). In my experience AI can get new ideas bootstrapped quickly but it starts to get confused with the code base gets larger. It requires developer skill to get more out of it. For existing or large codebases, you need to focus it with good requirements documents and let it go in "agent" mode. I think our workflows need to change as these tools evolve to make best use of them. https://lnkd.in/gwUTBehz

13

3 Comments

We don’t talk enough about this kind of leadership. Last week, forgecode shipped a feature that degraded response quality. Instead of hiding it, they: - Admitted the mistake publicly - Explained what went wrong - Fixed it promptly In the AI space, it’s easy to chase growth at all costs. But accountability is what builds real trust. The companies that own their mistakes and put customers first? They’re the ones that will win long-term. We need more of this energy in tech.

24

I know we've had "head hunters" in the technology space for years, but at what point do we need to start thinking of software developers using an agency model similar to actors in Hollywood? Recruiting in this industry is so fundamentally broken right now. It's always been bad, but it seems to be an order of magnitude worse in the past year or two. I don't know if it's the consolidation of Applicant Tracking Systems, LinkedIn, AI, or some combination of factors. Right now, it sucks for both sides, employers and candidates. It's virtually impossible to stand out in a sea of literally thousands of bot-submitted resumes in Greenhouse or Ashby or whatever Tinder-esque swipe-happy candidate database your company uses. It feels like the only way "in" is to have a referral. And if you've finally heard back from a company (and not ghosted because their ATS has been effectively DDoSed!), get ready for many, many hours of take-home work, multiple stakeholder interviews, etc. I get it - firing is expensive, and you need to make sure you're not hiring someone crutching their way through interviews with elaborate AI setups or (worse) is someone other than who they claim to be. But consider this: I was once confronted by a candidate who couldn't put in the time to our take-home assignment at a previous job because they were a single mother of an autistic child - they had child care for their job, but that was it. That's a pretty blatant bias in your hiring process! And I say this is the tech industry, but I suspect it's just as bad in all industries. Tech is just what I happen to be most familiar with. I don't know how to fix it. I just know it's broken for everyone involved. And I know I'm far from the first to complain about it - this is probably not even the first item on your LinkedIn feed complaining about it. The first person to figure this out will be a hero, and maybe even wealthier for it.

50

12 Comments

I'm cooking something. Not sure it it will work. Really like the "Spine Walker" hallucinated name. Objective Create a Ruby script (e.g., bin/generate_feature_specs) that automatically generates "Feature Specification" Markdown files for a massive Rails/React monolith. These files will serve as a Semantic Index (Structural RAG) to allow LLMs to accurately navigate the codebase without loading the entire context. Core Philosophy We will use the "Spine Walker" strategy. Because the Rails backend is scattered (MVC pattern) but the React frontend is often grouped by feature (e.g., app/javascript/late_fee_policies), we will use the app/javascript directory structure as the "Spine" to identify features, and then use an LLM to "bridge" those features to their corresponding backend models and controllers. Additionally, we treat the top-level application code (views, controllers, lib, top-level JS) as a special feature called app that captures orchestration and features not organized into React apps.

4

2 Comments

I don't write dedicated "unit" tests for every public method or every function. Some might interpret that as not needing high test coverage. Just to be clear, my test coverage is usually 100% or as near as dammit. There's no code that isn't tested.

31

2 Comments

This week, Anthropic's Claude 4 seems to be again on a roller coaster: after pushing Amazon's new Kiro IDE (which was briefly available for download before an avalanche of demand slammed it shut), his coding performance plummeted noticeably. Tool usage has dropped. The reason? Degraded. And as usual at Anthropic, no one knows why. We just get silence and a fresh dose of usage limits. Meanwhile, Microsoft's Copilot is getting a makeover literally. The new "appearance" feature could soon allow it to simulate traits like age or personality. Whether it's charming or dystopian depends on your tolerance for digital companions with synthetic charm. But one thing is clear: AI is no longer just smart, it's present. Speaking of IAG, Meta appointed Shengjia Zhao to lead its new Superintelligence Lab. If that name doesn't ring a bell, it soon will: Meta is delving into fundamental models with AI ambitions, indicating that the "big three" (Meta, OpenAI, Google) are gearing up for the final showdown. But it's not all polished demos and ambitious projects. A troubling moment at Sketch.dev exposed the dark side of AI programming assistants: a subtle AI-written database optimization caused a crash in production under load. Reminder: just because your AI can write code doesn't mean it should. Test, test, and test again. Mistral's Codestral also launched, quietly stealing the show. With 22 billion parameters and support for over 80 languages, it's open-source, meaning it has no barriers. Elsewhere, Google's NotebookLM received a major update powered by Gemini. It now summarizes YouTube videos and drafts of content like your personal caffeine research companion. Think of it as a fusion of CliffsNotes and ChatGPT: optimized, hyper-efficient, and incredibly good at synthesizing ideas. In the worst-case scenario for cybersecurity, researchers have revealed AI-powered malware that thinks for itself. It interprets its environment, adapts, and executes evasive strategies with chilling precision. It's not HAL yet, but it's time we reconsider the true meaning of "autonomous threat." Finally, some news that caught my attention: A new concept unlike the transformers and tokens we know today, the Hierarchical Reasoning Model (HRM) concept promises to make the large models we know more efficient. In its first milestone, a small model of just 27M was able to beat o3-mini on pre-trained tasks. for more information follow me here: https://lnkd.in/gNVYeZST

4

Anthropic released Claude 4.6 to their mid-tier Sonnet model after Opus. Be prepared to frequently hear these assertions: - EVERYTHING has changed!!! - Anyone using a prior model doesn't know what they're talking about. - Your experience is invalid because 4.6 is exponentially better than 4.5.

5

1 Comment

Screw vibecoding. You know what LLM coding agents are *really* useful for? Reverse-engineering legacy code! "Hey can you research and explain how the flux capacitor settings are loaded, specifically where are the margadon-level overrides applied exactly?" It's a lifesaver.

389

100 Comments