Jonathan DeLaigle - CTO & Software Engineer

10/20/2025

10/20/2025

Ever find yourself talking to a duck at 2 AM? 🦆
You might be a developer.

It’s called rubber duck debugging — when you explain your problem out loud, you usually find the answer yourself.
Who you talk to doesn’t matter, they just need to be a good listener... even a rubber duck.

Sometimes you just need to talk it out.

10/20/2025

I've been thinking about how we talk about technical debt.

Learning a better way to do something isn't really debt, it's an investment opportunity. The old code isn't wrong, it's just outpaced.

Over time though, even good investments can turn into debt if they block secure updates or make onboarding harder. The key is knowing when to reinvest and when to let go.

Interesting read:

Some history on term “technical debt” and on better language to use when communicating about it.

10/14/2025

Imagine this.
An LLM trains on public data like Stack Overflow, blogs, and forums.
Now imagine an attacker quietly slipping in just 250 crafted documents.

That might be all it takes to plant a backdoor.

Anthropic's new research with the UK AI Security Institute found that as few as 250 poisoned documents can teach models of any size a hidden trigger. In their test, the phrase made the model output gibberish.

We are not talking about one liners. We are talking roughly 420k tokens, about the size of A Game of Thrones by George R. R. Martin. Not huge, but still big. Yet small compared to a 13B parameter model's full training data.

Models in production today are not known to be susceptible.
But this finding changes a long held assumption.

It was thought an attacker would need to poison a large percentage of training data. Turns out the real number might be fixed and surprisingly small instead of millions.

If a poisoned trigger can survive pretraining, it gets pretty scary if that same model later gains tool access, code ex*****on, or API hooks.

Even if the effect is just noise today, the lesson is simple.
AI security starts with the source data and that has become a much bigger deal than before.

Full study: https://www.anthropic.com/research/small-samples-poison

Anthropic research on data-poisoning attacks in large language models

10/09/2025

You can’t just document what is being migrated -- you’ve got to document what’s not too.

You’d be surprised how much important stuff gets left behind until you spell it out.

10/09/2025

I read an article that said there are only two kinds of managers: empathetic and ruthless.

After about a decade leading engineering teams, I’ve seen both types up close, and I don’t think either one tells the full story. The best leaders balance empathy with ex*****on, heart with edge.

That’s where real leadership lives.

Article that sparked this:
https://www.seangoedecke.com/ruthless-managers/

--

10/09/2025

Commented-out code is just code in denial.
If you’re not gonna use it, delete it — git already remembers.

10/08/2025

One config to rule them all.

AGENTS.md is quickly becoming the README.md for AI coding agents. It gives all your tools a single source of truth for build steps, conventions, and system rules.

One file. All agents. Shared understanding.

That is how AI development suppose to work.

https://www.devshorts.in/p/agentsmd-one-file-for-all-agents

AGENTs.md - one file, work across many agents

10/07/2025

We’ve entered the cleanup phase of AI-generated code--and it's going to be messy.

AI is fast, but it's not wise. Without guidance, it'll produce more slop than structure. Senior engineers still need to steer the process and keep the architecture sound.

With great power comes great potential--and an even bigger mess if you're not careful.



👉 Read the full piece:

From hype to hard lessons in AI engineering

10/02/2025

AI agents should already be your new team member. They are here to stay. They can accelerate workflows, but like any teammate they need context and oversight.

I use them for drafting, ideation, and structuring workflows. They help me focus on higher-value decisions while they take care of the busywork.

Are you putting AI agents to work yet?

The article that got me thinking: https://www.oneusefulthing.org/p/real-ai-agents-and-real-work

09/30/2025

In the pickup line today. From the desk: the coding agent really is just a super fast junior dev. And of course, the secret to getting good results hasn’t changed — write a clear, solid ticket. Funny how some things never change.