AI Roundup: 2026 Q3
And here I am again with another installment of this. I’ve dropped the “Article” from the post name since I’m now including more than just article links. I hope someone finds it helpful.
Update: I originally included some personal thoughts about AI as well, but I’ve decided to move them into a new post instead.
Disclaimer: Each link’s inclusion simply represents information that I felt was interesting or helpful at the time that I noted it and does not necessarily indicate agreement with its contents.
Articles
- John Yang: Humans & AI (Ep. 1) - Gigachad Strikes: AI vs. Human showdown - who comes out on top? (2025)
- Avi Press: After 7 years in production, Scarf has reluctantly moved away from Haskell + HN + Reddit
- Fabien Sanglard: Don’t you mean extinct? + HN
- Kiran Gopinathan: Multi-agentic Software Development is a Distributed Systems Problem (AGI can’t save you from it)
- Sean Goedecke
- Fabien Sanglard: My agent.md to improve LLM-assisted code quality
- Earendil Product: What is a Harness?
- Terrible Software: AI Is Breaking This Thing We Call Trust
- Mark Seeman: On learning programming in an age of LLMs
- Chris Done: LLM-Complete: The next 700 programming languages
Discussions
- People Who Got Laid Off or Quit Their Jobs, What Are You Doing Now? — tangentially related, but interesting
- Question for devs using AI: How do you currently check AI-generated code for security vulnerabilities?
Quotes
The problem is that it sounds like you’re trying to offload the cognitive work to AI. This is something I often see. People get on the hype train, abuse LLMs, burn out because code review is exhausting, reading code is harder (and less enjoyable) than writing it, and they’ve let AI take over the interesting bits of their job, and then get on the doom train.
The rule is moderation and targeted application. You never leave the drivers seat. AI is often better at analysis than generation, especially as things get more complex. Use it for adversarial brainstorming, organising thoughts, clarifying complex topics, ramping up on new tools and repos, setting up infra, porting code to a new stack or language, writing boilerplate or simple (but tedious) changes, doing code review. They can augment your thought process, but they cannot take over judgement for you. And you control how much you give up to LLMs. Don’t choose to give up the parts that make you interested or fulfilled. Additionally, I often see people making the mistake of asking AI to be both the author and reviewer. You absolutely should never do that, for the same reasons we never let PRs merge without code review from at least one other person, ideally two. You can also use deterministic validation as feedback for the LLM to keep it on the rails. The borrow checker is one of the reasons why LLMs work particularly well with Rust.
It’s absolutely possible to use AI responsibly and constructively. I’ve actually been having more fun since I started using them because I treat them like a collaborator who can also do the bits I hate, not the bits I love.