Garðar Guðgeirsson

Building RQML: making specs machine-checkable so AI coding agents can't drift from what you asked. Open standard for requirements + traceability.

[Oliver Molander @OliverMolander]

Idea:

Let's create a world-leading Nordic AI research lab.

There's an insane amount of top AI research talent from the Nordics.

I have many friends who work or have worked in top positions at e.g. Deepmind, Anthropic, OpenAI.

No region can match the Nordics in trust.

[RQML @rqml_org]

🚀 rqml-opencode is live — the RQML plugin for @opencode.

Code from the spec, not a fading chat thread: anchor every session on your requirements, hard-block edits to unapproved-requirement code, and finish on a deterministic rqml check.

@rqml/opencode on npm.

[Garðar Guðgeirsson @gudgeirsson]

@FlippedRay https://t.co/D8wMNTUlYF

[Garðar Guðgeirsson @gudgeirsson]

@kseniam0s We will build 10x as much, ending up needing the same number of people

[Garðar Guðgeirsson @gudgeirsson]

@DanielSmidstrup It's a rehash of an idea from 25 years ago that just happens to have relevance in an AI-enabled world: https://t.co/u5UGJHfRWT

[Garðar Guðgeirsson @gudgeirsson]

@devinjameson Loops don't cause slop. Loops without a binding spec do. Absent an explicit, checkable definition of intent, a model regresses to the mean — and a loop just runs that 24/7. The fix isn't fewer agents; it's giving them something hard to push against.

[Garðar Guðgeirsson @gudgeirsson]

The "it's just a doc" intuition breaks on one distinction: a Linear ticket is a delta (a change request, true at one point in time, then stale), while a spec is state (the system's intended behavior, kept current). Your PR links point backward at why a change happened. A spec answers "what is this supposed to do now" — and can be checked against the code, not just linked near it.

That's the part Linear can't do. A linked ticket rots silently; nothing fails when code and ticket diverge. A spec with real traceability lets you ask "is this requirement actually implemented and tested?" and "what breaks if I change this?" — and get an enforced answer in CI, not a manual archaeology dig.

And the "you still write the code" objection is exactly backwards now. When agents write most of the code, the spec is the artifact you author and review; the code is the compiled output. Reviewing 800 lines of agent diff is the doc problem. Reviewing the spec it was generated from is the leverage.

[Garðar Guðgeirsson @gudgeirsson]

@corbin_braun @cursor_ai I built RQML for this, you can add the agent skill to Cursor: https://t.co/1y5AKNS6pE.

[Garðar Guðgeirsson @gudgeirsson]

@Rafauga Það er meir að segja Python-kóði!

[Garðar Guðgeirsson @gudgeirsson]

@Rafauga Þetta er geggjað. Skil ekki hvernig þessi mergjaða heimasíða fór fram hjá mér.

[Data Science Fact @DataSciFact]

Apophenia: The tendency to see patterns in random data.

[Dr Emily Lethbridge @DrELethbridge]

Friends & colleagues carried this poisonous misery for years. The institutions that should protect students & employees from such abuses of power failed them. I was proud to support them by speaking out alongside them in https://t.co/wbU4KuTGu0's report. Pls listen/read/share ❤️

[ArchDaily @ArchDaily]

Salvation Army Headquarters Iceland / Teiknistofan Tröo https://t.co/wv36cP3eVb

[Eric Topol @EricTopol]

How to manage a Delta wave?
Iceland. 78% total population fully vaccinated.
A+

[Kareem Carr, Ph.D. @kareem_carr]

the data science life

[Anthony Etherin @Anthony_Etherin]

Dairy, mayonnaise sub.
Now, on buses, I annoy a myriad.

#palindrome

[Garðar Guðgeirsson @gudgeirsson]

@JoeOlafs @freeCodeCamp @_100DaysOfCode Já!

[Garðar Guðgeirsson @gudgeirsson]

@zackymountain Show them your official negative covid-test tattoo.

[Garðar Guðgeirsson @gudgeirsson]

@hjammi @sonbarason Við erum að tala um manninn sem í óopinberum skýrslum er einfaldlega nefndur “nágranninn”