We’re basically outsourcing email judgment to AI, then trying to compensate by strengthening SPF/DKIM. That feels like hardening the locks while handing out more master keys.

I think there's several things going on at once in that space. This is just their vocalisation of it because it suits them.

The thing is Fastmail can't speak with absolute authority about email because Fastmail is not email. It's subordinate to it.

Genuine question: does the writing tool matter at all here if the exposition is clear and mathematically correct? I’ve seen great notes written in Word, LaTeX, and even slides—quality seems independent of format.

I would say it's not statistically independent. See https://scottaaronson.blog/?p=304 item #1. So we get to add another exception, which is fine.

That feels more like a selection effect than a property of the writing tool itself.

Well yes, if using TeX could turn me into a legit math researcher, life would be a lot easier. ;)

Interesting!

both no in principle, and when you're used to reading LaTeX, word is ugly. It's a milder form of how if these notes were handwritten it wouldn't matter, but it would also be less appealing than them being typeset well.

Really nice write-up. After reading through the survey, I’m curious if there’s any evidence that more complex heuristics (ML-based or multi-stage policies) actually outperform simpler budget + size + hotness models in production JITs, or if they mainly help at the margins.

Nice implementation. Have you thought about supporting LoRA fine-tuning on top of this, or is the design too low-level for that kind of extension?

Sure it could be extended to support LoRA finetuning but this implementation has the goal to be as lean and efficient as possible for a pre-training stack as you can be.

At work I mostly use Codex, while for personal projects I've settled on OpenCode. The split is less about model quality and more about context. Work projects benefit from consistency and predictable workflows, whereas side projects are where I experiment with different models, prompts, and agent setups.

I'm curious whether this will materially reduce costs for local authorities or whether the benefits are primarily in expanding payment options.