=== What's next

I'll try another message source (maybe Whatsapp or Discord) and see if this triage approach generalizes beyond email.

Full post: adamwiggins.com/posts/fireh...

So it works! But why hasn't Gmail, Superhuman, or anyone else implemented something like this?

Best theory so far: the "annoying homework" of initial setup and ongoing review labeling doesn't work for a mainstream product.

=== Not so good

Emails sent to “Review” often feel like annoying homework when more than one or two at a time

The 1.8% of mislabeled emails triggered some irrational anger in me, signal that algorithm accuracy is a critical part of the UX

=== Good stuff

Loving that the priority inbox surfaces ~3 emails a day, always directly relevant to my current projects / life priorities

It was a joy to delete decades worth of accumulated Gmail and Fastmail filters

All peer-to-peer systems have to compromise one way or another on this: they can decide that direct P2P is essential and simply fail, they can provide centralized fallback alternatives, or they can try to find/convince/incentivize another peer to route traffic on your behalf.

Hole punching for NAT traversal is a key technology for p2p, definitely. But there are vastly more technical issues beyond that. @iroh.computer is doing some good work here.

😎

Save the date (July 12-14 in Berlin) and sign up for the mailing list or join the Discord to find out when tickets go on sale.

www.localfirstconf.com/

Malleable software, open social protocols, self-sovereign identity, and local models are a few topics on our mind.

What do you think should be on-topic for Local-First Conf 2026?

In 2026 we want to expand a bit beyond CRDTs and sync.

Our industry is rapidly changing due to LLM-assisted coding. But our core values of data ownership, user agency, and user empowerment are as salient as ever.

In 2025 we expanded to two days and double the capacity, while trying to keep the same energy.

This was a year that "sync engines" really came on the scene, and one whole day was dominated by that topic.

In 2024 we wanted to bring together academics, idealists, and builders to see if we would all gel.

We did, and a community was born! We sold out the 150-seat venue in just a few days.

By @cultrepo.bsky.social, filmed at @localfirstconf.com firstconf last year

Featuring builders from our community including @martin.kleppmann.com @rsms.me @pvh.ca and many more

I covered that approach in the post. LLM with tuned prompt and a few examples scores pretty badly. Fine-tuning on the other hand gets pretty close to the classifier approach, so that's an option.

Great post, hits home with my family's experience. I'm a US citizen originally, now dual US/German.

Love this:

> They use language like “this is the future of coding” when talking about their product. It’s snake oil — empty, self-important hype. This is something we’ve worked hard to avoid in our group.

Interesting parallel to my "actionable emails" goal!

I also want to be informed about changes in local or national policy that affect my life, as these may be actionable. For example, changing winds in German politics encouraged me to pursue dual citizenship last year.

This approach is the first one I tried before the classifier :)

I didn't like it because it led to exactly what you describe here, which is basically custom prompt tuning. That's the sort of "futz with your gmail filter rules" that I want to avoid.

But yeah maybe worth further experiments!

What's your goal for consuming news? e.g. find interesting/cool articles like the AI ranching example. Or being informed on whatever's in the news zeitgeist in the moment without being too drawn into it. Or something else?