iiinova — where your digital self keeps evolving

I’ve been using omlx on the same MacBook. Qwen 3.8 27B is incredibly slow but output is decent, can’t fit much more than a 128K context window though. 3.6 35b a3b 4bit is noticeably less capable but soooo fast

030

johanna @jola.dev · 20h

The many many sins

040

Reposted by johanna

Lewis @oyster.cafe · 20h

teamwork makes the dreamwork

test: introduce test jail to hide lewis' sins
845eee12
jola.dev
4 days ago

3716

johanna @jola.dev · 28/09/2026

Oh, and think about what you're trying to accomplish. Is the goal just to churn out code and features, or is the goal to build a community of people that discuss and collaborate? The two are in direct conflict.

120

johanna @jola.dev · 28/09/2026

PR descriptions that explain _why_ you're doing something opens the door for a much more interesting conversations than ones that just say _what_ or that are empty. For the same reason, never let LLMs write the description.

110

johanna @jola.dev · 27/09/2026

I guess we could tag `:main` or something as a moving tag for the latest version, but it's not impossible that that can be broken at any given time as stuff gets merged. Even then, there's nothing that gets your app to restart and pull latest, you'd still have to do that yourself

110

johanna @jola.dev · 27/09/2026

oooh, I don't think container registries generally support that use case, they're primarily designed around you asking for a specific version. You could set up a workflow to fetch latest tranquil git commit sha on man and then try pulling the image for it

110

johanna @jola.dev · 27/09/2026

When submitting interesting blog posts to shelf.cafe you can now click the little button to fetch the title for you, cause who has time for typing anyway

010

johanna @jola.dev · 27/09/2026

Tranquil PDS now pushes builds to container registry on merge to main, in case you want to grab a recent ref and live dangerously tangled.org/tranquil.far...

181

johanna @jola.dev · 27/09/2026

Is the video game Lylat Wars based on the movie Independence Day?

000

Reposted by johanna

shelf.cafe @shelf.cafe · 26/09/2026

shelf.cafe now renders your profile picture if available, piggy-backing on Bluesky's CDN. Long-term it would be nice to have a self-hosted image blob CDN. Note that we do not depend on Bluesky for fetching the profile and all a Bluesky outage can cause is for the fallback image to be displayed!

031

Reposted by johanna

fig (aka:[phil]) @bad-example.com · 23/09/2026

PART TWO: All the other reasons i don't like CARs for atproto repo exports! In this one: ambiguities and inefficiencies. (the biggest problem is already hopefully getting fixed in the IETF draft though, yay!)

data-service-diaries.pckt.blog

War on CARs: the rest of it

Part 2: streaming struggles, bad blocks, worthless weight, and superfluous structure in Content Addressable aRchives

6538

johanna @jola.dev · 22/09/2026

I work for incident.io if you're looking for alternatives, happy to chat!

360

johanna @jola.dev · 22/09/2026

In that case the source code might be better yeah haha

010

johanna @jola.dev · 22/09/2026

Reading the source code instead of reading the docs often means the model suggests solutions and answers questions based on internal details that are subject to change, instead of the implicit public API of what the author chose to document and "guide" you towards

120

johanna @jola.dev · 22/09/2026

I assume it depends on harness/model, but I think they're trying to avoid it because more and more websites block their traffic, so training the model to read the code instead is lower friction

110

johanna @jola.dev · 22/09/2026

I’ve been thinking a lot about what are almost like dark patterns in LLMs that make them much harder to use for learning about coding than they need to be. They’re designed around taking action, and it can be hard to get them to just support in you in learning jola.dev/posts/ai-ant...

jola.dev

AI is antithetical to learning

LLM agents can write a lot of code for you very fast, but they also take away opportunities to learn and grow.

1335

johanna @jola.dev · 21/09/2026

Okay, one more! Kite now supports zstd compression. It fetches the latest dictionary on startup, seamlessly handles dictionary updates, and it works without any new dependencies, instead relying on the OTP 28 `:zstd` module tangled.org/jola.dev/kite

tangled.org

jola.dev/kite

An Elixir atproto Jetstream V2 subscriber library

0110

johanna @jola.dev · 21/09/2026

When I migrated to tranquil I noticed that two blobs were missing from my `listBlobs`. After going down a deep rabbit hole and making two big PRs, after upgrading to latest tranquil I now have my two blobs back!

0170

johanna @jola.dev · 21/09/2026

Aggressive application of mutes, blocks, and labelers, and it’s not bad

010

Reposted by johanna

Tranquil PDS @tranquil.farm · 20/09/2026

The even more interesting (for us) part is that with this update we are also expanding the maintainer team to officially include @jola.dev !! We are very excited to have her on board for the shenanigans! Lastly we've seen an upshot in community contributions and we are very thankful for that! :3

0412

Reposted by johanna

Tranquil PDS @tranquil.farm · 20/09/2026

So we don't usually make version update posts but this is a bit of a special one. We just released v0.6.7 which in and of itself isn't immensely interesting, it's a very bug fix heavy release just like v0.6.6. A lot of those bug fixes should fix compatibility with some apps namely Colibri. But ...

10527

johanna @jola.dev · 20/09/2026

You can now define multiple endpoints to subscribe to. If it fails to connect it fails over to the next one. Since Kite uses timestamp cursors, with the rewind failover "just works".

000

johanna @jola.dev · 20/09/2026

It's not easy to blaze trails, really appreciate the things you contribute to the community!

111

johanna @jola.dev · 20/09/2026

The recent jetstream v2 east downtime encouraged me to implement failover for my Elixir subscriber library. On @calabro.io's recommendation I switched to timestamps for live subscriptions, and I added a little rewind when switching between instances to handle clock skew tangled.org/jola.dev/kite

tangled.org

jola.dev/kite

An Elixir atproto Jetstream V2 subscriber library

2140

johanna @jola.dev · 20/09/2026

Also you're doing archive replay after the 10s rewind, you can just pass the rewound timestamp and go from there. And it looks like if you don't configure an API key it will just fail to rewind and skip to live?

110

johanna @jola.dev · 20/09/2026

Thank you for the library link, I did look at it when I implemented failover for Kite. I think you can simplify it a bit. Timestamps are valid cursors, you don't need to translate them to seq to failover normal live subscriptions

120

johanna @jola.dev · 20/09/2026

I understand the reasons for wanting a lot of work to be client side and relying heavily on the PDS, but it's not a great user experience when your website downloads 25MB of data just to render some shitposts. The atproto community could do with caring more about performance and bandwidth

030

johanna @jola.dev · 20/09/2026

You standing up for what you believe is appreciated. It means so much to so many people.

090

johanna @jola.dev · 20/09/2026

Yeah, I’m thinking of kind of steady state subscribing to live or near live data, not during a backfill. My assumption was that the sequential cursor was the default mechanism and timestamp more of a fallback. Timestamp being recommended makes failover easier!

121

johanna @jola.dev · 20/09/2026

Do you mean use timestamp based cursors over seq in general, or just for the cutover?

120

johanna @jola.dev · 19/09/2026

You can pry the “I think” from my cold dead hands. I think.

030

johanna @jola.dev · 18/09/2026

hmm, that shows it as up? I haven't received a single event from it since 10pm BST last night, and any attempts to connect to it now returns 503

110

johanna @jola.dev · 18/09/2026

I guess the next feature I add is automatically falling back to the other jetstream...

120

johanna @jola.dev · 18/09/2026

If you've got Elixir on your machine, you can try staring straight into the sun with this script that prints the unfiltered jetstream straight into your terminal, using tangled.org/jola.dev/kite tangled.org/strings/did:...

tangled.org

stare_into_the_jetstream.ex · by jola.dev

A quick demo script that subscribes to the jetstream and prints all events live

3161

johanna @jola.dev · 18/09/2026

ironically `jetstream.us-east.bsky.network.` is returning 503 for me right now

110

johanna @jola.dev · 18/09/2026

I've extracted an Elixir Jetstream V2 library from @shelf.cafe. Define multiple subscribers, filter on collections and dids, keep track of cursors. All the good stuff! Makes it trivial to connect to the live data stream for atproto in #ElixirLang. Check it out! tangled.org/jola.dev/kite/

tangled.org

jola.dev/kite

An Elixir atproto Jetstream V2 subscriber library

1293

johanna @jola.dev · 18/09/2026

Sounds like it doesn’t need a lot of resources but a generous bandwidth allowance would be nice? I haven’t run anything in OVH yet, most of my stuff is on Hetzner. How’s the experience/reliability etc?

110

johanna @jola.dev · 18/09/2026

Have you decided what you’re moving to?

120

johanna @jola.dev · 17/09/2026

Hopefully the session going stale shouldn't happen anymore. Let me know if you see it again!

110

johanna @jola.dev · 17/09/2026

Your latest RSS feed entry has a category that's just an empty string, which my parser wasn't handling. Fix is rolling out now tangled.org/jola.dev/ann...

tangled.org

Handle RSS feeds with empty categories · jola.dev/annot.at@d1a1c03

annot.at is a service for syncing static websites, such as blogs, to atproto and standard.site

120

johanna @jola.dev · 17/09/2026

Oh I spoke too soon, I found the issue! Fix incoming

200

johanna @jola.dev · 17/09/2026

Hmm, I just tried a publish and it worked. Could you try logging out and in again, just in case it's related to the session. I've seen issues in the past and I thought I fixed them, but worth a try

100

johanna @jola.dev · 17/09/2026

Agreed! Standardization is really hard and even where it works it might not have the result you’re actually looking for. Lots of lexicons is more fun

050

Reposted by johanna

Macro Mayhem @macromayhem.fm · 17/09/2026

🎙️ Just published a new episode of Macro Mayhem: 005 - OTP Security Releases, ElixirConf US, and Goatmire Badges. Have a listen:

share.transistor.fm

005 - OTP Security Releases, ElixirConf US, and Goatmire Badges

We recap ElixirConf US in Chicago, unveil this year's AtomVM-powered...

095