iiinova — where your digital self keeps evolving

Me mumbling to the AI has been a daily ritual for a while now. Also while sitting behind the computer it's just a quicker way to get info into the model. They understand STT artifacts surprisingly well.

000 0

Damn, that's just harmful. We're already anthropomorphising these things to much.

It all depends so much on which model.

000 0

Yeah I read it too. Turns out we can't have nice things. Node did catch up though, it's quite nice too these days.

000 0

Sssshhhht, they are listening 🫣

I'm looking forward to a new GLM model. Can't take long anymore, what do you think, two weeks?

Johan Schuijt @resolve.works · 08/10/2026

I think sharing heavy state from orchestrator to subagents is a trap. Those output tokens are expensive compared to re-reading often. What's your thoughts there?

111 0

Johan Schuijt @resolve.works · 08/10/2026

Yeah, I've been using a pattern similar to this dynamic workflows thing, and it's really efficient. I've been thinking about solidifying it into something built On top of pi durable.

000 0

HackerOne | Daniel Stenberg | 54 comments

Here's what the curl lead said on this 6 months ago. www.linkedin.com/posts/daniel... I think the 50k driveby's are from people that don't understand development yet, and that you'll always have those. But meanwhile, most software is now improved by the people that use these tools responsibly.

linkedin.com

Over the last few months, we have stopped getting AI slop security reports in the #curl project. They're gone. Instead we get an ever-increasing amount of really good security reports, almost all don...

000 0

I can really only recommend it. The setup is a bit of a exercise, but once done, it's an amazing distro. If you're setting up, their guide is great. It helps that you already have experience with another distro. Good luck!

If that's your reasoning, you'd better stop using the Internet, because Linux also has Ai generated code in there. And everything runs on Linux.

000 0

I think that you hit the nail on the head there, but that the answer here differs per project.

I think if you just now started programming with these models, And don't read your own generated code. You probably have no idea how other people do all this and what maintenance burden means. I think it can all be explained by a lack of understanding.

The name use is the part that makes my eyebrows raise.

000 0

Releases · resolveworks/pathfinder

github.com/resolveworks... I built a local android agent, it allows me to research anything while walking around using open models. It's all there in the commit and release logs.

github.com

Minimal native Android agent. Contribute to resolveworks/pathfinder development by creating an account on GitHub.

000 0

In what way is the resource use high?

000 0

Some of the parties require a specific license if you want to run at a certain scale. Besides, I think it's somewhat like the open core FOSS approach, that the moat is in your datasets and knowledge, which enterprise will pay for.

120 0

Load is up, but the quality of AI assisted PR these days is getting good. I think this tech is making the barrier to software dev a good bit lower. I see a lot of people build stuff now that would not have done so previously.

100 0

It all depends on that perspective. Is AI training fair use? I think so, and that it's hard to argue the opposite.

000 0

SteamOS is secretly Arch 😬

That's the idea right? For the world to learn from your efforts. If you can have a machine do that, to make my work easier, that's cool. Would be nice if you release the weights, but if not, still cool 😄

000 0

It really depends on the project. For some projects, yes, everything that goes to production, I check. For other projects, I maintain seventy thousand lines of Kotlin without knowing Kotlin.

000 0

Yes, this. It's very hard to keep the robots from repeating everything the code does in comments.

Quite a few different ways, mostly pipelines. Now with the agents, I feel it's more important that you have some chain of provenance. Some log of what happened so that the agents can check.

I fully condone training models on my code just as I condone others learning from my code. I would argue that art and code are two very different things. Code is a means to an end, if you can make my life easier by improving my tooling to get to that end, I'm all for it. That does not apply to art.

"That's because now we use pipewire instead of pulseaudio and jackd to output to alsa, oh and pick the ffmpeg backend over gstreamer. Did you already install the XWayland compatibility layer?"

I think Linux for gaming is getting better and better. Did you ever try it out? Most distros you can boot from a USB really easily just to take a look.

000 0

Linux takes some getting used to, but it's really not that bad 😬. Once you have it working, performance is so much better that Windows. Also, AI is really good at Linux, so any error you can just put into the chat and get the solution.

000 0

Do it, make the switch. You'll be amazed at how an OS feels that respects you as a human being 😁. What distros are you thinking of?

Roll your own stack with Arch. Though I wonder if staying on the X stack while the ecosystem slowly seems to move to Wayland will bring you much.

That's exactly my setup. I have a extremely lean foss pixel that I daily drive, and my "consumer phone" at home with all the banking / eID stuff.

There should be a compatibility layer in there called xwayland that allows for X applications to run under wayland. If your X applications aren't working that might be the thing to check. That said, just snapd alone justifies moving away 😄

000 0

And tomorrow you will again as you forgot to setup full disk encryption 😅

Thinkpads have always been a default, it's nice machines. For the phone I'm on Google Pixel with GrapheneOS, the choice there depends more on how deep you want to commit to doing stuff yourself vs want google to do it for you.

Hahaha, love the robots in the readme!

I'm getting more and more unsolicited AI generated emails. All of them completely missed the point and are close to gibberish. There is real people tying their real names to these, though. I don't understand why. Sending these does not make you look good in any way.

021 0

I don't think they use this at OpenAI. I believe they use internal models. I have also been a bit disappointed with their latest models.

000 0

Yeah, the workflow is the harness. This is why I'm using the extensible pi harness. It can change with me over time as my assumptions change.

000 0

read-memories: Search past Claude Code session logs to retrieve context and insights

You can also analyze all your previous sessions yourself and use the results of that analysis to update either your global agents file or skill files. There's also skills out there that help you with this. For example, skills.rest/skill/read-m...

skills.rest

Search past Claude Code session logs to retrieve context and insights.

Meanwhile, OpenAI should finally be profitable on inference. Courtesy of the latest DeepSeek papers.

000 0

Looks interesting. Did you write somewhere about what I'm looking at? How it all ties together?

100 0

Yeah, using a harness from a company that sells tokens always seemed a bit weird to me.

000 0

Why would you use compaction at at task boundaries? why wouldn't you use sub-agents? Doesn't compacting also use tokens?

100 0

I think roughly fifty percent of my prompts is shaped sort of like this. I'm continuously trying to understand my codebase better. Or hunt for tech depth. If you have any tips on how to deterministically search for complication, let me know.

I never thought about forking the context for subagents i always have them start clean. That's an interesting take.

000 0

Yeah, this is roughly what I'm doing. I don't use pull requests, I use git hooks. And use the main tech lead as a reviewer and merge. I wrote about it here. resolve.works/articles/pi-...

000 0