chicagobuss

my 'local llm on a budget' has cost me just about $2,400 in total, spread over a bout 3 years of accumulating stuff.

I got my first 7900xt 20gb on a good sale shortly after they came out in an open box deal on ebay for $590.

After another 7900xt 20gb popped up on facebook marketplace recently for $420, I sniped that and then got a microcenter deal on an asus proart board, 32gb of ddr5, and a 9950x for $1050.

A couple more deals for used cases, an older 1tb nvme ssd I had lying around, and a nice PSU finished the build.

I run Qwen 3.8 27b on linux and have tuned the llamma.ccp / rocm setup substantially with the help of codex and claude. I get ~80-110 tok/s depending on all kinds of factors and full 256k context windows with a 6-bit quant.

It's surprisingly capable and I love that I can run it indefinitely, tuning as it goes to behave more like a frontier model.

When I try talking for prompting my coding agents, I'm always severely disappointed with the specificity of the prompt. It's just much easier for me to "talk about" coding specifics with a keyboard than it is for me to speak about them, for some reason.

I suspect this has to do with the nature of the work you're doing; I dunno, just count me firmly in the camp of people who only rarely feel like I can express a thought sufficiently for coding agents with their voice.

same

> Multi-agent is mostly dead

Wow that's a huge take. Seems like you're on your own with this one.. everyone I know is scaling up the number of agents they're able to keep fed with good, useful work.

Ha, your columns look a lot like why I made my own as well - https://github.com/chicagobuss/crusteo

I loved the utility of paseo but like the OP didn't want to build on web frameworks, so I re-wrote all of it in rust with zed's ui framework for the frontend desktop client.

More on my personal blog about it here: https://chicagobuss.dev/blog/my-setup/

you... don't read hackernews much I take it? This is like 10% of all submissions lately, lol.

when something is that cheap, just remember, you're giving something valuable to the people selling you that thing

I think claude code is just way, way, way more than "just a harness" at this point.

You have to remember all of the surrounding features it works with now, like the infra behind the cloud agents, or how remote control works, the mobile apps, their connectors, team skills and mcp servers, on and on....

why does the pareto cost per task chart not have 6.1 sol?

awesome job to the whole team. Congrats Bryan and everyone!

I agree with you and disagree with the OP in similar ways.

I work on my own "agentically-engineered" projects and the pursuit of certain ideas for what you're calling intrinsic motivation, and I don't particularly care if others find it interesting or even find it at all or not, because the ideas themselves are cool to me.

It's kind of like I can suddenly tinker at 2,000% efficiency and share what I'm doing, _and if someone else finds it interesting, that's cool too_

Yes

It's entirely possible that in their testing of newer models, the whole problem is that even if it's doing better in benchmarks, maybe it's insufferable to work with, thus they're not releasing it.

go ask 100 random people across the world and I bet less than 20 will know both those company names.

I'm confused, isn't a self-hosted cloud something like openstack or kubernetes?

what a cool answer

It's also very convenient to not have to remember a bunch of extra combos or perform the gymnastics to press them, especially if you need other modifiers too like shift+pgup, shift+ctrl+pageup, etc

Tracker is a lightweight document and work-tracking system designed around three goals:

1. *Totally self-hosted and 100% private*: I run it over ZeroTier to connect my systems, but Tailscale or just your personal LAN works great as well.

2. *Flexible, immediately usable, & extensible*: It can use S3/R2/etc for the actual document storage if you want object storage, or you can just use local files. I've been running RustFS myself on a small local k8s cluster which emulates S3. It comes with migration tools to switch from one storage system to another, and it is ready to run in many different forms all designed for AI agents to be pointed at it and immediately self-bootstrapped. "Look at the tracker tool at `<ip:port>` and configure your own mcp and skill using the instructions within" is all you need to tell an agent.

3. *Very lightweight, reasonably fast, and designed more for agents to use it well rather than humans*: That being said, it does have a minimal web UI included so you can also browse your documents.

Furthermore, it has a few design quirks which I really wanted—some people will probably love them and others might hate them. You have to remember that I have a... problem? with simply loving computers, so I tend to work from several different machines simultaneously, and that's a major motivating factor here.

- *Flat structure with tags & folios*: Tracker is only slightly opinionated on document storage hierarchy. You can store a massive collection of raw docs in a flat format, or you can use *folios*—which are just documents of `kind='folio'` whose metadata tracks its files (themselves documents tagged with `folio:<slug>`). There is no other way to introduce additional hierarchy; to me, this is a major benefit—I just wanted one big flat space.

- *Explicit locking on everything*: All changes go through an explicit locking mechanism, even simple metadata edits. This is to prevent an agent from reading a stale document before taking action.

- *Strict attribution*: All edits require an actor ID (sent via `X-Actor: <role>@<hostname>/<ip>`). Because I work across Linux, macOS, and Windows simultaneously, it's critical to know exactly which coding agent is doing what, from where, and on which machine.

More details on my personal blog: https://chicagobuss.dev/projects/personal/tracker/

Duckbill is hiring three-five engineers for 3 roles with hybrid work in downtown SF. We're in-office MWF and WFH Tuesday and Thursdays. Relocation is an option but sponsorship / remote is not.

We're working on Skyway, a new kind of enterprise-grade finops tool born out of the Duckbill Group's consulting business. We're staying small (team-wise) but focusing on the largest of the SaaS spenders and their problems exclusively. We ship daily and have several unfair advantages.

1. Senior Front-end Engineer (TS/JS/React + bit of Python back-end)

2. Senior Back-end Engineer (Python + bit of front-end)

3. Senior Data Engineer (Python, airflow, duckdb, s3, bigquery)

More details are on our careers page (apply through this): https://www.duckbillhq.com/careers/

why not just use llm by simon willison

perfection

internal reports from current AWS engineers seem to be confirming all of the speculation in this article. Shit's rotten from the inside out and you can pretty evenly blame AI, brain drain, and good old fashioned "big company politics"

https://forums.theregister.com/forum/all/2025/10/20/aws_outa...

Duckbill engineering is hiring a full-time ONSITE senior data engineer in downtown San Francisco. We WFH Tuesdays and Thursdays.

We're tackling the hardest finops challenges for the biggest enterprise customers head-on and have a killer team along with the advantage of the established and industry-leading DuckBill Group.

Help shape the leading edge of an entire industry with us while we're still just at seed stage and under 6 engineers!

Comp range is generous on the equity and between 180k and 210k base depending on seniority. We're cooking with flask, react, postgres, clickhouse, parquet and airflow on AWS at the moment.

---

Apply here: https://www.duckbillgroup.com/careers/

Or email: [email protected]

Duckbill engineering is hiring a full-time ONSITE full-stack senior SWE in downtown San Francisco. We WFH Tuesdays and Thursdays.

We're tackling the hardest finops challenges for the biggest enterprise customers head-on and have a killer team along with the advantage of the established and industry-leading DuckBill Group.

Help shape the leading edge of an entire industry with us while we're still just at seed stage and under 6 engineers!

Comp range is generous on the equity and between 180k and 200k base depending on seniority. We're cooking with flask, react, postgres, and clickhouse on AWS at the moment.

---

Apply here: https://www.duckbillgroup.com/careers/

Or email: [email protected]

this is awesome. I especially appreciate the detailed testing notes.

Peering is an option as well but it's a whole different ballgame of complexities to set up vs. private link

I'm surprised this take isn't gaining more ground in these discussions.

people keep talking about how "this is what all these Americans want" - bologna I say. They're just voting how they've been programmed to believe.

Find me a Trump supporter that has only researched him from first-hand credible sources and has not been influenced by friends, family, social media, or mainstream media. I would be very surprised if any person like this exists.

last time the checks and balances were more prepared to handle (block) him.

this time, they're primed to enable and embolden their plans (see project 2025).

I don't expect these four years to be anything like the last time. I legitimately won't be surprised if this is the last somewhat-normal presidential election in the united states