Xantier

I have a server in the closet running AMD Ryzen 5 3600, 9070 (non-xt) and 64GB of DDR4-2400. Running Qwen3.8-flash-next on Strata and pushing ~60 t/s decode and reading ~1.5k t/s prefill. Works well for the tasks I'm doing with it. For external access I'm exposing the OpenAI compatible endpoint via a CloudFlare tunnel and using it on a hosted Viberglass instance.

The cost was around 850 for the GPU and memory a few months ago, other hardware has been in the closet running my server for around 5 or so years.

I built an internal tool that does this. Having the ability to spin up sandboxes on demand to run a harness(es, depending on the task) and having a collaborative thread which multiple people can chip in with their opinions. Mostly the attempt is to make a coding harness part of the discussions that we as humans usually have during the day (on Slack, ticketing systems, GitHub etc.) and trigger the harness to do the grunt work based on those discussions. First creating a plan what is iterated on and then doing the actual coding based on that approved/finalized plan.

Thanks for the numbers. It's interesting to see how big the difference between the top end and the smallest ones are. I think the ideal approach is likely using the bottom end with different harnesses for different kind of tasks eventually, where the system prompt and toolset is predefined, maybe wrapped by something like a harness of harnesses which decides the correct one to choose. Matters a lot for local hosting of the models as well since speeds are usually a bit slower than what big providers offer.

Still kicking around with https://viberglass.io to assist the whole development team (or smaller companies) to work together on development. Hoping to open source a better multiplayer experience on it over the next few weeks and then it's moving self-hosting options towards other cloud providers.

Which alternatives are you using for AirDrop on Linux? I haven't been able to find a good one for this yet.

Tinkering around building https://viberglass.io. It started off as a way to allow my wife to keep her website up to date without any coding skills, but ended up growing arms and legs.

It is a hosted ticketing system/agent harness platform with integrations towards other ticket systems and chat apps. It allows triggering agentic (coding) tasks without the need to context switch and/or know anything about installing the wanted tools, SDKs, IDEs etc. Ephemeral workloads in isolated containers or cloud compute. Trying to help commoditize small scale development tasks and and prevent them from getting lost in the void of the backlog.

Open source with local or AWS self-hosted, full IaC attached.

I've done a bit of research around trying to find the best, simple clients that don't need an engineering degree to set up but haven't really stumbled upon a good one yet. Claude desktop app advertises that they support all features (they don't), all the others with closer to full spec compliance are related to software development (VS Code and Continue etc.).

I find the context forge quite nice from IBM. I used a few approaches they've implemented within my Ctxpack context management solution as well.

I think the agentic use cases aside, the client side of MCPs are still lacking quite a bit and would need to mature to be able to catch up to the spec. I feel a lot of use cases exist outside of fully automated agentic approaches, since we can't really rely on LLMs yet to produce at a human level.

The underserved cases rely a lot on prompt and resource management at the moment. Being able to iterate and share those across teams to provide easy starting points to delegate tasks is something I feel would be workable for the current iteration of AI assisted work, outside of pure software engineering.

Hopefully other clients join VS Code Copilot to allow more varied approaches than just simple tool calls here. I think Copilot's approach on prompt and resource management isn't quite the best approach either though. It is still early days for MCPs in general so i think we'll see a lot of experiments in this space.

Is this fine-tuning with a more hyped naming? Let's say a company produces 100 documents per day that are appended to a knowledge base. What's the cost and delay to be able to ask questions about these docs whenever a single one is added in?

I feel differently about next.js after their latest shenanigans. Their end goal clearly seems to be vendor lock-in to the Vercel/next.js platform, whichever way it ends up happening. The current attempt looks to be to tweak and assert control over the React library, to make it Vercelified.

Neither BBC nor Medium are blocked in Malaysia. I don't understand the reason for spreading disinformation about this.

Had the same issue as well. Dry air in the apartment and flickering screen when interacting with an IKEA chair.

Solved by adding a "tail" to my chair using a piece of old speaker cable. There is a metal part on the bottom of the chair that the tail can be attached to.

Lyon - Lille flight on the other hand for tomorrow is 147 euros (cheap price compared to other days it seems). The flight takes a little bit over an hour, getting to Lyon airport from the city takes about an hour as well (18.20€ with public transport, 50+ with taxi/ride hailing). On the landing end transport to city is luckily cheaper and faster, 45 minutes with public transport, 15 mins and ~20€ with a taxi.

You can use coroutines with Kotlin JS. Your sentence doesn't really make sense to me though since coroutines and React have very little to do with each other.

I've been doing few projects with it and find it very lovely. Interop with Java works flawlessly in my experiences and RxKotlin extending RxJava makes that side work nicely as well. For one project[0] I used Ratpack as the web tier and that with Guice locked in place very easily. DB side of things I have used Kotlinquery which did its job nicely for my small use case.

In another I used Spring Boot where Kotlin felt like a native in that environment.

I'm guessing if you are building something with ORMs or similar where Java Bean convention is needed, you might lack the beauty of immutability and data classes (there are workarounds though) but other than that most pieces drop in place without a hitch.

0: https://github.com/Xantier/trycatch

We have blue ocean installed on our Jenkins instances and I periodically check when it is suitable for our use. Currently our promotions to environments etc. use `input` step on the pipeline. This is, for now, the main thing that blocks us from using blue ocean.

This is a very interesting comment and I like to think the same way. That is the way that seems to bring most happiness to me in professional life. The question is, how do we recognise these places before joining them?

Sourcetree, if you are not on Linux like me :(

For end to end usage samples and somewhat comparing to other DBs I can recommend my kitchen-sink project NERD-stack: https://github.com/Xantier/nerd-stack