Martin Thwaites

Quick reminder that people are still finding the fediverse, so keep coming back to you contact lists on the various sites you follow them to see their new handles. Mine is @martindotnet@hachyderm.io

@eddiejaoude I think we do it pretty well, none of us actually feel like "devrel". We just... do what needs to be done. A lot of "DevRel" people get really precious about what devrel is, and lose focus on the fact that the job is relationships inside and outside the org.

It is interesting to see how different companies do DevRel differently. Which companies do you think do it really well and why?

Can confirm, seeing this on the stage was amazing

@aHev @davidfowl @kellabyte @aspiredotdev You are definitely missing a lot of the project if thats your view. Logs, traces, events, metrics, profiles, SDKs, global semconv, local semconv, proxies, schemas, validations, protocols. Naming headers isn't even otel, that's W3C.

@davidfowl @kellabyte @aspiredotdev Use it all the time and love it. Agreeing on headers is hard!

This was a pretty incredible journey. I rememeber when .NET had its own version of distributed tracing (didn't everyone before otel??). Now it's a no brainer, if you aren't using otel for tracing you should be. aspire.dev/dashboard/stan… #otel #aspiredev @aspiredotdev

@kellabyte @davidfowl @aspiredotdev Thats not an "otel" problem, it's a tracing problem Otel as a project is way more than tracing. For your specific usecase, I'd advice looking at the events spec, its more aligned to emitting "start/end" Between logs, events, metrics, traces and profiles, we've got you covered

@davidfowl @aspiredotdev One major drawback to OpenTelemetry traces is that it’s only useful for tracing completed requests. You can’t correctly view partially completed long running request traces to diagnose where it’s stuck. Design of OTel format kinda sucks in that way.

@Aaronontheweb OpenCode+ollama+gaming GPU is surprisingly effective.

The thing I really appreciate about local AI and home-rolled agent harnesses is that this is where the people who are in software for the love of the game are creating a garage-hacker movement. OSSing everything. Building and sharing tools. Experimenting. 100000% contrast to the AI doomer / grifters / slop-maxxing / course sellers. If you are feeling down or worried about how AI is going to impact your career, this is where you're going to find your love of building + learning things again and probably drastically increase your market value as a developer right now too. You can probably jump in with an old gaming rig or a Mac and start right away - you won't be replacing frontier models with that gear, but you'll be surprised what you can automate and accomplish even with small models.

@Aaronontheweb Qualcomm is my long term bet to be honest, even longer is Qant but thats more like a 5 year thing. There's a lot riding on the next generation of the halo platforms (beyond gorgon). APU and NPU combinations are the future.

@MartinDotNet I'm long on AMD in that space but I think they have the most catchup to do on the software stack

A great question in the Netclaw Discord: "what should I use this for?" I'm a business owner and have dozens of use cases that apply to me professionally; I'll share those AND some personal use cases.

@Aaronontheweb Yeah this is how the MBP is staying ahead in portable inference.

@MartinDotNet The other bit is using the native hardware supported quantizations - smaller models can suddenly perform a lot better on Sparks despite their parameter counts because of this

@Aaronontheweb This is the whole ball game right now. Model choice, context size, quantitative, etc. This is what will make or break a good experience with LLMs and everyone else is trying to one shot everything with Opus 4.6 (or even 4.7). The concurrency stuff is the hard part imo.

@MartinDotNet I am very torn between running cheaper MOE models with high concurrency to handle lots of parallel agent work (i.e. sales / marketing agents doing lots of schedule and event-driven work) or using my RDMA connection and running Deepseekv4 flash or something

@Aaronontheweb Theres an Instruct model that might be better for that. I can't remember which is have loaded for it, I can check later.

@MartinDotNet this was my experience also, so for anything programming or diagnostic I stick the 27b. I'm going to run the 35b MOE model for stuff like our marketing / sales automation because it's probably acceptable for those workloads

Giving Qwen3.6:27b a shot with the model token prediction (MTP) support that just shipped on llama.cpp This should be a significant speed-up over what I've been using currently. My AMD Radeon 9700s max out at ~24t/s for small prompts w/ thinking on github.com/ggml-org/llama…

@Aaronontheweb I have a strix halo with 128GB, just a single one. I would love me some DGXs, or any of the DB10s.

@MartinDotNet I also have a pair of DGX-Sparks I'm prepping for use with the additional agents I'm spinning up

@davidfowl Protocol is stable now, thats the part that matters for Aspire ;)

@MartinDotNet These are just traces but I’m also capturing profiles (not otel ones yet). I don’t know where the spec is on profiling

Massaging the profiling telemetry to make sure we can diagnose issues with the aspire cli using the aspire dashboard. #aspiredev

@Aaronontheweb Interesting, although i was more interested in your personal hardware. The prefill and TPS can only really come from behind the openai endpoint as its not exposed to the caller. Those ones I'd love to see them have gen_ai semconv tracing and metrics. I shall try it next week.

@MartinDotNet The other dynamic context that got fetched here was probably from our memory system netclaw.dev/architecture/m…

@Aaronontheweb @AkkaDotNET @htmlcsstoimage @getsdkbin I should show you the new canvas for doing those investigations with visuals and human interaction. Not to say Netclaw can't do a good investigation.

- Marketing automation. Every GitHub release of @AkkaDotNET , make a graphic using @htmlcsstoimage , a twitter thread, a LinkedIn thread, and post them all only after I approve. - Engineering automation. Receive webhooks from Seq when failures occur in @getsdkbin and investigate

@davidfowl I've been toying with the idea of hooking up wispr flow to Copilot during these meeting with the prompt "any time a feature is discussed, spawn a sub-agent to build it and update the agent as the discussion unfolds, when the work is pushed to a PR, tag everyone in slack".

Now I have meetings with people that try to explain why we need more meetings, and in parallel, I spin up more sessions to do that work 🙃. The next 5 years are going to be wild.

My AI psychosis has subsided.

@davidfowl I'm pretty sure it runs docker-compose build && docker-compose push /joke

@davidfowl Does asking my AI Coding Agent to deploy everything count?

How many of you have a single step, fully automated release process?

Putting together my itinerary for June and have the following open evenings where I'd love to give a talk at a meetup or user group: Belgium (Antwerp): Tue 9 June Belgium (Antwerp or Brussels) or London: 15 and 16 June Let me know, details at: ted.dev/where