Brian Artigas

I help businesses stay secure & run smarter through IT, cybersecurity & a bit of AI magic | Founder @AllstateCompute | Always learning, always improving

[Brian Artigas @brianartigas]

@lxe You'll be hearing from Linus's vast IP legal team. They've already publicly posted their cease and desist letter here: https://t.co/TCVWDzHnee

[Brian Artigas @brianartigas]

@neerajjj6785 ChatGPT, not because it's the best at any one thing but because it's good enough at everything and the others have gaps in those areas where OpenAI is good enough.

[Brian Artigas @brianartigas]

My Qwen inference nodes are currently limited to 256K and 160K context windows, and sometimes I need more reach than that. That said, modern coding harnesses, context management, and persistent memory are making that less of an issue.

For large codebases, I also find frontier models useful for architectural decisions, complex debugging, and reviewing the work of local agents.

I haven't had a chance to run GLM 5.3 Flash locally yet. I don't have the hardware ready, and eBay and Amazon are taking their sweet time delivering it. 😂

[Brian Artigas @brianartigas]

@bekahj After all these years we have self driving cars that talk to us and now turbo boost and pursuit mode. We finally got K.I.T.T., it just a Tesla not a Trans Am.

[Brian Artigas @brianartigas]

$100/month for OpenAI. Local AI does all my actual coding. ChatGPT handles PR reviews and orchestrates my local coding agents.

My Qwen 27B and Qwen Flash Next agents rip through hundreds of millions of prompt tokens and tens of millions of generated tokens daily.

We don't need to talk about the capital expenditures. 😂

[Brian Artigas @brianartigas]

@Morrowmake Week 40, I'll just take out a second mortgage on my house for a B300.

[Brian Artigas @brianartigas]

B) A router. Explaining it isn't really that difficult. 192.168.1.10/24 and 192.168.2.20 are on different networks. If you want two different networks to communicate, you need a router.

Now, if we're talking about 192.168.1.10/24 communicating with 192.168.2.20/22, that's a different story. That's how you identify someone who truly understands subnetting and routing rather than someone who just memorized the answers.

[Brian Artigas @brianartigas]

@komal_uk01 I use Zorin because I work on both Windows and Linux computers, and it makes switching between them easier on the muscle memory. Mint does too, but I wanted non-experimental Wayland support.

[Brian Artigas @brianartigas]

@PeasantSmith I guess it's a good thing LG uses a the modern equivalent to a Z80 to run their TVs. That thing can barely open Netflix without crashing . Immune through sheer computational incompetence!

[Brian Artigas @brianartigas]

@superalesha Amazing work!

[Brian Artigas @brianartigas]

@PeasantSmith I was playing with Minimax H3 and made an 80 second trailer for a YouTube mini series I want to make... This music video is proof of how terrible I am at prompting this sort of creative work. 😂 He did a great job on this!

[Brian Artigas @brianartigas]

Knowing how to code helps, but I think systems thinking is more important. A good automotive technician, for example, already understands dependencies, troubleshooting, inputs/outputs, and failure modes. Give them some basic software knowledge and a capable AI coding agent, and I bet they could build a functional app.

The harder part isn't writing the code anymore. It's understanding what you're building, how the pieces fit together, and recognizing when something is wrong.

Security, race conditions, scalability, etc., are still concerns, of course, but their importance depends on the type of application.

[Brian Artigas @brianartigas]

@devops_prashant I read every semicolon in that Python code!

[Brian Artigas @brianartigas]

@axhiii88547 Better for what exactly? That's like asking, “What's the best tool?” and declaring an adjustable wrench the winner.
Cool. Here's a Phillips-head screw. Your wrench sucks. 😂

[Brian Artigas @brianartigas]

@PeasantSmith @unbug Yes, those are US eBay prices right now. I'm specifically looking at DDR4 ECC 3200. I'm planning a dual EPYC system so I can use all of my PCIe slots without BAR allocation headaches, and memory bandwidth will matter if I'm involving the CPUs in inference or offload.

[Brian Artigas @brianartigas]

The only thing keeping me from replacing my dual Ivy Bridge Xeon motherboard and CPUs at home is DDR4 pricing. $1,200 for 128GB of RDIMMs is just nuts. I'm not even considering anything that takes DDR5 for the home lab.
I'm getting ready to try one of those SXM2 NVLink interconnect boards and see what four V100s can do when they can actually talk to each other at hundreds of GB/s.

[Brian Artigas @brianartigas]

Been playing with Gufo on my Ryzen AI MAX+ 395 / Strix Halo box and I'm pretty impressed so far.
I'm running Qwen3.8-Flash-Next locally on the Radeon 8060S with ROCm 7.2.4. With prompt caching doing its job I'm seeing ~33 tok/s decode, cached prompts restoring in ~11ms, and ~368ms TTFT on a 5.2K-token request.
What interests me more than the raw tok/s is that Gufo is actually making Strix Halo feel usable as an inference endpoint. Prefix caching, speculative decoding, and the low memory overhead make this little box much more interesting than I expected.
Still testing concurrency and longer agent workloads, but so far this has been a fun setup.
Apparently I needed another inference engine to obsess over. 😂

https://t.co/XzvijjEoNJ

[Brian Artigas @brianartigas]

I tried to run Strata, but this poor old dual Ivy Bridge Xeon home lab server doesn't have AVX2 support, only AVX. I've been scrolling eBay for a dual Epyc to replace the dual Xeon setup, but having difficulty coming to terms with DDR4 pricing. I have DDR3 RDIMMs coming out of my ears.

I'll spin it up at my office and let you know. I have a lonely 3090ti sitting on the bench waiting to be put to work.

[Brian Artigas @brianartigas]

So far I think I've found my ideal local model for an RTX 3090: Qwen 3.8 27B 3-bit GSQ.
It fits at full context in vLLM and I'm seeing ~40–50 tok/s decode, with cached prompt processing reaching several thousand tok/s. More importantly, the code it produces is actually good - which is not the case with several other highly lobotomiz... uh, I mean quantized models.
I've tried several models/quantizations across llama.cpp and vLLM, and this is the winner so far:https://t.co/y1uwpJvJov