Partly. Another factor in the equation is load.

Running a VM is just a qemu-system-whatever command. They usually get pretty lengthy, but it's nonetheless an ordinary shell command.

Strange. A good engineering manager would see that "way too big" PoE daughter board design as exactly what one would want in a reference design that will be used to test and integrate your preferred PoE solution. Power product life cycles are so short and availability problems so frequent that a good engineering manager knows that their engineers will be reworking power solutions with some regularity.

A good engineering manager would also know that UI development for commercial products is not optional. The engineering manager will expect that marketing will want branding at the very least, that differentiating features will need to be surfaced, etc., and that all of this will need to be integrated into build, test and the package system, and QA'd on real hardware. Basic stuff for an engineering manager.

This is wrong. OpenWRT is fostering several manufacturers that are using OpenWRT as the factory platform for their products. This is a reference design (one of several, this particular one from 2024 is now dated and newer designs are available,) provided by OpenWRT, and they've thoughtfully made it available to anyone that might want one: you can just go buy some with no NDA bullshit and get your developers moving in your lab or doing UI development or whatever. The not-cost-optimized PCB is what you want for this, in addition to the ample RAM+Flash. The "useless" POE is another aspect of this: access points use POE ubiquitously, which is a key OpenWRT use case.

That reality is carefully left unclear in all this "silicon sovereignty" narrative. It's a nice plant with new tooling cranking out 300mm wafers, but it's not the same game as a TSMC or Intel fab making cutting edge, high margin silicon, and there are a number of competitors making similar power devices around the world. And yes, the "AI" fluff is pure marketing nonsense; everyone needs lots of power devices for everything. Yes, they'll obviously seek lucrative contracts to supply exotic power devices for AI applications, but that stuff gets commoditized quickly.

If you had 1024 of these, you'd have a grand total of 1 MiB of RAM. You won't be compute clustering much with 1 MiB of RAM.

Using this for a long lived "developer environment" would be extraordinarily expensive anyhow. Scaling the vCPU + RAM cost of these to the same shape compute optimized Graviton On-Demand EC2 instance (16 vCPU x 32 GB RAM) shows about 4x the cost.

So don't do that. Just use an EC2 instance.

Broadly speaking yes, this is the business model. IBM has been at this for many years with technology transfers, licensing agreements, support and other arrangements. Rapidus, Samsung, GlobalFoundries, ST, SMIC, AMD, etc. have all used IBM R&D work at various times for various nodes and products.

The cutting edge of semiconductors is a writhing mass of copulating tapeworms, and IBM lives deep inside that ball. For IBM, what this means is that when you buy one of the ASML machines to make products with this process, you'll pay IBM for the knowledge and support to actually get it working, or give them a cut, or something else, TBD, as circumstances warrant.

Taalas has a running demo here: https://chatjimmy.ai/

It's eye opening: generated an AVX-512 optimized Mersenne Twister in C in 0.076s, 13,706 tok/s. Too fast for the tok/s to be terribly accurate.

Know that the increase in CPU utilization may mean you've improved the performance of your "database server," because now your CPU cores are waiting less on IO. It also may not mean this, but just looking at htop won't tell you either way.

There really isn't an excuse, and it isn't possible to hate null/nil/undefined/etc. enough.

Seems like a nice, basic, affordable platform for workgroup/SME stuff. Not NetApp/Pure Storage "enterprise" grade though.

Consider that this shared sense of rigour you have in mind is illusory, and LLMs and their context struggles are simply revealing this. I see precious little rigour in any of the 'tech' world I've lived in for decades. The tools proliferate, paradigms emerge and die and reemerge, and whatever stick you consider using to measure any of it has competitors with different units. Past the physics of power and signaling, and the prevailing cost of a silicon wafer, we are almost all, relative to a small number of much older disciplines, muddlers of various degrees of skill.

I've found dealing with context limits relatively easy: specify and confine. LLMs need clear specifications and strong guidance to produce good work.

But that's just my current muddling take on the practice. Perhaps, 90 days from now, even this burden will be gone, and a simple prompt will generate world class operating systems, programming languages and a formal basis in mathematics for both.