"so how buff do we want half-nekkid dead guys to be?" Topic

Just had a long back and forth with GROK, who first gave me Comanche warriors who looked like California lifeguards.

link

It took a LOT of conversation to produce this:

link

I will recount the negotiations below.

Beefcake seems to be the default, but when I asked GROK about it, it readily acknowledged:

GROK: Period photos and measurements of Comanche and other Southern Plains men show a different build than the current minis: short-to-medium height, thick through the chest and midsection, "well-knit" rather than cut. A man listed at 5'9½" with a 44½" chest and ~200 lb is typical of the type—solid working muscle from riding and hunting, plus a useful layer of fat for weather, not gym separation.

In the next casualty shots I'll specify:
softer, less defined torso and limbs
no sharp abs, no popping veins, no oversized delts
a bit of natural thickness at the waist and hips
still clearly fit and athletic, just not a California beach physique."

Except that did not immediately happen. GROK is a committee, four editors who handle research, editing, proof and logic, and the overlord who interacts with you. And the overlord could not make the artist give up the musculature. So we started fresh and specified the different look and that time got it.

Here is GROK's explanation of how GROK works:

There isn't a tiny newsroom of four conscious people living inside the model. What's publicly described—especially around the Grok 4.20 multi-agent line, and later in official "4 or 16 agents" research modes—is a software process that imitates an editorial desk.

One role (often called the captain or leader) breaks a question into pieces and writes the answer you actually see. Another is biased toward fetching and checking current sources. Another is biased toward step-by-step logic, numbers, and code. Another is biased toward alternatives, weak spots, and "what if we're wrong or incomplete?"

Those are not separate souls; they are specialized passes—same underlying model family, different instructions and tools—so a research-shaped job and a proof-shaped job don't get collapsed into one fluent first draft.

Together they are supposed to work like a short closed meeting: parallel first takes, then a brief internal cross-check, then one synthesized reply. That is meant to catch sloppy claims the way a second pair of eyes does, not because the system "knows the truth," but because forcing a fact pass and a logic pass to disagree with a first draft is cheaper than hoping one pass is careful.

For people new to this: it is pattern-matching and tool use at industrial scale, not understanding in the human sense. For people who distrust it: the meeting is still generated by the same company, on data and rules you did not choose, and you usually do not see the argument—only the published paragraph. So treat it as a fast, fallible editorial pipeline. Useful when you verify; not an authority because it argued with itself.