blob: 1d87e897303569ab624f6926b8ae3b610ea91c6e (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
|
## HEADLINE LOCKED (2026-07-11, user):
"The largest neural network ever trained without backpropagation — and every component is analog-compatible."
- Armor the claim with "trained FROM SCRATCH" (kills the MeZO/zeroth-order comparison: those fine-tune
BP-pretrained models; largest from-scratch BP-free to date ~ tens of M (DFA transformers, EP ConvNets)
-> 1-3B is 1.5-2 orders of magnitude beyond the field).
- "All components analog-compatible" per standing doctrine: network fwd/feedback physics analog;
optimizer state digital-side (factored-Adam co-processor) — phrase as "the learning physics is
analog end-to-end" if pressed.
- Reply to Ben WAITS for: stage1b endpoint (today) + bf16-cascade gate (queued after stage1b on GPU1)
-> then email = measured cost table + two-tier menu (single 3B run in-cap / $40k program w/ scaling
ladder + BP twin + benchmarks), total-envelope question, BP-twin attribution question.
|
