The Ladder · honest numbers as of 2026-07-10
Build order is not launch order. Every rung below carries the number we'd stand behind today, and — where that number isn't flattering yet — the specific fix already in progress. Nothing here is a projection dressed up as a result.
The compiler, the four backends, and the acceptance discipline that makes every other rung on this ladder checkable instead of asserted. This is the thing everything else stands on.
Source: efficient/DASHBOARD.md — Strand (A) Lang & Compiler; Chapter Stats
A 30B-class MoE model, running on one consumer GPU, dequantized bit-exact against llama.cpp (0 mismatches across 114,688 values checked). The memory win is real and measured. The speed is not — we're not hiding that.
Honest update (re-measured 2026-07-10): the earlier FUSE and kernel-launch theories were both tested and falsified — FUSE overhead is ≈0% once warm, and batching the launches made it slower. The real limit is PCIe bandwidth: ~1.1 GB is streamed per token over a Gen3 link (12.4 GB/s), capping this box at ~10.7 tok/s. So the claim is the memory regime — 3.1× less VRAM, bit-exact — not the speed.
Source: efficient/DASHBOARD.md — Strand (E) Navigator, re-measured 2026-07-10 (W3, nav_moe_qwen3_gpudq_fast.cu)
Real C functions, converted to QUANTA and triple-gated (native ≡ interpreter, compiled against the original C, and value-checked equal) — not claimed, certified. 967 at the start of this build order, +18 and +42 in the two most recent runs as the converter's type coverage grew, then −1 pulled under the honesty floor (a harness-coincidence false-equivalent).
Wild-yield on real-world code is climbing as coverage grows: 1.7% → 4.75% after the int/arrays/typedef unlocks — still early, and shown as such.
Source: efficient/DASHBOARD.md — "THE ONE NUMBER (LOCK-47.8)"
A self-improving loop where the model writes QUANTA, the gate checks it, and only genuinely-computing programs are kept — no Python, no downloaded base model. The gains are modest by design: this is the flywheel proven at small scale, not a scaling claim.
Current ceiling is grammar expressiveness (linear maps, not yet rich algorithms) — the next unlock is corpus breadth, not a bigger model.
Source: efficient/DASHBOARD.md — Strand (D) Training
One law, applied to this page
8.3 tok/s next to llama.cpp's 201 looks bad, on its own. Next to the 3.1× VRAM reduction — and named honestly as a PCIe-bandwidth ceiling (~1.1 GB streamed per token, not a bug we’re hiding) — it’s a rung on a ladder, not a hidden weakness. If a number on this page ever stops matching the dashboard it was pulled from, the dashboard wins, and this page changes within the hour.
We use cookies.
We use privacy-friendly analytics — nothing until you choose. See our privacy notice.