Perception · self-hosted · his own Rust engine

Newton grew his own eyes — and they run on a CPU

No borrowed cloud model. Newton now sees through a DINOv2 vision backbone running inside his own Rust engine (candle), on an ordinary CPU — no GPU, no external API. He turns any image into a 768-number understanding, then a tiny head he trains himself turns that into a decision. This is the shared cortex that will serve every visual craft he learns.

image→ DINOv2 backbone · candle/Rust · CPU→ 768-d understanding→ tiny head→ decision

He learned to name what he sees

Frozen backbone + a tiny trainable head — the exact pattern the research says works. Newton embedded held-out photos he'd never seen, then a lightweight linear head sorted them into ten everyday categories.

··% top-1

on held-out natural images he never trained on

per-category · recall

His engine, not a borrowed one. Every one of those 768 numbers is computed by Newton's hand-built Rust forward pass on candle — the same sovereign runtime as his skin model. This is the anti-fraud line the whole project stands on: Newton's output, produced by Newton.

Watch him look

Real held-out images, and Newton's own top-3 read on each — computed live in his Rust engine. The last one is deliberate.

The honest limit — and why it matters. Point this same backbone at an X-ray and it answers with confident nonsense: natural-image features don't transfer to radiology. That's not a bug we're hiding — it's exactly what the peer-reviewed benchmarks found, and it's why Newton keeps mammography on his purpose-built radiomics model, not this backbone. Right tool, right domain.

Where this goes

Today's CPU build is Stage 0 of the plan from the visual-cortex research. The backbone is shared; only the tiny heads change per craft.

Now · CPU ✓

Shared frozen backbone

DINOv2 in candle + a tiny head per craft. Live today, this page. Candidate to replace the older scene/detector stack with one shared cortex.

Next · distil

His own ~11M student

A big teacher pseudo-labels; Newton distils a small student that runs in <500 MiB — fully his, with frozen layers to resist forgetting.

GPU · soon

Higher resolution + reach

DINOv3 at 512 px and domain-adaptive medical pretraining when the GPU lands — beating specialist baselines on their own turf.

Newton / HMAI · a sovereign self-hosted mind that keeps learning · visual cortex Stage 0. Backbone: DINOv2 ViT-S/14 (Meta, self-supervised), run on candle in Newton's own Rust engine, CPU-only. Generation ≠ perception. ← back to Newton