cd /news/artificial-intelligence/same-net-same-seed-100-vs-50-on-unse… · home topics artificial-intelligence article
[ARTICLE · art-110912] src=structured-intelligence.grok.me ↗ pub= topic=artificial-intelligence verified=true sentiment=· neutral

Same net, same seed – 100% vs. 50% on unseen magnitudes

Structured Intelligence, a company with a patent-pending encoding method, reports that its structured place-value input encoding achieves 100.0% accuracy on unseen magnitudes in Task A, compared to 52.7% for one-hot and 50.0% for learned embeddings, using identical networks, data, and random seeds. The company claims its method generalizes to completely unseen inputs without retraining, citing a held-out comparison where structured encoding reaches 98.6% unseen accuracy versus 0.65% for one-hot and 0.9% for learned embeddings. Structured Intelligence is offering a public benchmark, API access, and licensing under NDA, with Dr. Jarrod S. Segura as the contact.

read3 min views2 publishedAug 25, 2026
Same net, same seed – 100% vs. 50% on unseen magnitudes
Image: source

Structured Intelligence · patent pending

Billionaire-backed embeddings buy a slot per token. Ask a magnitude the net never saw and the slot is chance. Place value already has the column. Identical network. Identical data. Identical random seeds. The sole variable is the input encoding.

Task A · unseen

Structured

100.0%

train 100.0% · h=48

Task A · unseen

One-hot

52.7%

train 99.2% · h=16

Task A · unseen

Embedding

50.0%

train 100.0% · h=16

01

Run it

Same net, same seed. Only the encoding changes. Unseen magnitudes expose the slot. You do not take our word.

02

Try it Sign in, mint a key, POST integers. Public place-value rungs. No charge to try.

03

Buy it

License the API, purchase, or stake. The composed family stays under NDA. You are not buying a PDF.

Product factory

Apps, not source #

Counsel holds the closed family. The factory ships products that sit on it: a saturation clock, success mirrors, a cockpit. Clients pay for the run. They never see the codes.

Open the factory Live miniature · seed 7

Task A — is this reading odd? #

Taught on 0–127. Asked on 512–1,023 — never seen. Same logistic readout. The only change is the input encoding.

Structured

Train 0–127

Unseen 512–1,023

One-hot

Train 0–127

Unseen 512–1,023

Embedding

Train 0–127

Unseen 512–1,023

A slot that was never trained holds whatever it was born with.

Full protocol · measured

Held-out comparison #

Both structured and one-hot reach 100% on the training set. Training-set performance is not the claim. Generalization to completely unseen inputs is. That is why you pick this over a table with a billion-dollar budget.

Encoding Seen Unseen Parameters
Structured (fixed place value) 100.0% 98.6% 1,673
Learned embedding 37.1% → 100% 1.7% → 0.9% 3,721 → 51,721
One-hot 100.0% 0.65% 33,417

Embedding seen accuracy rises only after capacity is grown; unseen accuracy falls as parameters increase.

Task A · measured

Odd reading, unseen magnitude

  • Structuredunseen 100.0%
  • One-hotunseen 52.7%
  • Embeddingunseen 50.0%

Task B · measured

Add two readings, 32× jump

Taught 0–15. Asked 128–511. Composed rungs trained nothing new — same primitives, one column wider.

  • COMPOSED (place-value rungs)100.0%
  • Structured, monolithic0.0%
- One-hot, monolithic0.0%

Live compose check on this page: 200 of 200 (place-value add, no retraining).

For a training run

Train it to know. Not to guess. #

Today you train a model to guess. Show it enough examples and hope the next one looks like the last. When the next one is a size it rarely saw, it guesses — that is the 50% on this page. Chance in a nice jacket.

This is training to know. Unseen is still a thing. Same net, same seeds: 100% on a number it never trained. They guess. This knows.

What the rivals need #

Atoms in the observable universe: ~10⁸⁰. A place-value code needs 12, 16, 32, 64, and 256 inputs.

Range Values One-hot slots Place-value inputs
12-bit 4,096 2¹² 12
16-bit 65,536 2¹⁶ 16
32-bit 4.29 billion 2³² 32
64-bit 1.8×10¹⁹ 2⁶⁴ 64
256-bit 1.16×10⁷⁷ 2²⁵⁶ 256

Try. Then buy. #

The result is public. The family is not. Run the bench, mint a research key, POST integers. If you want the API in production, a license, purchase, or a stake — that is a named conversation, not a download.

Buy · named evaluator

License, purchase, or stake

Commercial use and the full encoder. NDA. You are buying a project that runs on the codes, not the codes.

Email Dr. Jarrod S. Segura

Dr. Jarrod S. Segura

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @structured intelligence 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/same-net-same-seed-1…] indexed:0 read:3min 2026-08-25 ·