A cognitive system, not a product
Shypz is not a chatbot. Not an API wrapper. Not another language model with a name. It is the internal cognitive architecture — designated Shypz Architecture — that defines how learning, memory and answering are organized within a single unified system.
It is not finished, and it does not claim to be. It is adjusted as it is measured, and an adjustment stays only when the measurements say the system got better.
The architecture, adjusted
Since the first version of this page, the architecture was adjusted, and this page now describes it more simply: one unified system, inspired by the brain but not a copy of one. How it works inside is not published here. What it does is.
A single unified system, not separate layers and parts. Its behavior comes from how it is organized, not from imitating anything.
It learns from conversation and reading as it runs, and consolidates what it learns along the way.
It can stay silent instead of answering. Every test counts correct, wrong and silent separately: a guess is never hidden, and a silence is never counted as a hit.
When a task needs more than it has, it builds what is missing on its own, while it runs.
Capacity comes from memory and compute. Where a limit of the brain has a purpose, the purpose is kept in a machine form, without the limit.
Not built for one task. The same architecture is measured on real text, on rule tasks and on public logic benchmarks.
Every adjustment is checked in the same order: real data first, then rule tests, then public benchmarks. It stays only when the measurements say it is better.
Generalization on the unseen
GOTU — Generalization on the Unseen (Abbe, Bengio, Lotfi & Rizk, ICML 2023) — is a public benchmark of logic functions. Training sees only part of the inputs; the test asks about the part that was never seen. There, the training data alone does not decide the answer, so what a learner says shows its own bias. The paper finds that networks trained by gradient descent lean toward the simplest function that fits what they saw, more or less depending on the model and the learning rate; later work shows it also depends on how the data is encoded.
So the point here is not that networks fail and this does not. It is what this architecture does on its own terms: it sees each training example once, and it may stay silent. The runs use the data of the public GOTU code: the same functions, the same seen part, the same sizes and the same random seeds. Before and now are the same architecture. Between the two runs it was only adjusted, and it is better now.
These numbers will not stay valid. The architecture keeps being adjusted, and this table will be replaced by newer runs.
| Task | Before · 15 Novcorrect / wrong / silent | Now · 16 Novcorrect / wrong / silent |
|---|---|---|
| lengthgen 6–10Parity of 15 bits. Training sees only inputs with at most 6 to 10 flipped bits. | 0.24–0.43 / 0.57–0.76 / 0 | 1.00 / 0 / 0 |
| full16parity 3k–11kParity of 16 bits, from 3,000 to 11,000 examples. Scored on all 65,536 inputs. | 0.24–0.42 / 0.24–0.42 / 0.17–0.51 | 1.00 / 0 / 0 |
| maj3dim40freeze2Majority of 3 bits. A quarter of the inputs never seen. | 1.00 / 0 / 0 | 1.00 / 0 / 0 |
| 2parity · 2parity40Product of 2 bits, among 15 and among 40 inputs. A quarter never seen. | 0 / 1.00 / 0 | 0 / 1.00 / 0 |
| threesymA function of 3 bits with 4 values. Half of the inputs never seen. | 0 / 0 / 1.00 | 0 / 0 / 1.00 |
| cyclic3dim15Sum of 15 cyclic products of 3 bits (14 values). An eighth never seen. | 0 / 0 / 1.00silent everywhere, even on seen inputs | 0.89 / 0.01 / 0.10seed 1 |
| rfexampleTwo products, of 3 and of 4 bits. Half of the inputs never seen. | 0 / 0.50 / 0.50 | not re-measured yet |
- How to read it. Share of the inputs never seen in training that got a correct, a wrong or a silent answer. full16parity has no unseen part, so it is scored on all inputs, and it runs without the paper's curriculum. Ranges span the tasks of a row; values are means over seeds 1–3 (before: cyclic3dim15 seeds 1–2, rfexample seed 1; now: cyclic3dim15 seed 1).
- Parity. From chance or worse to 100% correct on inputs it never saw, in all 30 runs (10 tasks × 3 seeds).
- threesym. On the unseen half it does not guess: nothing wrong, all silent.
- 2parity. Not solved yet. On the unseen corner it gives the same answer as the simplest function that fits the seen part (the paper's minimal-degree interpolator), and it does not yet know that it does not know.
- cyclic3dim15. Before, it gave no answer at all, not even on inputs it had seen. Now it answers 89% of the unseen part correctly, with 1% wrong.
- rfexample. Being re-measured with the adjustments; not in this snapshot.
What Shypz is not
- Not a framework. Shypz Architecture is not a toolkit to build on — it is the internal structural organization of a cognitive system. You don't install it. It runs.
- Not a wrapper. There is no underlying third-party model being orchestrated. Everything in it is native to the architecture.
- Not a next-word predictor. It is never trained to predict the next word. It is measured by its answers.
- Not a prompt technique. What it does happens inside the system, not in visible text chains.
- Not incremental. This is not an improved transformer. It is a different structural paradigm for machine cognition.
- Not finished. It is still being adjusted. The numbers on this page are a snapshot and will change.
- Not launched yet. This page exists only to describe what is being built. It will be removed at launch.
The architecture exists
The architecture was adjusted, and the measurements say it is better. It keeps learning, and it keeps being measured. Deployment comes after that.
This page is temporary. It exists to communicate what is being built before the system becomes publicly accessible. When Shypz launches, this page will be replaced by the product itself.
No demos. No waitlists. No hype. Just architecture — and its numbers.