Lattice
Thinking without a neural network.
Lattice answers questions from readable facts: no training, no weights, and a reason behind every answer.
Classic AI models store their knowledge in billions of numbers, and every answer has to push a large part of them through the processor. Knowledge and thinking are inseparably mixed.
Lattice separates the two. The knowledge is stored as individual, readable facts. A small program picks exactly the facts each question needs and derives the answer from them.
In numbers
- neural networks, training runs and weights
- 0
- facts in the standard model Omni-Compact
- 27M
- questions answered in the test catalogue
- 96.2%
- memory per question with 351 times more knowledge
- ×1.05
How Lattice thinks
01
Every answer is explained
Every answer comes with the full path: which facts were used and how they connect.
02
Multi-step reasoning
Lattice combines several facts into new conclusions and searches specifically for what is still missing for a proof, not for something similar.
03
Understands exceptions
Penguins are birds, but they don't fly. The more specific statement wins; if two statements of equal rank contradict each other, Lattice reports a contradiction instead of guessing.
04
Learns instantly
New knowledge takes effect from the next sentence, with no retraining. What Lattice only assumes is never stored permanently.
05
Makes nothing up
Lattice phrases its answers from fixed sentence templates. Anything that does not follow from the facts is not claimed, and unknown words stay unknown.
06
German and English
Lattice answers in the language of the question and also understands colloquial phrasing.
Intelligence Paging
More knowledge, same effort.
Like an operating system with virtual memory, Lattice only keeps the tiny parts of its knowledge in memory that a question needs right now. The rest stays on disk.
Measured across two and a half orders of magnitude: with 351 times more facts, the number of active facts per question stayed practically the same, and multi-step questions were still answered in full.
Models
A model is a knowledge base that Lattice reads.
Standard
Omni-Compact 1.5
27.3M facts · 96.2% answer rate. The balanced standard model.
Retired
Ares 1.7
37.1M facts · 95.5% answer rate. No longer developed.
Mehr erfahren“An earlier model with 351 million facts answered fewer questions correctly than Omni-Compact with 27 million. Beyond a certain point, more knowledge only adds noise.”
Frequently asked questions
Is Lattice a chatbot?
No. Lattice is a reasoning engine for questions and answers. It understands a limited grammar and builds its answers from sentence templates instead of writing freely, and in return every answer can be traced.
Does Lattice need a graphics card or the cloud?
No. Lattice runs on an ordinary computer and needs only a little memory. Without the optional web lookup, no data leaves your computer.
What happens when Lattice doesn't know something?
Then it says so. For us, invented answers count as errors just as much as missing ones.
How can I try Lattice?
Get in touch with us through the contact page.