Open Science

Science in the Open

Multiscalar Intelligence is a research driven company, and we publish what we find. We believe open science is how a field advances: proofs, benchmarks, code and results released where anyone can check them, including the results that did not go our way. Our contributions begin with the foundations the rest of our work depends on, mathematics, mechanism design and economics, and extend to the multi-agent systems built on top of them.

We are opening a community of top tier researchers and engineers to collaborate on scientific work. If you would like to be part of it, write to hello@multiscalar.ai with your research interests and CV.

July 2026 Open Problems

Six More Erdős Problems, in Five Days

Found by GPT-5.6 under the Multiscalar research prompt

Erdős problem 690 took months of harness engineering. These six took five days, and almost all of the work went into the prompt. Each one states exactly what a full proof has to establish, names the traps, and sends adversarial agents at every candidate argument. We attempted about thirteen problems, so a little under half worked out. Every proof is posted on the erdosproblems.com forum for public attack, and the prompt that produced it is published next to it.

Proposed solutions, published for attack rather than certified. Source, prompts and Lean formalisations are at github.com/ShouqiaoW/erdos.

May 2026 Number Theory

A Complete Answer to Erdős Problem 690

Discovered by the Multiscalar Fields System

We prove that the natural density dk(p), of integers whose k-th smallest prime divisor is p, is not unimodal for every k ≥ 4, completing Erdős' classification. The proof was discovered by the Multiscalar Fields System with limited human interaction.

Read the proof →
June 2026 Neural Compression

Neural Compression of Satellite Imagery

Discovered by Multiscalar Dynamo

A Sentinel-2 earth-observation tile compressed by our learned codec with no visible loss. Drag the slider across rate points. Even at the highest ratio, the reconstruction stays visually identical to the original.

compressed satellite tile