03 / Programs / Chisei

Efficient Intelligence Program

Named from the Japanese word for intelligence (知性, chisei), this program investigates how much general-purpose capability can be concentrated into the smallest practical computational budget.

Foundational Reading

Before reviewing CHISEI, readers are encouraged to read:

Intelligence Density

Towards Measuring Capability per Unit of Computation

SGT-PUB-2026-0001 · Publication No. 001
Read White Paper

Rationale

The field of artificial intelligence has, for over a decade, operated under an assumption that has become so pervasive it is rarely examined: that capability is primarily a function of scale. Larger models, larger datasets, larger training runs. This assumption has produced impressive results — and it has also produced an extraordinary concentration of research effort around a single variable.

The Efficient Intelligence Program proceeds from a different premise. We ask whether the relationship between parameter count and capability is genuinely linear, or whether structural and methodological choices can achieve disproportionate results at constrained scales. The question is not whether small models can match frontier systems on every benchmark — they cannot. The question is whether we understand the mechanisms of capability well enough to concentrate it deliberately.

This is a scientific question, not an engineering optimisation. It requires the construction of formal measurement frameworks, reproducible methodology, and willingness to publish results that challenge prevailing assumptions — including negative results.

Central Research Question

"How much general-purpose intelligence can be concentrated into the smallest practical computational budget?"

Registered Projects

CHISEI-1.5B-R0

An Experimental Study in Maximum Intelligence Density

Research Preparation
Permanent Identifier
SGT-RP-2026-0001
Program
Efficient Intelligence Program
Revision
R0
Status
Research Preparation
Registered
Last Updated

Research Question

"How much general-purpose intelligence can be concentrated into the smallest practical computational budget?"

Abstract

CHISEI-1.5B-R0 investigates methods for maximizing general-purpose capability within an approximately 1.5-billion-parameter language model. The study proceeds from the premise that the relationship between parameter count and capability is not linear — that significant capability can be unlocked through deliberate architectural choices, curated training data, and systematic optimization, rather than through scale alone. The project name derives from the Japanese word for intelligence (知性, chisei), reflecting the program's commitment to studying the nature of capability itself rather than its quantity. CHISEI-1.5B-R0 represents the initial revision of this experimental line; findings will establish a reproducible baseline for subsequent compact-model studies.

Objectives

  1. 01.

    Develop a reproducible compact-model training methodology that can serve as a reference standard for future Intelligence Density research.

  2. 02.

    Establish an Intelligence Density evaluation framework enabling rigorous measurement of capability per unit of computational budget.

  3. 03.

    Measure capability gains against compute-matched controls to isolate the contribution of architectural and methodological choices.

  4. 04.

    Validate practical deployment on accessible hardware to ensure findings are transferable beyond laboratory conditions.

Note

This page presents program-level context. Project data — including abstracts, objectives, and metadata — is drawn directly from the central research registry. The permanent record for each project is authoritative: SGT-RP-2026-0001