Active research  ·  Research program

CIR

Cost efficient intelligence research.

CIR investigates architectures and learning systems that may lower total cost. The aim is equivalent model capability for less economic cost.

Status

CIR is active research. No benchmark results have been published yet. Current experimental architectures are not the final CIR architecture.

01 · Research question

Can a different architecture reach the same capability for substantially less total cost?

The comparison point is a strong Transformer system. Q stands for matched useful model capability.

CIR searches for designs that make the ratio below meaningfully smaller than one.

minimize  TotalCost(CIR, Q) / TotalCost(Strong Transformer, Q)

A ratio of 0.5 would mean equal capability at half the total cost. This is the objective, not a result.

02 · Cost to capability

Proxies are useful. None of them is the cost.

CIR does not optimize any single proxy on its own. It asks a more economically meaningful question. How much total cost does a given level of capability require?

  • FLOPs

    Ignores memory traffic and utilization

  • Token throughput

    Says little about capability reached

  • Parameter count

    Active compute can differ greatly

  • Bits per byte

    One view of capability, not all of it

  • CPU time

    Depends on hardware and implementation

  • Inference speed

    Leaves out training cost entirely
03 · Method

How CIR evaluates a candidate.

Baseline

Why the Transformer

It is the strongest, most studied and most optimized design available. Beating a weak baseline proves little.

Metric

Why BPB alone is insufficient

Bits per byte measures compression of text. Two models with equal BPB can differ in recall, reasoning and long context use.

Capability

Capability is multidimensional

Candidates are compared across several capability dimensions. A gain on one dimension cannot hide a loss on another.

Hardware

Hardware aware evaluation

Cost is measured on real CPUs and GPUs. A design that saves FLOPs but stalls on memory can cost more.

Rigor

Matched and fair comparisons

Baselines get the same tuning effort and budget. Results are checked across seeds and scales before any claim.

Honesty

Measured, estimated, projected

These three kinds of numbers stay clearly labeled. Unexpected failures are recorded, not hidden.

04 · Areas investigated

Where CIR looks for cost.

  • 01Neural architecture

    Model structure

  • 02Recurrent and state mechanisms

    Model structure

  • 03Attention

    Model structure

  • 04Memory

    Model structure

  • 05Learning efficiency

    Learning

  • 06Optimizer interaction

    Learning

  • 07Parameter activity

    Learning

  • 08Data efficiency

    Learning

  • 09CPU and GPU hardware behavior

    Systems

  • 10Memory traffic

    Systems

  • 11Training systems

    Systems

  • 12Capability evaluation

    Evaluation
05 · Scope of claims

What CIR does not claim.

CIR is part of DotrixAI, not the whole lab. Other programs may follow.

  • It has not replaced the Transformer.

  • It is not proven at frontier scale.

  • It claims no fixed cost multiple over Transformers.

  • It claims no proven reasoning advantage.

  • It makes no claim of outside adoption.

  • Today's experimental designs are not the final architecture.

Detailed benchmark evidence may be published as the program matures.

Our evidence standard All research