Skip to Main Content

We test ideas until they hold.

Spetia SI Labs is Spetia’s foundation model lab. We do foundational and experimental research on how models learn, and turn what holds up into models and products, for Spetia’s businesses and as products of our own.

FoundationalHow models learn, from pretraining to fine-tuning, and where their abilities come from.

ExperimentalNew methods and systems, each tested against a control before we trust it.

AppliedModels and products for Spetia’s businesses, and products of our own.

Foundational and experimental research

Foundational research asks how models learn and why they get better. Experimental research tries new methods, architectures and systems, and keeps only what survives a fair test: a matched control, several seeds, and the uncertainty written down.

  • Foundation models

    Active

    How do models learn what they learn?

    Pretraining, training dynamics and scale: where a model’s abilities come from, and what each stage of training actually adds.

  • Experimental rigor

    Active

    Is the improvement real?

    Matched controls, paired seeds and an interval on every claim, so a lucky run is never mistaken for progress.

  • Experimental intelligence

    Exploring

    Can the experiment run itself?

    Systems that propose the next experiment, run it against a control and report what it settled, with people approving every step.

Built for Spetia, and beyond it

The lab adds to what the Spetia group already builds. Its research becomes models and tools for Spetia’s own businesses, and separate products of our own.

For Spetia’s businesses

Models and tools for the work the group already does: digital engineering for buildings, industrial plants and data centers at Spetia Engineering, and construction and design at Studio Spetia.

Products of our own

Separate products for teams anywhere that train or rely on models, built on the same research. Nothing has shipped yet. When something does, it will be announced here.

Papers, published in the open

Our papers will be published as preprints on print.spetia.com, the lab’s own preprint server. Each one will carry its method, its uncertainty, and enough detail for anyone to check the result.

On the bench right now

The open questions we are testing. When one is settled, the answer goes up on print.spetia.com, even when the answer is no.

  • Is a one-seed win ever enough?

    Plenty of decisions rest on a single training run. We are measuring how often those wins survive a rerun.

    Experimental rigor

  • Which settings does a model actually care about?

    Learning rate, batch size, warmup: a few matter a great deal and most barely move the result. Knowing which saves weeks of tuning.

    Foundation models

  • How much compute goes to questions already answered?

    Teams rerun experiments that someone has already settled. We want to know how much that costs.

    Experimental rigor

  • Can a model choose its own next experiment?

    A system that proposes the next run, tests it against a control and reports what it found, with a person approving each step.

    Experimental intelligence

Toward experimental intelligence

Most progress in AI rests on an experiment someone ran once. We are working toward intelligence that runs the experiment properly: it sets the control, reads the result, and knows when it does not know.

How we work

  • Show the working

    Every result comes with its uncertainty and the evidence behind it.

  • Negative results count

    A settled “no” saves the next team a month of compute.

  • Run it twice

    If it cannot be reproduced, it is an anecdote, not a finding.

  • People decide

    Our tools advise and propose. People approve.

  • Spend compute carefully

    The cheapest experiment is the one you no longer need to run.

  • Publish in the open

    Methods, uncertainty and the runs that did not work go into the paper too.

Bring us a question

Have a result you’re not sure about, or an idea you’re willing to experiment on? Bring it to the lab. A researcher reads every one.

I want to know if changes on .

Or write to contact@spetia.com in your own words.