Sundai Hack #133 · Autolab × Maritime

A model that
grows while it learns.

Instead of building a network at full size on day one, we let it add parameters as it trains — and let an AI agent work out the best way to do it.

training time model size
The motivation

Every model you have trained was born at its final size.

Scaling laws tell you how many parameters to buy. They quietly assume you keep all of them, the whole time.

HOW MODELS ARE BUILT TODAY Pick a size from a scaling law Train at that size for every single step WHAT WE ASK INSTEAD What if the size is allowed to change as it learns?
The idea

Training cost is the area under the size curve.

full cost, every step parameters training time
Constant size — the conventional way.
less area, same finish parameters training time
Growing size — both models end identical.

A model early in training is learning word frequencies and basic grammar. Paying for a large model from step zero buys capacity long before there is anything to use it for.

How we built it · the two platforms

One agent runs the research. The other puts it in front of you.

AUTOLAB — runs the science Research agent invents a growth plan, reads the score, decides Our machine trains the models and returns one number plan result It ran 8 experiments on its own while we built the site. No human chose what to try after the first one.
MARITIME — shows the world Always-on machine serves the live results page, day and night Anyone, anywhere watches the experiments as they finish Our laptop can close. The results stay up. That is what makes this a demo and not a folder.
The experiment · what we actually compare
BASELINE full size from step 0 OURS grows 1.2M → 5.9M HELD IDENTICAL FOR BOTH ◆ the same compute budget, counted in operations ◆ the same data, in the same order — WikiText-103 ◆ the same final architecture at the finish line ◆ 3 models trained per side, to separate signal from luck Measured by validation loss — how surprised the model is by text it has never seen.

Only one thing differs: whether the model is allowed to grow.

Baseline gets
2,433
training steps
Ours gets
2,908
training steps, same budget

Being small early buys 19% more training for the same money. That is the whole mechanism.

Result
Both lines stop at the same compute budget. Lower is better.

Ahead from the start, and never overtaken.

Every growing model beat every baseline. The gap holds even after we gave the baseline its own better learning rate.

Baseline, tuned
4.51
± 0.015
Growing
4.11
± 0.053
Gap
0.40
nats, no overlap
Then we tried to break our own result
FIRST NUMBER WE GOT 0.91 nats better  ·  “102× more stable” WHAT SURVIVED THE AUDIT 0.40 nats better, still no overlap …and one claim died outright an unconstrained 1.2M flat model scores 3.905 — better than either
We ran the checks we had only named, and reported what came back.

Two of our own claims did not hold.

The baseline was mis-tuned. It ran at a learning rate chosen for the growing model. Given its own, it became more stable than ours — and the gap halved.

The finish line was too big. We pinned both arms to a size roughly 10× above what this budget wants. Remove that and a plain small model wins.

Growth wins when the target size is fixed — not in general.

What the agent discovered on its own
Every point is one full experiment the agent designed, ran and scored.

When a model grows matters far more than how finely it grows.

Growing too early and growing too late were both worse. Splitting growth into more, smaller steps changed nothing.

What won was finishing growth about a third of the way in, and alternating between wider and deeper rather than doing all the widening first.

It then stopped and told us the remaining differences were smaller than the noise between runs.

Where this goes

Three things we would do next.

grow here

Let the model choose the moment

Growth runs on a fixed timetable today. It should fire when progress stalls — the sign that the current size is used up.

wider deeper ?

Let it pick wider or deeper

The agent showed alternating beats widening first. A model could try both and keep whichever helps more right then.

today useful scale

Prove it at a real scale

These are small models trained briefly. If growth still wins on a model people would actually use, it changes what a training budget buys.

GrowLab

The best size
is not a number.
It is a shape.

Everything below is live right now — the agent is still searching, and the results page updates as it goes.

Live results · hosted on Maritimeapi.maritime.sh The agent's search · Autolabapp.autolab.ai/projects/qsimeon/growlab Code · run it yourselfgithub.com/qsimeon/growlab