WEBVTT

00:00:01.290 --> 00:00:02.345
Hi, I'm Cypress!

00:00:02.345 --> 00:00:06.566
Poolside's Laguna report ships two coding models, but its real subject is

00:00:06.566 --> 00:00:08.325
the factory that built them.

00:00:09.808 --> 00:00:13.222
The idea: treat model building as an industrial process.

00:00:13.222 --> 00:00:17.775
Data, training, evaluation and inference form one versioned loop, and a research

00:00:17.775 --> 00:00:20.051
win ships by flipping a flag.

00:00:22.008 --> 00:00:23.390
Hive generates synthetic data.

00:00:23.390 --> 00:00:26.845
Blender streams the data mix, and Titan trains on it.

00:00:27.265 --> 00:00:32.043
Atlas serves inference and RL rollouts, reading Titan's model definitions bit for bit.

00:00:32.043 --> 00:00:34.615
What you evaluate is what you serve.

00:00:35.035 --> 00:00:40.203
One code execution platform, spanning about a million repositories, supplies synthetic tasks,

00:00:40.203 --> 00:00:43.649
runs the evals and scores the RL rewards.

00:00:44.069 --> 00:00:47.040
Every run is committed code with its own ID.

00:00:47.040 --> 00:00:51.332
So any checkpoint, or even one training token, traces back to its source.

00:00:53.192 --> 00:00:55.065
Even the data mix is automated.

00:00:55.065 --> 00:00:57.875
AutoMixer turns a matter of taste into a search.

00:00:58.296 --> 00:01:02.763
Sample mixtures from more than fifty dataset groups, staying near a hand-made prior.

00:01:03.183 --> 00:01:06.190
Train about sixty small proxy models, one per mixture.

00:01:06.610 --> 00:01:09.382
Score each one on coding, math, knowledge and more.

00:01:09.802 --> 00:01:13.274
Fit a regressor per capability, predicting scores from the mixture.

00:01:13.694 --> 00:01:17.120
Then optimize the mixture, weighting the capabilities you care about.

00:01:18.425 --> 00:01:22.195
The evidence: they built the smaller model from scratch, inception to

00:01:22.195 --> 00:01:23.566
release, in five weeks.

00:01:25.142 --> 00:01:29.714
The report's bet is that iteration speed, not an architecture trick, is the lever.

00:01:29.714 --> 00:01:32.979
The factory's job is to make the second model boring.

00:01:34.525 --> 00:01:38.351
To recap: one versioned loop, research that ships by a flag, and

00:01:38.351 --> 00:01:40.582
a data mix chosen by an optimizer.

00:01:41.002 --> 00:01:43.181
Every source is in the full article.

00:01:43.181 --> 00:01:43.804
I'm Cypress.

00:01:43.804 --> 00:01:44.115
Bye!
