Primus for X

Primus for Machine Learning

Machine learning is where Primus got started. Describe something you’re curious about in one sentence, and Primus scopes the question, reads the literature, runs experiments on GPUs, and writes the paper.

A Primus project mid-run: research stages on the left, live GPU jobs and budget in the middle, and Primus narrating its work on the right

One sentence is enough

Your prompt is a direction, not a spec. You don’t need a dataset lined up, a model picked out, or a clear idea of which experiment to run. Vague is fine: Primus asks follow-up questions and sharpens the project with you before the work starts.

You don’t need to be an ML engineer either. If you are one, think of it as a team of interns: it does the literature reviews, sets up the experiments and babysits the clusters.

Writing research prompts →

From prompt to paper

01
Scopes the question
Turns your idea into a well-posed question, and checks it with you.
02
Reads the literature
Studies what’s already known, and pins the papers that matter to your project.
03
Runs experiments
Designs them, runs them on GPUs, and iterates on what it learns.
04
Writes the paper
Methods, results, code and citations, including when the hunch doesn’t hold up.

Questions that tend to work

Each of these shapes has produced a finished project. The one habit worth having: aim narrow, not for the moon. Name a specific weakness and improve it without breaking everything else.

Does X cause Y?

“Does the RL clipping mechanism control how diverse the generated outputs are?”

Is that claim true?

“A vendor says their model wins on this metric. Does the win survive a better metric?”

Open question, pick a side

“Does post-training move a model’s representations toward or away from how brains process language?”

Beat the incumbent

“Beat directed evolution at proposing high-fitness protein variants at the same query budget.”

Build the yardstick

“There’s no trusted automatic score for 3D-mesh quality, so build one and show it tracks human judgment.”

Fix a known failure

“Use preference optimization to fix the word-skip and repeat failures TTS models are known for.”

Port it

“Replatform this model onto architecture X and verify it holds up across the key benchmarks.”

Optimize one dimension

“Find the best quantization of model X, judged on X and Y, without losing accuracy on Z.”

Shape a hunch

“Maybe injecting a self-reflection token reduces hallucination, so let’s design something to test it.”

It keeps going, and knows when to stop

Experiments run on GPUs Primus provisions for you. By default it runs in Full Self-Driving, making its own routine sign-offs while you do something else, and you can chat with it at any point to steer: “focus on the smaller model first,” “stop tuning and write it up.” It emails you when it needs you, which is in three situations:

STOP 1