The work behind the model

The public conversation about frontier AI focuses almost entirely on the models. The work that goes into training them, especially the expert human work, is treated as invisible infrastructure. It should not be.

RemotebridgeJuly 7, 20264 min read

The public conversation about frontier AI focuses almost entirely on the models. Which model just shipped. Which model is stronger at what. Which model is cheaper per token. The work that goes into producing those models, especially the specific human work of experts correcting, evaluating, and adversarially testing them, is treated as invisible infrastructure. That opacity does not serve anyone.

This piece is a plain description of what the expert-work side of the frontier actually looks like, in practice, from the perspective of a specialist doing it. It is written for professionals who might be considering this work and who want to know what they are being asked into before they sign up.

What you actually do

The exact structure varies by lab and by project, but most expert work falls into a few shapes.

Evaluate model responses. You are shown a prompt in your field and one or more model outputs. You judge them against criteria the lab provides, or against your own criteria if you are asked to define them. You explain your reasoning in writing. This is the largest single category of expert work today.
Author prompts and questions. You write the specific inputs the lab wants to train or test the model on. In medicine, this might be clinical vignettes. In law, contract clauses or case summaries. In mathematics, worked problems. The value is in the depth and realism of what you produce, not the volume.
Adversarially test. You try to elicit specific failure modes: cases where the model gives a confident, plausible answer that is actually wrong. You document what worked and why. This is fast-growing.
Review rubrics. You audit the criteria the lab is using to grade model outputs and flag anything that would let bad reasoning through. This is a leverage role, often assigned to the more senior experts on a project.

The work is asynchronous. You choose when you sit down. Sessions typically run in blocks of two to four hours. Deliverables are documented.

What the day looks like

You log into whatever platform the lab uses. You see a queue of items in your field, at a difficulty level appropriate to your background. You pick one, open it, produce your response, submit. You move to the next. There is no manager on a call. There are quality reviewers checking your output against internal criteria and reaching out if something is unclear, but the interaction is written, not real-time.

Most experts we hear from do this for three to eight hours a week alongside their primary practice. Some do more. Some do less. It is designed to be additive to a working career, not to replace one.

What it pays

Rates are set by the labs and are in USD. They are typically indexed against Western professional scales, not against local labor markets. For an expert in an emerging market, this means the per-hour rate is often several times higher than what the same time would earn in local practice. For an expert in a Western market, it is competitive with consulting.

The rate you see is the rate you receive. There is no cut taken by Remotebridge from your compensation. Our model is a referral payout from the lab on successful placements, reinvested into expanding access. That structure is deliberate.

Where it is honest to have concerns

There are legitimate concerns worth naming.

Some experts wonder if their contributions are training models that will eventually make their own field obsolete. In our experience the fields where models are strongest are also the fields where the demand for expert oversight is highest, not lowest. Capability and evaluation move together.
Some worry about intellectual property. The standard structure is that the specific outputs you produce belong to the lab (as work-for-hire) but the reasoning and expertise remain yours. Any project you take on will be explicit about this in the contracting.
Some ask whether the work is ethically comfortable. Every project has scope, and you can decline any specific project or role. The catalog surfaces what is open; the decision is always yours.

Why we write this

The people layer of AI has been under-explained. That is bad for the field and it is bad for the professionals whose work is currently making it possible. This site exists to shorten the gap between qualified professionals and the roles that are open to them, and to write plainly about the work itself so the decision to enter or not is informed.

If any of this describes you, browse the catalog or sign up and let the platform surface what fits your background. If none of it does, thanks for reading anyway.

transparency how it works

Explore open roles.

Upload your resume once and run an honest fit check against any role in the catalog. Access is free for candidates, always.

July 8, 20263 min read

What happens when a new model ships

Model launches are treated as consumer events. Inside frontier AI labs, they are the moment expert evaluation work explodes. Here is what actually changes, and why the demand for domain expertise ratchets up with every release.

July 8, 20264 min read

What labs mean when they say 'evaluation'

Evaluation is the single most important word in frontier AI work, and the one professionals entering the field understand least. Here is what it actually means, what evaluation work looks like day-to-day, and why it is priced the way it is.

July 10, 20264 min read

How to spot legitimate remote AI work

There is real, well-compensated remote work in AI right now. There is also a growing ecosystem of scams that mimic it. If you are a qualified professional in an emerging market, being able to tell them apart is table stakes.

What you actually do

The exact structure varies by lab and by project, but most expert work falls into a few shapes.

Evaluate model responses. You are shown a prompt in your field and one or more model outputs. You judge them against criteria the lab provides, or against your own criteria if you are asked to define them. You explain your reasoning in writing. This is the largest single category of expert work today.

Author prompts and questions. You write the specific inputs the lab wants to train or test the model on. In medicine, this might be clinical vignettes. In law, contract clauses or case summaries. In mathematics, worked problems. The value is in the depth and realism of what you produce, not the volume.

Adversarially test. You try to elicit specific failure modes: cases where the model gives a confident, plausible answer that is actually wrong. You document what worked and why. This is fast-growing.

Review rubrics. You audit the criteria the lab is using to grade model outputs and flag anything that would let bad reasoning through. This is a leverage role, often assigned to the more senior experts on a project.

The work is asynchronous. You choose when you sit down. Sessions typically run in blocks of two to four hours. Deliverables are documented.

What the day looks like

Most experts we hear from do this for three to eight hours a week alongside their primary practice. Some do more. Some do less. It is designed to be additive to a working career, not to replace one.

What it pays

Where it is honest to have concerns

There are legitimate concerns worth naming.

Some experts wonder if their contributions are training models that will eventually make their own field obsolete. In our experience the fields where models are strongest are also the fields where the demand for expert oversight is highest, not lowest. Capability and evaluation move together.

Some worry about intellectual property. The standard structure is that the specific outputs you produce belong to the lab (as work-for-hire) but the reasoning and expertise remain yours. Any project you take on will be explicit about this in the contracting.

Some ask whether the work is ethically comfortable. Every project has scope, and you can decline any specific project or role. The catalog surfaces what is open; the decision is always yours.

Why we write this

If any of this describes you, browse the catalog or sign up and let the platform surface what fits your background. If none of it does, thanks for reading anyway.