Data for frontier intelligence

Frontier models are no longer limited by how much they can read. They are limited by what they learn from, and by how well anyone can tell whether the learning worked. The gap between a model that answers well and one that finishes hard, long-running work is largely a gap in data: its shape, its grounding, and whether its outcomes can be verified.

Phemeral Data is a team of frontier researchers. Most of us have built agents that run against real systems, and most of us spent earlier years on the less visible work underneath them: collecting, labeling, cleaning, and auditing data at a scale where small errors compound. We are bringing that combined experience to a question we think the field has not explored carefully enough. How should synthetic and real data be built as one system, rather than as two competing sources?

Where the literature leaves a gap

A lot of recent work frames training data as a choice. Either models generate their own curricula, or only human-produced data carries the signal that matters. We think both views miss the mechanism. Synthetic data is cheap and can be aimed at specific weaknesses, but it inherits the blind spots of the model that wrote it. Real data is grounded, but it is slow to collect and thin exactly where a model most needs help. The useful question is not which one wins. It is how each can correct the other's failure modes, and how to know when it is doing so.

How we work

• Ground synthetic data in real environments. Generated tasks should be checkable against something outside the model, such as a running system, a test suite, or a recorded outcome.
• Learn from real agent work. Observed sessions, with their context, tool calls, and failures, are a source of hard problems that are difficult to invent from scratch.
• Measure what a dataset does. We judge data by whether it moves a model in the direction we claim, through controlled comparisons, not by how much of it we have.

Why this matters for AGI

We do not expect general intelligence to arrive from a single training recipe. It is more likely to come from systems that can reliably take on longer, messier work and improve from the record of doing it. Data is the substrate for that improvement. If the loop between generation, observation, and evaluation is built well, each stronger model produces better problems for the next one, instead of running out of problems worth solving.

Get in touch

We are interested in researchers and engineers who have built agents, evaluation harnesses, or data systems, and who care whether the data behind a model is actually good. If you have a dataset, an environment, or a problem we should look at, write to founders@phemeraldata.com.