The problem

Sales decisions rest on three questions: which company to pursue, why now, and what evidence supports it. A useful answer is specific to the buyer’s requirements, grounded in recent signals, and backed by sources a person can check. A fluent answer that cannot be verified is not useful.

A decentralized research network

Contributors improve our sales harness in parallel using their own compute. Each change is evaluated on tasks the contributor has not seen, and changes that hold up become part of the shared harness.

Independently funded, parallel experimentation supplies both harness improvements and training data, so model development moves faster and at lower cost.

Figure 1. Contributors improve our sales harness, evaluation keeps what holds up, and every run leaves a trajectory for future model training.

How results are evaluated

Each result is judged on what matters in a real sales decision: whether the company fits the request, whether the signal behind it is timely, and whether the supporting evidence can be confirmed at its source. We also measure what each result costs to produce.

Evaluation tasks stay hidden until results are in, so improvements reflect general capability rather than familiarity with the test.

Trajectories

Every experimental run produces a trajectory: a record of the task, each step taken, the tools used, the evidence found, and how the result was judged. A trajectory shows not only whether an answer was right, but how it was reached.

Figure 2. An illustrative trajectory: the request, the research steps, the evidence cited, and the judgment. Sources are placeholders.

Evaluated trajectories are the training data we are building toward: examples of good sales research, and of what went wrong when it fell short.

Open questions

  • Grounding. How can every claim a model makes be traced to a source a person can verify?
  • Timing. How can a model recognize when a signal makes a company worth contacting now?
  • Efficiency. How can good decisions be reached with fewer, cheaper steps?
  • Learning from experience. How can evaluated trajectories train models that reason better than the system that produced them?

Why competition

Shared benchmarks have accelerated AI before. ImageNet united researchers around one evaluation, and the best top-5 classification error fell from 28.2% in 2010 to 6.7% in 2014.[1] We are bringing that competitive spirit to sales reasoning.

Figure 3. Winning ImageNet top-5 classification error by year, for entries trained on the provided data.[1]

Get involved

Contribute compute and improvements to our sales harness, or join the list for early access to our models.

Start contributingRequest early access

References

  1. Russakovsky, O., Deng, J., Su, H., et al. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision 115, 211–252 (2015). arXiv:1409.0575