Collaborate with us

Joint research

We welcome work on social intelligence, personality and context modeling, extended interactions, continuous video generation, multimodal control, and evaluation. Start with a specific question and agree on hypotheses, experiments, and milestones.

Describe the limits of current methods, the mechanism you want to test, the resources you need, and how you will assess results. A focused question and a feasible experiment give the collaboration a clear starting point.

Efficient, real-time systems

Real-time experiences require models and systems to be optimized together. Our interests include generation acceleration, caching and scheduling, streaming, multi-GPU coordination, compression and deployment, as well as ways to maintain image quality and behavioral stability under limited resources.

An engineering collaboration should jointly define the hardware configuration, input distribution, how latency is measured, concurrency and quality thresholds. Interim results should include reproducible baselines and controls, to help separate algorithmic gains from configuration changes and infrastructure differences.

Data, evaluation and human feedback

High-quality interaction research depends on real and diverse language, situations and evaluations. We focus on how to organize, clean and annotate data suited to multi-turn dialogue, and on evaluation designs that can reveal differences between models.

A data collaboration should first establish the source of the data, its authorization, how personal information is handled and the scope of use. For work that relies on human feedback, it should also describe the raters’ tasks, how informed consent is arranged, quality control and how the results will be used, so that data whose rights cannot be verified does not enter the process.

Applications and creative projects

AI companions, real-time virtual characters, games and interactive storytelling, film and content creation can all provide concrete settings for applying understanding and generation capabilities. We want to work with teams in these fields to discover which experiences truly need continuous interaction, and which parts need human creative work and oversight.

A proof of concept can start with an experience of limited scope, with participants, content rights, terms of use and stopping criteria clearly defined. Projects involving real people’s voices or likenesses, minors or important decisions need additional assessment before they begin.

How we work together

Outline the problem, prior work, proposed roles, resources, and deliverables. Before work begins, agree on confidentiality, data use, intellectual property, attribution, and publication.

Use milestones to test key assumptions, then decide whether to expand the experiments or integrate the results into a product. Revisit goals, costs, and risks at each stage.

• Problem and context. Describe who it serves, the current limitations and the experience you want to change.
• Evidence and methods. Share related work, prototypes, baselines or reproducible results.
• Resources and boundaries. Describe data authorization, hardware needs, division of roles and constraints.
• Results and evaluation. Agree on the form of deliverables, the criteria for success and what will be made public.

Get in touch

Collaboration inquiries should be submitted through the official channels published by OneZero Labs. An initial introduction can use a non-confidential summary outlining the team’s background, the problem and the proposed approach, along with any validation results that can be shared publicly. A mailing address and submission process will be added to this page once confirmed.

Until both parties have put clear confidentiality and data processing arrangements in place, please do not submit trade secrets, unpublished personal information, account credentials or data you are not authorized to share. The information on this website does not in itself constitute confirmation of a collaboration, a commitment to purchase or a license to any results.