about
The Open-World Assumption states that agents should expect novel, unanticipated situations to arise over time. Despite this, the current industry standard is to pretrain models on a wide data distribution, learn a policy that generalizes to the data at hand, and expect it to be able to perform long horizon tasks. We believe that this approach will fail to produce economically viable agents.
We are exploring a different direction: pushing energy-based models in a way that is compatible with continual learning, without giving up the practical advantages of modern large-scale training.
Concretely, methods exist that can be bootstrapped on top of existing pretrained models (transformers, SSMs, etc.) to turn them into systems that can update themselves over time, instead of being frozen snapshots. This tackles a problem that should be obvious: Every agent must assume that it's current knowledge is incomplete and future observations may violate training assumptions.
jobs
(email your resume to hello@basiliskeng.com)
Research Scientist
Drive a focused research agenda on turning pretrained sequence models into continually learning systems. You will take internal prototypes that already work in small settings, scale them up, and run clean experiments that inform where the approach breaks and how to extend it.
Requirements: Strong background in machine learning (PhD or equivalent), hands-on experience training and evaluating modern models, familiarity with energy-based models and continual learning, and the ability to write clear, reliable research code that others can build on. Computational neuroscience background preferred, especially experience with predictive coding, associative memory, and complementary learning systems.
Research Engineer
Build and operate the infrastructure that lets the research team move quickly: training pipelines, experiment tracking, evaluation harnesses, and lightweight serving for models under active development. You will work closely with research scientists and also contribute directly to method implementations and diagnostics.
Requirements: Strong software engineering in Python (PyTorch experience required), experience with distributed training and ML infrastructure, solid DevOps/observability skills, and interest in contributing directly to research code as well as the surrounding systems.
blog
coming soon