https://t.co/13kLGu4zGA
We are also actively looking for researchers, engineers, and managers who want to help us in this journey. Our goal is to create inherently safe, adaptable, and aligned models, entirely through mechanistic and algorithmic interventions.
Our second blog post, focused on recovering lost performance in image based ResNet18 models on image recognition datasets.
This is another price of evidence toward our belief that continual learning can be solved via simple, lightweight interventions.
A new technical preprint focused on leveraging transport keys for ResNet-18 and ViT style models. We see up to 90% recovery on basic image classification tasks such as CIFAR-10 when we train a simple image model on a set of sequential tasks and induce catastrophic forgetting.
Our core thesis is centered around building adaptable AI models, which we define as models that can automatically align to production needs. We are heavily focused on safety, optimization, and continual adaptability. You can learn more on our main website: https://t.co/uJwEusaUFK
Continual learning is perhaps the largest open problem in AI today.
We are unveiling some early results on a new research technique intended to address catastrophic forgetting on small scale language modeling and agentic tasks by aligning model internals: https://t.co/etjdyzgNbF
We were inspired by concepts centered around model stitching and interpretability. Our initial implementation of transport keys was centered around recovering performance in ResNet-18 style models and ViTs on CIFAR-10 style datasets. We plan on releasing this paper soon!