Lab
Experiments, in the open.
We run experiments on new models, methods, and tools, and publish every result with the method and data behind it.
Read the findingsDrag to draw
Findings
Recent results.
- 14 Oct
Can a small model learn UI design from visual feedback?
67.7%of blind judge votes over the untrained model - 29 Sept
Does encrypted compaction use fewer tokens than text summaries?
22%fewer tokens than text summaries - 24 Sept
Can Jev mask personal data better than an LLM?
1.43×GPT-6 Luna’s score on personal-data masking - 30 Sept
Can Jev build structured objects without generating text?
4×faster than GPT-6 Luna at producing structured objects