learning workshop · work in progress
Palimpsest
Keep the thought. Keep the correction.
A local workshop for investigating models, keeping exact observations and following how an idea changes. The notebook works; the larger instrument is still taking shape.
Runnable notebook and separate experiments · learning direction under revision
An explanation can feel clear and still leave you unable to build the thing it describes. Palimpsest asks for a different kind of understanding: change an input, inspect what happens, follow the relationship down to the operations, then use it to make something of your own.
The work should accumulate. A failed attempt, a corrected observation and the reason for the correction are all useful. A polished summary should not rub out the route that produced it.
What did the model actually see?
The working notebook gives that question an inspectable answer. Save a rough note, revise it, and select the particular revision you want to send. Preview shows the assembled context before an explicit model request. The saved receipt separates what you supplied from what the model returned.
That makes a small, useful experiment possible: select an earlier observation instead of its correction, inspect the changed request, and compare the outputs. Keeping a note in the application is different from putting it in context. Both are different from changing model weights.
Question-specific notebooks keep investigations connected without mixing every source into every request. Selected notes can become HTML or Markdown editions with a source bundle, while the original revisions remain available.
One base, a different prediction
A separate experiment applies the intervention idea to DNA. In One base, Evo 2 7B scores the sequence: swap, insert or delete a base, then compare conditional probabilities along the original and edited windows. The standalone viewer source is also available beside this workshop.
This requires its own Evo 2 installation and CUDA hardware. It is not running on this website, and its GPU path was not re-tested for this publication. Changes in predicted sequence probability do not establish a biological effect.
The notebook is a beginning
The first attempt built the notebook, records and atlas before it delivered the desired encounter with a manipulable mechanism. Those are useful tools, but they do not fulfil the whole ambition. Palimpsest is being reconsidered around instruments that let a learner follow a model’s behaviour down to its primitives and compare the consequences of changing them.
That direction remains open. The public source contains the working prototype and separate experiment, not a claim of a finished LLM curriculum or validated teaching method.
Open a notebook
Clone the repository and run:
python3 -m palimpsest serveOpen http://127.0.0.1:8766. The notebook needs only Python’s standard library. Ollama is needed when you explicitly send a request to a local model. Capture, revision history and exports work without it. Setup and the separate experiment’s requirements are in the README.