diff --git a/docs/learning.md b/docs/learning.md
index aac132e..2d37481 100644
--- a/docs/learning.md
+++ b/docs/learning.md
@@ -38,3 +38,87 @@ Learning objects are instantiated using the following arguments:
   only sensible if the labeling agrees with the given Reference instance. (Same level of
   theory, same basis set, grid settings, ... ).
 
+
+<figure markdown="span">
+  ![Image title](wandb.png){ width="900" }
+  <figcaption> Illustration of what the Weights & biases logging looks like.
+  The graph on top simply shows the force RMSE on each data point versus a unique
+    'identifier' per data point. The bottom plot shows the same data points, but now
+    grouped according to which walker generated them. In this case, walkers were sorted
+    according to temperature (lower walker index were lower temperature), and this is seen
+    in the fact that walkers with a higher index generated data with on average higher errors,
+  as they explored more out-of-equilibrium configurations.</figcaption>
+</figure>
+
+
+The core business of a `Learning` instance is the following sequence of operations:
+
+1. use walkers in a `sample()` call to generate atomic geometries
+2. evaluate those atomic geometries with the provided reference to obtain QM energy and
+   forces
+3. include those geometries to the training data, or discard them if they exceed
+   `error_thresholds_for_discard`. Reset walkers if they exceed
+   `error_thresholds_for_reset`.
+4. Train the model using the new data.
+5. Compute metrics for the trained model across the new dataset and optionally log them to
+   W&B.
+
+Currently, there are two variants of this implemented: passive and active learning.
+
+## passive learning
+
+During passive learning, walkers are propagated using an external and 'fixed' Hamiltonian
+which is not trained at any point (e.g. a pre-trained universal potential or a
+hessian-based Hamiltonian).
+
+```py
+model, walkers = learning.passive_learning(
+    model,
+    walkers,
+    hamiltonian=MACEHamiltonian.mace_mp0(),     # fixed hamiltonian
+    steps=20000,
+    step=2000,
+    **optional_sampling_kwargs,
+)
+```
+Walkers are propagated for a total of 20,000 steps, and samples are drawn every 2,000
+steps which are QM evaluated by the reference and added to the training data.
+If the walkers contain bias contributions, their total hamiltonian is simply the sum of
+the existing bias contributions and the hamiltonian given to the `passive_learning()`
+call.
+Additional keyword arguments to this function are passed directly into the sample function (e.g. for
+specifying the log level or the center-of-mass behavior). 
+
+The returned model is the one trained on all data generated in the `passive_learning()` call as well as all data which was already present in the learning instance (for example if it had been initialized with `initial_data`, see above).
+The returned walkers are identical to the ones passed into the method, but this is done to
+emphasize that internally, they do change due to calling `passive_learning` (because they
+are either propagated or reset, or their metadynamics bias has changed because there are
+more hills present than before).
+
+## active learning
+
+During active learning, walkers are propagated with a Hamiltonian generated using the
+current model. They are propagated for a given number of steps after which their final
+state is passed into the reference for correct labeling.
+Different from passive learning, active learning *does not allow for subsampling of the
+trajectories of the walkers*. The idea behind this is that if you wish to propagate the
+walker for 10 ps, and sample a structure every 1 ps to let each walker generate 10 states,
+it is likely much better to instead increase the number of walkers (to cover more regions
+in phase space) and propagate them in steps of 1 ps. Active learning is ideally suited for
+massively parallel workflows (maximal number of walkers, with minimal sampling time per
+walker) and we encourage users to exploit this.
+
+```py
+model, walkers = learning.active_learning(
+    model,                      # used to generate hamiltonian
+    walkers,      
+    steps=2000,                 # no more 'step' argument!
+    **optional_sampling_kwargs,
+)
+```
+## restarting a run
+
+`Learning` has first-class support for restarted runs -- simply resubmit your calculation!
+It will detect whether or not the corresponding output folder has already fully logged the
+each of the iterations, and if so, load the final state of the model, the walkers, and the
+learning instance without actually doing any calculations.
diff --git a/docs/wandb.png b/docs/wandb.png
new file mode 100644
index 0000000..896c4b8
Binary files /dev/null and b/docs/wandb.png differ