X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;ds=sidebyside;f=README.txt;h=af96ee9d73173cc05be312e3e4d0d95882a8f8ea;hb=4ec52fe66419a6e1d2b231108ccbb45902395fcc;hp=d4740f386def252fb3b3229bcb97ba95b40fb3dc;hpb=59600257e0eda86816a43676c5ffbe598d78bdb5;p=culture.git

diff --git a/README.txt b/README.txt
index d4740f3..af96ee9 100644
--- a/README.txt
+++ b/README.txt
@@ -1,23 +1,54 @@
 
-======================================================================
-For the stack experiment:
+Trying to make GPTs build their own "culture".
 
-./main.py --task=stack
+* Motivation
 
-Takes ~1h10min on a 4090.
+The original motivation of this experiment is the hypothesis that
+high-level cognition emerges from the competition among humans in the
+space of language and ideas.
 
-======================================================================
-For the arithmetic expressions experiments
+More precisely, communicating agents try to out-do competitors by
+creating stuff that is smart but doable, e.g. some other agents get
+it, but not all. Then, that smart thing is added to the "culture",
+they all learn and get to understand it, and it repeats.
 
-# 38M parameters / 250k samples
+* Setup
 
-./main.py --task=expr
+It starts with a "world model" that they got before they communicate,
+and from there, they try to "be smart" by proposing quizzes that can
+be solved but not by everybody.
 
-# 352M parameters / 2.5M samples, reaches 99.80% after 12 epochs, the
-  learning rate schedule is obviously terrible
+There are 5 competing GPTs.
 
-./main.py --task=expr --nb_blocks=48 --dim_model=1024 --nb_train_samples=2500000 --result_dir=results_expr_48b_d1024_2.5M
-======================================================================
-25.07.2023
+The "world" is a 6x8 grid with one or two "birds" moving in a straight
+line and bouncing on the world's borders. The colors correspond to a
+fixed "z-buffer order". It could be another "world", but this one has
+objectness, occlusion, and motion.
 
-./main.py --task=sandbox --nb_train_samples=10000 --nb_test_samples=1000 --nb_blocks=4 --nb_heads=1 --nb_epochs=20
+Given a random world state, and the state after two iterations of
+birds moving, a "quiz" is to predict the second frame, given the
+first, or the opposite.
+
+My home-baked GPT-37M trained with 250k solves this with ~99% success.
+
+At every iteration, we select the GPT with the lowest test accuracy,
+and run one epoch. If its test accuracy got higher than 97.5%, it will
+create new quizzes. To do so, it generates a large number of pairs of
+frames, and checks which ones of these quizzes are hard but not too
+hard, which means
+
+[THIS IS THE IMPORTANT BIT]:
+
+it can be solved, in both time directions, by all the other GPTs **but
+one**
+
+The both time directions is to avoid a simple type of quizzes which is
+simply to deal with noise in the first frame.
+
+The GPT generates 1000 of such quizzes, that are added to the
+"culture", i.e. the training set.
+
+Then training resumes.
+
+The hope is that interesting concepts emerge (connectivity, symmetry,
+interior/exterior, shape vocabulary, etc.)