X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=README.txt;h=b34b5ea6617210c28a440fc49e414ff274ca7be1;hb=239a52ec7face6fcd4515916e80813702fbdf49b;hp=bf6180a1c5b3edbacc5da38493528b921bdcd871;hpb=45f841397dd6f8fb163f2c5b0793cdc250072c73;p=culture.git

diff --git a/README.txt b/README.txt
index bf6180a..b34b5ea 100644
--- a/README.txt
+++ b/README.txt
@@ -1,16 +1,64 @@
 
-For the stack experiment:
+[This file may describe an older version than the current code]
 
-./main.py --task=stack
+Trying to make GPTs build their own "culture".
 
-Takes ~1h10min on a 4090.
+Francois Fleuret
+Jun 21st, 2024
 
-For the arithmetic expressions experiments
+* Motivation
 
-# 38M parameters / 250k samples
+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.
 
-./main.py --task=expr
+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.
 
-# 352M parameters / 2.5M samples
+* Setup
 
-./main.py --task=expr --nb_blocks=48 --result_dir=results_expr_48b_d1024_2.5M --dim_model=1024 --nb_train_samples=2500000 --learning_rate_schedule="1:2e-5,3:4e-6"
+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.
+
+There are 5 competing GPTs.
+
+The "world" is a 6x8 grid with three "birds" moving in a straight line
+and bouncing on the world's borders. It could be another "world", but
+this one has objectness and motion. There are ten colors and 4
+directions of motions, so roughly (6x8x4x10)**3 ~ 7e9 states.
+
+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. The starting and ending states are chosen, by
+rejection, so that there is no occlusion.
+
+My home-baked GPT-37M trained with 250k solves this with ~99% success
+[to be verified with the new setup].
+
+At every iteration, we select the GPT with the lowest test accuracy,
+and run one epoch.
+
+* Creating new quizzes
+
+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.
+
+We update the test accuracy of all the GPTs, and then we go to the
+next iteration.
+
+The hope is that interesting concepts emerge (connectivity, symmetry,
+interior/exterior, shape vocabulary, etc.)