Introducing AutoMuse ChatGPT Working on Temporal Hierarchical Memories for AGI ACE systems_transcript.csv

text,start,duration
hey everyone David Shapiro here with an,0.539,3.601
update,2.94,3.18
um so before we get started I will plug,4.14,5.039
my patreon this will take just a second,6.12,5.58
um I'm able to do AI full time because,9.179,4.92
of the support that I get on patreon so,11.7,4.08
if you uh want to support me jump over,14.099,4.02
support me on patreon also if you want,15.78,4.68
any direct help if you sign up for the,18.119,5.58
higher tiers on patreon I am happy to,20.46,5.579
jump in and give you advice or help you,23.699,4.561
solve problems and this goes as far as,26.039,4.441
looking at your fine tuning data talking,28.26,4.139
about architecture and solving whatever,30.48,4.14
other problems you have,32.399,4.801
um okay so that's it for patreon second,34.62,3.72
is,37.2,4.74
um I want to uh plug the r artificial,38.34,5.879
sentience subreddit so this is a,41.94,4.38
subreddit that I and a few other,44.219,3.481
um thought leaders on cognitive,46.32,3.96
architecture created,47.7,4.199
um and it's not my subreddit,50.28,3.48
specifically someone else created it I'm,51.899,4.561
just a member but the primary thing is,53.76,4.74
creating autonomous cognitive entities,56.46,3.9
we would have picked that but that is,58.5,4.079
too long of a subreddit title so we have,60.36,4.32
artificial sentience,62.579,3.621
um jump in over here talk about,64.68,3.9
artificial cognitive entities artificial,66.2,4.66
sentience and so on,68.58,5.219
um next update is uh for the Raven,70.86,5.04
project so I posted an update I pinned,73.799,3.36
it here,75.9,4.14
um the the tldr is that like I'm super,77.159,4.081
burned out,80.04,4.14
um and so I am slowing down that being,81.24,4.08
said,84.18,3.36
um you know this this project has almost,85.32,4.92
700 stars and there's a huge amount of,87.54,5.52
Interest both in AGI today uh in,90.24,5.46
cognitive architecture and also keeping,93.06,6.599
it open source uh specifically so,95.7,6.54
um if you want to jump in uh jump in,99.659,4.801
here we're still getting organized we're,102.24,4.98
figuring out a consensus policy and I'm,104.46,4.32
hoping that with the correct consensus,107.22,3.48
policy it will be more distributed and,108.78,4.08
more participatory,110.7,4.559
um so you know it is slow that's part of,112.86,5.34
the design though now finally to the,115.259,5.4
topic of today's video if you're a chat,118.2,4.62
GPT user you have probably been,120.659,3.721
frustrated over the last few days,122.82,4.56
because all your chats are gone,124.38,6.84
and what I realized is that uh this is,127.38,5.82
really unfortunate because I've been,131.22,5.58
working on my novel and I use it um I,133.2,5.82
create a new chat every time I'm working,136.8,4.38
on a new chapter and what I do is I go,139.02,3.84
back to the old chapters and say write,141.18,4.559
an executive summary of this chapter and,142.86,4.5
then I can just use that to quickly copy,145.739,4.981
paste it into an into a new chat,147.36,5.34
um so that the uh the new chat can help,150.72,4.86
me it's caught up on the story and goes,152.7,4.679
right along but I'm like okay I can,155.58,4.26
automate that it's the same uh set of,157.379,5.58
procedures every time I use an executive,159.84,5.22
summary so that it knows about the,162.959,3.601
characters and the story and so on and,165.06,4.08
I'm like why don't I just go ahead and,166.56,5.16
work on this,169.14,6.36
um uh offline right I can I can manage,171.72,6.42
my own files and if you watched uh my my,175.5,4.379
next most recent video you know that I'm,178.14,5.7
working on a QA chat bot and so these,179.879,6.121
actually have a lot in common it has to,183.84,4.44
deal with organizing an arbitrarily,186.0,4.62
large amount of data a large Corpus of,188.28,5.16
data and the QA chatbot case it is,190.62,6.479
scientific papers in the auto Muse case,193.44,6.2
the auto Muse chat GPT case it is,197.099,4.681
organizing a story,199.64,5.08
and so but because of the similarity,201.78,4.379
right the the thing that they have in,204.72,4.379
common is it has to summarize and search,206.159,4.681
an arbitrarily large amount of memory,209.099,3.601
and it has to do it automatically in the,210.84,4.02
background so these these projects are,212.7,3.84
actually much more similar than you,214.86,3.599
might think from the perspective of,216.54,5.339
cognitive architecture so since chat GPT,218.459,5.821
is down and I can't really use it it's,221.879,3.78
kind of actually stalled my forward,224.28,3.3
progress on my novel,225.659,4.14
so I'm like all right well why don't I,227.58,4.019
just fix this,229.799,5.281
um so what I have done is let me just go,231.599,5.341
ahead and show you the file show you the,235.08,4.26
repo so here is the um here's what I'm,236.94,4.5
working on with the auto Muse chat GPT,239.34,4.58
project,241.44,2.48
so you might have remembered that I I,243.959,3.901
was working on auto Muse a lot last year,245.76,3.72
and then I stopped because I'm like I,247.86,3.54
don't want to hurt anybody I don't want,249.48,3.78
to you know unemploy my friends or,251.4,4.08
whatever now people are using chat gbt,253.26,4.14
to write like crazy and it's all garbage,255.48,4.8
right so or mostly garbage,257.4,4.38
um so I'm not I'm not worried about that,260.28,4.979
anymore I you know when when all the,261.78,5.76
publicate all the all the Publishers are,265.259,4.741
using AI to detect whether or not it's,267.54,4.56
written by AI or to detect the quality,270.0,3.96
speaking of if there are any Publishers,272.1,4.14
out there and you want to learn how to,273.96,3.66
use,276.24,5.04
um AI to to rapidly like,277.62,5.76
um measure the quality of someone's,281.28,4.74
fiction let me know,283.38,3.36
um,286.02,2.58
because that can help you quickly sift,286.74,4.56
through sift the uh the the the,288.6,5.28
um the grain from the chaff,291.3,6.839
um so anyways point being is there's a,293.88,7.08
few components of this that are that are,298.139,5.221
all the same right so you got the the QA,300.96,4.32
I will be working on this it's been,303.36,3.66
three days I had I was really tired over,305.28,3.06
the weekend,307.02,3.179
um burnout is real,308.34,3.6
um but I am recovering and everyone has,310.199,3.421
been very kind about that so I'm I'm,311.94,3.24
grateful and I said like I need some,313.62,2.94
time everyone's like take your time,315.18,2.34
buddy,316.56,3.18
all right anyways going down a rabbit,317.52,4.92
hole point being is that these projects,319.74,4.679
have a lot in common and so what I'm,322.44,4.259
working on is okay how do I create a,324.419,4.801
memory system that works for both of,326.699,5.401
these right because we humans have one,329.22,5.039
memory system and it works for whatever,332.1,4.02
task that we're doing whether you're,334.259,4.38
remembering you know birthdays or how to,336.12,4.919
do your job or whatever and I okay yes,338.639,5.221
it's not one memory system we have a,341.039,4.621
memory operating system you know the,343.86,3.36
hippocampus and all sorts of other,345.66,3.84
connections and representations and,347.22,4.74
abstractions but the point is is that,349.5,5.46
with the correct format or system of,351.96,5.34
abstractions and representations you we,354.96,3.66
should be able to create a memory system,357.3,3.42
that is good for an arbitrary number of,358.62,5.22
tasks so back in the day,360.72,5.18
there was this concept called temporal,363.84,5.34
hierarchical memory which this was like,365.9,6.4
back in 2005 I think people thought that,369.18,5.7
this was going to lead to AGI,372.3,4.86
um and while it didn't pan out it died,374.88,4.74
of quick death the concept is still very,377.16,3.96
helpful and it's actually coming back,379.62,3.66
with Neuroscience,381.12,4.919
um and so basically here is a simple,383.28,5.28
representation of what I mean so we have,386.039,5.1
all your all your narrative logs right,388.56,5.34
so you have narrative memory which is,391.139,5.461
the the the chronologically linear,393.9,5.34
series of experiences that you have,396.6,4.56
right and that's you know that's what a,399.24,3.72
chat log is right it is chronologically,401.16,3.36
linear,402.96,4.799
um but one problem with chat GPT is we,404.52,4.739
have no idea what is going on in the,407.759,4.021
background one rumor is that it uses a,409.259,4.44
scratch pad I don't know that that's,411.78,4.139
panned out someone else told me that all,413.699,4.44
it does is using uses a rolling window,415.919,5.101
of 8 000 tokens which a rolling window,418.139,4.861
of 8 000 tokens that's a lot of text,421.02,5.04
8000 tokens is like,423.0,4.56
um let's see that's about it's like what,426.06,4.8
3.6 characters per token on average so,427.56,6.84
that is over 24 000 characters,430.86,5.58
um which then you you think that the,434.4,3.72
average word is five to six characters,436.44,2.819
long,438.12,3.359
so we're looking at what can I do math,439.259,4.201
right now I think that's about 6 000,441.479,5.821
words so 6 000 words and you know 250,443.46,6.9
words per page uh so that's uh four,447.3,6.66
times six so that is 24 pages worth of,450.36,6.059
text that's a lot of text if that's how,453.96,5.76
it's working now the chat GPT API,456.419,5.22
um the in the documentation they say,459.72,3.539
that you only have four thousand tokens,461.639,3.96
even still four thousand tokens is 12,463.259,4.921
Pages you can summarize quite a bit in,465.599,4.681
12 pages and this is also going to go up,468.18,4.799
over time right they're they're doubling,470.28,5.52
the window size pretty regularly all,472.979,4.741
right so with all that said,475.8,5.1
we have a linear chunk of logs and so,477.72,4.74
what happens sometimes when you're,480.9,3.299
having conversations is it's like it,482.46,3.42
gets something wrong and then you,484.199,3.361
correct it and then you're like no no no,485.88,3.9
that's not what I wanted and so what I'm,487.56,5.22
what I'm doing with this is I've got a,489.78,5.4
few prompts so in a in in a previous,492.78,3.66
chat,495.18,3.54
um chat uh bot project here let me bring,496.44,3.9
that up I'll show you real quick,498.72,2.9
do,500.34,3.84
repositories come on,501.62,4.24
um so in the original,504.18,4.739
um salience one so this this is this is,505.86,5.519
the the progenitor work right this is I,508.919,3.961
probably won't update this because this,511.379,4.561
repo is just for demonstration purposes,512.88,5.7
um but it shows you the idea of using,515.94,5.399
salience to summarize conversations as,518.58,5.579
well as anticipation to uh predict,521.339,5.221
what's going forward so salience is,524.159,5.041
memory and anticipation is going forward,526.56,4.92
right so that is the beginnings of a,529.2,4.5
cognitive architecture so then you take,531.48,4.14
this forward and say okay we have the,533.7,4.199
same exact uh salience in anticipation,535.62,6.3
but it is organized around,537.899,6.781
um writing right so in this case I'm an,541.92,5.34
AI named Muse my primary goal is to help,544.68,4.02
the user plan brainstorm outline and,547.26,2.639
otherwise construct their work of,548.7,3.36
fiction right so it's got a very clear,549.899,4.141
goal and I won't show you uh the,552.06,3.66
conversation I had because it's you know,554.04,4.38
it's private I'm working on my own story,555.72,4.98
um but having setting the decision the,558.42,4.56
the system go to this and then updating,560.7,4.319
it with the salience and the,562.98,3.299
anticipation,565.019,4.561
it is if all you do is just run this as,566.279,6.0
it is right now you will find that it is,569.58,4.92
very very helpful much more helpful than,572.279,4.261
than the plain vanilla,574.5,4.26
um chat GPT at helping you plan your,576.54,4.38
story and in fact,578.76,5.1
um it is so good that you like man it,580.92,5.099
won't need much help in terms of cueing,583.86,4.32
up the correct memories in order to make,586.019,5.88
it even better now all that being said,588.18,7.38
uh you have a a an arbitrarily long chat,591.899,5.761
log right because writing a novel takes,595.56,3.719
a long time,597.66,2.94
um and it's going to be way more,599.279,3.321
information even just you know,600.6,3.84
brainstorming and stuff right because,602.6,3.28
what happens when you're brainstorming,604.44,4.32
you have uh false positives right you,605.88,4.38
have things that you change things that,608.76,2.94
you say no I'm not gonna I'm not gonna,610.26,4.259
go that way so on and so forth and so,611.7,5.94
what happens is it a chat log is,614.519,4.921
basically like a transaction log so if,617.64,3.54
you're familiar with like SQL databases,619.44,4.2
a transaction log tracks all the changes,621.18,5.279
in the database but it builds on one,623.64,4.92
another right they're all sequential and,626.459,4.201
so we have to keep track of that in,628.56,4.98
chronological order so the the next step,630.66,5.28
is to summarize chunks of the,633.54,4.859
conversation as concisely as possible,635.94,4.62
while extracting the most Salient bits,638.399,3.961
and so that's where these two new,640.56,4.08
prompts come in so I have write an,642.36,4.8
executive summary so it's as simple as,644.64,4.379
it says write an executive summary of,647.16,3.48
the following chat log executive summary,649.019,4.741
so this writes a very concise just,650.64,5.04
here's here's a narrative sequence of,653.76,4.259
what happened and so this allows it to,655.68,4.56
keep track of okay this is this is how,658.019,4.801
the conversation has gone with while,660.24,4.86
excluding a lot of the irrelevant,662.82,5.34
details and then secondly story details,665.1,4.739
so write a summary of what we learn,668.16,3.119
about the story from the following chat,669.839,5.401
log now this should be in the long run,671.279,7.381
this prompt won't work for other tasks,675.24,5.76
right this is this is hard-coded this is,678.66,4.32
geared towards writing fiction because,681.0,4.2
it says specifically,682.98,4.799
extract information about the story in,685.2,4.139
the future the cognitive architecture,687.779,4.5
should design this task or or do task,689.339,5.94
selection to figure out what exactly is,692.279,5.101
most relevant to summarize that being,695.279,4.921
said because Auto Muse is purpose built,697.38,4.92
for fiction we can go ahead and hard,700.2,3.3
code it,702.3,3.659
um but what this does is it ignores the,703.5,3.66
conversation,705.959,3.12
and just says what do we know about the,707.16,2.76
story,709.079,3.121
and I tested this out and it actually,709.92,4.56
works really well so like you can take a,712.2,4.74
long conversation that's 2 000 tokens,714.48,4.68
long and you end up with a 200 token,716.94,4.019
summary of what you learn about the,719.16,2.7
story,720.959,3.601
so that those those summaries the,721.86,4.02
executive summary and the story,724.56,3.42
extraction those are going to be,725.88,4.5
summarized in these chunks right and so,727.98,4.32
you get a 10 to 1 reduction in terms of,730.38,3.48
files because that's how I've got it set,732.3,4.5
right now where every time you get to a,733.86,5.4
10 dialogues back and forth and this is,736.8,4.8
an arbitrary number in fact I should,739.26,3.54
probably,741.6,2.46
um just have this as the the,742.8,3.0
conversation length,744.06,3.24
um I should parameterize that but,745.8,4.44
anyways 10 is fine because you don't,747.3,4.68
want to go right up to the to the window,750.24,4.32
limit because one you don't know how,751.98,4.859
long the conversation is going to be and,754.56,4.62
two you want to regularly summarize,756.839,3.601
what's going on,759.18,4.32
okay so then once we get and this is as,760.44,5.22
far as I got I tested those prompts and,763.5,4.56
if you look at the the two Do's,765.66,4.739
um the the the the very next to do is,768.06,4.68
I've got to actually populate the chunk,770.399,3.601
summaries you see there's nothing here,772.74,3.719
right I've got the chat logs and so now,774.0,4.86
I need to integrate the summaries and,776.459,4.88
put those summaries here,778.86,5.46
and for those I don't know whether or,781.339,4.541
not I need to use semantic embedding,784.32,5.04
right because the the the the primary,785.88,5.88
organization feature here is timestamp,789.36,4.32
so you see chat underscore timestamp,791.76,5.34
user Muse user Muse user Muse and so,793.68,5.7
that is probably all the organization,797.1,4.62
that I actually really need,799.38,4.62
um because again with a temporal,801.72,4.08
hierarchical memory,804.0,4.92
um uh time is the primary organizing,805.8,5.219
Factor not semantic search we actually,808.92,3.599
don't care what's in it semantically,811.019,3.901
because we're going to algorithmically,812.519,4.861
just recursively or not recursively but,814.92,5.58
repeatedly summarize these chat logs so,817.38,4.56
those are going to end up here in the,820.5,2.7
chunk summaries,821.94,3.24
and the chunk summaries are about five,823.2,5.52
to ten times shorter than the originals,825.18,6.3
so that that allows us to really kind of,828.72,4.679
distill down okay what is the most,831.48,3.96
important information and then we're,833.399,4.081
going to do another layer above that,835.44,5.22
which is we're going to extract the,837.48,5.88
topics from these and then the topics,840.66,4.799
are not going to be linear right because,843.36,4.979
you know you think about like okay what,845.459,5.161
do I know about Abraham Lincoln that is,848.339,4.081
not anchored to time that is detached,850.62,4.2
from time now your understanding of,852.42,4.14
Abraham Lincoln might be updated right,854.82,3.36
you might have like you can think of it,856.56,3.48
as an internal Wikipedia page in your,858.18,4.14
head and so what we'll do is we'll,860.04,4.68
transition from this temporal,862.32,4.86
hierarchical part to a more knowledge,864.72,5.28
graph or topical component and so that,867.18,4.68
is the third layer of abstraction So,870.0,5.04
eventually what I want is for an auto,871.86,6.06
Muse it's going to have a top like the,875.04,5.88
whoops yeah come back so the primary you,877.92,4.8
know so the chat logs that is like low,880.92,5.039
level uh uh raw memories then there's,882.72,5.22
the chunk summaries which is one layer,885.959,3.5
above that so you can think of that as,887.94,4.74
medium term memory and then long-term,889.459,6.221
memory is going to be this kg model this,892.68,4.92
knowledge graph model of topics I don't,895.68,3.3
know that it's actually going to be a,897.6,3.359
Knowledge Graph because I don't with,898.98,3.419
semantic search I don't really care,900.959,2.761
about the links,902.399,5.161
between it and also because because the,903.72,5.28
story is going to be updated and,907.56,3.66
summarized repeatedly I don't know that,909.0,4.32
we actually even need semantic search so,911.22,4.919
we might not do any embeddings with this,913.32,5.1
um just by virtue of the memories are,916.139,4.32
going to be so well organized and it's,918.42,3.419
going to be organized around one,920.459,3.901
specific project so here's the thing,921.839,4.62
when we get to the to the to a fully,924.36,4.02
fledged cognitive architecture like,926.459,4.021
Raven where it's like oh hey Raven let's,928.38,4.259
talk about my story you absolutely will,930.48,3.96
need some kind of search so that Raven,932.639,4.621
can locate the correct set of memories,934.44,4.86
um but even still we should probably,937.26,5.28
have some kind of hierarchical,939.3,5.399
um automatic organization in the,942.54,4.26
background ideally it's all kept in,944.699,4.44
natural language because again when we,946.8,4.74
get to the level of fully autonomous AGI,949.139,4.921
you want its thoughts to be entirely,951.54,4.799
transparent and entirely interpretable,954.06,4.8
by humans because like if you have your,956.339,4.261
AGI whether it's Raven or another one,958.86,4.02
thinking hey Gee maybe I should go get,960.6,4.2
access to nuclear launch codes you want,962.88,4.139
to be able to see that in clear text you,964.8,4.74
don't want AGI memories to be in some,967.019,4.861
abstract embedding or AGI specific,969.54,4.62
language we want it all to be in in pure,971.88,4.92
natural language so that one it can,974.16,4.2
understand us and we can understand it,976.8,4.5
and we'll have trust so anyways,978.36,4.68
that's where we're at I didn't get too,981.3,2.94
much further today because again,983.04,3.659
recovering from burnout um but I'm,984.24,3.839
really happy with this especially some,986.699,3.361
of the tests that I did and I apologize,988.079,3.361
I can't show it but what I encourage you,990.06,2.579
to do is,991.44,3.72
um just go ahead and clone this down and,992.639,5.94
run chat it should work as it is,995.16,4.919
um actually you might you might want to,998.579,3.541
comment this out because it'll reset,1000.079,4.56
your conversation every uh 10 dialogues,1002.12,5.399
but just with the anticipation in,1004.639,5.88
salience and having it um having it,1007.519,5.76
having this as the objective it's really,1010.519,5.161
good already it'll help you plan out,1013.279,5.521
your story it won't write Pros for you,1015.68,5.64
um maybe we'll get to that one day uh,1018.8,4.86
but yeah so that's that and then the QA,1021.32,3.239
bot,1023.66,2.88
so talking talking through the memory,1024.559,5.941
here in this case the memory is abstract,1026.54,6.72
it's not it's not narrative the memory,1030.5,6.419
is in the form of papers right and so,1033.26,5.88
this these text documents that's the,1036.919,4.5
memory that we need to work on but then,1039.14,4.919
we need to have other kinds of cognitive,1041.419,4.681
tasks like you know what am I going to,1044.059,4.801
read or um you know what am I going to,1046.1,4.079
summarize or what am I going to,1048.86,3.3
synthesize and so I've been thinking,1050.179,5.581
through this and um and the direction,1052.16,6.18
that I'm going with it is the very next,1055.76,4.44
thing that's going to happen is there's,1058.34,4.38
going to be a cognitive control module,1060.2,4.859
and so as you're talking to it the,1062.72,5.1
cognitive control module will have a,1065.059,4.86
it'll have it'll have a menu of various,1067.82,4.44
things that it can do one it can just,1069.919,3.961
ask you a question ask you a follow-up,1072.26,3.659
question say hey do I know what you're,1073.88,3.179
doing,1075.919,3.961
um it can read papers it can summarize,1077.059,5.341
papers it can synthesize new information,1079.88,4.86
based on information that it's got and,1082.4,4.68
then it can generate hypotheses I think,1084.74,4.319
that's those are the primary cognitive,1087.08,4.68
tasks now it's going to automatically,1089.059,6.061
pick which task it wants to use based on,1091.76,5.46
where you're at in the conversation so,1095.12,3.299
the first thing that happens in the,1097.22,4.02
cognitive control module is that it will,1098.419,4.5
update its task,1101.24,3.12
um so it'll basically have a task log,1102.919,3.721
right and so the task log will be,1104.36,4.559
similar to the scratch Pad,1106.64,4.32
um and it'll be like so we'll have,1108.919,4.321
instead of instead of the chat log Write,1110.96,3.959
a brief summary of the most Salient,1113.24,3.9
points of the conversation we're going,1114.919,3.481
to have,1117.14,3.659
um update the task log based on task,1118.4,5.159
salience right so task sailing says what,1120.799,5.581
is it that we're doing and also instead,1123.559,6.301
of instead of having the goal uh,1126.38,7.62
uh you know I'm an AI Muse my my goal is,1129.86,6.48
to plan brainstorming instead the goal,1134.0,5.4
will be to advance science and so by,1136.34,5.699
tell just by telling chat GPT your goal,1139.4,4.56
is to advance science it will adopt a,1142.039,4.441
new model and a new approach,1143.96,4.64
um to what it's doing and so anyways,1146.48,4.559
these projects will be there's going to,1148.6,4.42
be some very distinct differences and as,1151.039,4.621
I learn about it I'll be able to create,1153.02,4.74
um a more pragmatic higher level,1155.66,3.78
abstraction,1157.76,3.659
um that can be uh more universally,1159.44,3.72
applied,1161.419,5.221
so anyways that's the update for today,1163.16,5.759
um I know that this is probably not the,1166.64,4.38
kind of update you're used to but it is,1168.919,3.781
what it is and uh let me know what you,1171.02,4.399
think in the comments,1172.7,2.719